diff --git a/.zenodo.json b/.zenodo.json new file mode 100644 index 0000000..9e38da7 --- /dev/null +++ b/.zenodo.json @@ -0,0 +1,15 @@ +{ + "title": "CoReNet: Combinatorial Residual Networks for multijet topologies", + "upload_type": "software", + "description": "Reference implementation of CoReNet, a configuration-driven combinatorial residual network for jet assignment and event classification in multijet final states, with the HH to bbbb and HH to bbWW examples of the accompanying paper.", + "creators": [ + {"name": "Khanal, Aniket", "affiliation": "Carnegie Mellon University", "orcid": "0009-0007-5557-9821"}, + {"name": "Alison, John", "affiliation": "Carnegie Mellon University", "orcid": "0000-0003-0843-1641"}, + {"name": "Gomez Espinosa, Alejandro", "affiliation": "Carnegie Mellon University", "orcid": "0000-0002-9443-7769"} + ], + "license": "MIT", + "keywords": ["jet assignment", "Higgs boson pair production", "machine learning", "CMS", "CoReNet"], + "related_identifiers": [ + {"identifier": "10.5281/zenodo.23567532", "relation": "isSupplementedBy", "resource_type": "dataset"} + ] +} diff --git a/CITATION.cff b/CITATION.cff index fae8536..74ac8ba 100644 --- a/CITATION.cff +++ b/CITATION.cff @@ -1,4 +1,4 @@ -# Placeholder: the owner fills in authors, DOI and references before the release. +# The software DOI and the preferred-citation (article) are filled in after the first release. cff-version: 1.2.0 message: "If you use CoReNet, please cite it as below." title: "CoReNet: Combinatorial Residual Network" @@ -8,7 +8,17 @@ date-released: "YYYY-MM-DD" # to be confirmed license: MIT repository-code: "https://github.com/[owner]/CoReNet" # to be confirmed authors: - - family-names: "[to be confirmed]" - given-names: "[to be confirmed]" + - family-names: "Khanal" + given-names: "Aniket" + orcid: "https://orcid.org/0009-0007-5557-9821" + affiliation: "Carnegie Mellon University" + - family-names: "Alison" + given-names: "John" + orcid: "https://orcid.org/0000-0003-0843-1641" + affiliation: "Carnegie Mellon University" + - family-names: "Gomez Espinosa" + given-names: "Alejandro" + orcid: "https://orcid.org/0000-0002-9443-7769" + affiliation: "Carnegie Mellon University" # doi: "10.xxxx/zenodo.xxxxxxx" # to be confirmed # preferred-citation: # to be confirmed (article describing the architecture) diff --git a/README.md b/README.md index 082ef44..daedef5 100644 --- a/README.md +++ b/README.md @@ -21,8 +21,8 @@ pixi shell # activate; or prefix every command with `pixi pixi itself is a single static binary (`curl -fsSL https://pixi.sh/install.sh | sh`) that needs no root access, so the same commands work on a laptop and on a cluster front end. -This installs the command-line tools `corenet-validate`, `corenet-train`, `corenet-evaluate` and -`corenet-export` (each has `--help`), and the Python package `corenet`. The unit tests run with +This installs the command-line tools `corenet-validate`, `corenet-train`, `corenet-evaluate`, +`corenet-export` and `corenet-report` (each has `--help`), and the Python package `corenet`. The unit tests run with `pytest tests`. ## Dependencies @@ -35,6 +35,7 @@ This installs the command-line tools `corenet-validate`, `corenet-train`, `coren | h5py | 3.0 | input and prediction files | | PyYAML | 5.4 | configuration files | | scikit-learn | 1.0 | AUC / ROC in `corenet-evaluate` | +| matplotlib | 3.5 | figures of `corenet-report` | | onnx, onnxruntime | 1.14, 1.15 | `corenet-export` (optional extra `export`) | | pytest | 7 | tests (optional extra `test`) | @@ -62,9 +63,12 @@ corenet-evaluate --checkpoint runs/hh4b/model_fold0.pkl --input examples/data/hh # 4. export to TorchScript and ONNX, and check the exported models against PyTorch corenet-export --checkpoint runs/hh4b/model_fold0.pkl --input examples/data/hh4b_small.h5 \ --output-dir runs/hh4b/export + +# 5. one HTML page with the training curves, ROC curves, pairings and input checks of the run +corenet-report --run-dir runs/hh4b --input examples/data/hh4b_small.h5 --output runs/hh4b/report.html ``` -The same four commands with `bbww` in place of `hh4b` run the HH → bb̄WW example. +The same five commands with `bbww` in place of `hh4b` run the HH → bb̄WW example. [docs/HH4b.md](docs/HH4b.md) walks through the HH → bb̄bb̄ example step by step; [docs/BBWW.md](docs/BBWW.md) does the same for HH → bb̄WW against tt̄. @@ -90,7 +94,7 @@ combine: [HH] `jets` is the jet collection: `max: 12` is the number of jet slots read per event, and the input file must have at least that many. The listed `features` are the inputs of the network: these columns of the input -file, and only these, are fed to it. \ +file, and only these, are fed to it, for the `b` jets and for the `other` jets alike. \ `jets btag 4` selects the four jets with the highest `btag` (alternatively `jets pt 4` would select the four leading jets): any column of the jets in the file can be used this way, and `jets mass 2 lowest` would take the two jets with the lowest value instead. The selector reads its column from the file for this selection only; it @@ -98,7 +102,7 @@ need not be a feature, and a feature named `btag` would be one more network inpu file holds all jets of the event, and the b jets are chosen when it is loaded. For more details on specifying the event decay, check [docs/EventInfo.md](docs/EventInfo.md), and for the input file itself, check [docs/Dataset.md](docs/Dataset.md). The full-statistics files (454,261 HH → bb̄bb̄ events, 292,285 HH → bb̄WW events) are distributed -separately ([DOI / URL to be added]); download them into `examples/data/` as `hh4b_all.h5` and +separately on Zenodo ([doi:10.5281/zenodo.23567532](https://doi.org/10.5281/zenodo.23567532)); download them into `examples/data/` as `hh4b_all.h5` and `bbww_all.h5`. ## Usage @@ -169,6 +173,7 @@ Output: |---|---| | `eval_fold0.h5.json` | the metrics | | `eval_fold0.h5` | the predictions, as below | +| `report.html` | every result of the run on one page, written by `corenet-report` ([Report](#report) below) | | dataset in `eval_fold0.h5` | content | |---|---| @@ -222,6 +227,27 @@ and `event` (N, 2); for HH → bb̄WW `b` (N, 5, 2), `q` (N, 5, 2), `lepton` (N, | `views_` | (N, V) | probability of every pairing of a hypothesis with several pairings | | `head_` | (N,) | the extra binary score of every output after the first | +### Report + +```bash +corenet-report --run-dir runs/bbww --input examples/data/bbww_all.h5 --output runs/bbww/report.html +``` + +* `--folds 0 1`: report these folds only (default: every fold found in the run directory). +* `--max-events 500000`: events of the input file used for the input and weight distributions (random subsample above it; `0` = all). +* `--config my_config.yaml`: event description to use instead of the one stored in the checkpoint. + +To see the training and validation folds in the ROC and score figures (overtraining check), first write the +predictions of every event with `corenet-evaluate ... --output runs/bbww/eval_fold0.h5 --fold -1`. + +Output: + +| file | content | +|---|---| +| `report.html` | self-contained page: run summary, training curves, ROC / AUC / score distributions, view probabilities and pairing accuracy against `TARGETS`, the inputs as the network embeds them, the raw inputs and the event weights ([docs/Report.md](docs/Report.md)) | + +Open `report.html` in any browser; from a cluster, copy it to your machine first (`scp :runs/bbww/report.html .`). + ```python import onnxruntime as ort from corenet import load_config diff --git a/docs/BBWW.md b/docs/BBWW.md index b65b4e7..ef71361 100644 --- a/docs/BBWW.md +++ b/docs/BBWW.md @@ -11,7 +11,7 @@ with its own score. `examples/data/bbww_small.h5` (20,000 events: 1,198 HH, 18,802 tt̄; 10.2 MB) is a random subsample of `examples/data/bbww_all.h5` (292,285 events: 17,511 HH, 274,774 tt̄; 149 MB), which is distributed -separately ([DOI / URL to be added]); download it into `examples/data/`. The layout +separately on Zenodo ([doi:10.5281/zenodo.23567532](https://doi.org/10.5281/zenodo.23567532)); download it into `examples/data/`. The layout ([Dataset.md](./Dataset.md)): | path | shape | meaning | @@ -127,7 +127,14 @@ corenet-evaluate --checkpoint runs/bbww_full/model_fold0.pkl --input examples/da --output runs/bbww_full/eval_fold0.h5 ``` +```bash +corenet-report --run-dir runs/bbww_full --input examples/data/bbww_all.h5 --output runs/bbww_full/report.html +``` + +writes the results page of the run ([Report.md](./Report.md)). + The commands, their options and their output files are the same as in the [HH → bb̄bb̄ guide](./HH4b.md). The prediction file additionally holds `views/tt` (the two tt̄ pairings) and `heads/ttbar` (the tt̄ score), and `class_probabilities` has three columns (rest, -signal, ttbar). The exported model takes `b`, `q`, `lepton`, `met` and `event`. +signal, ttbar). The exported model takes `b`, `q`, `lepton`, `met` and `event`. The report adds the +confusion matrix of the three classes, the ROC curve of every class against the rest and of the tt̄ score. diff --git a/docs/Dataset.md b/docs/Dataset.md index 86aa270..02dfad9 100644 --- a/docs/Dataset.md +++ b/docs/Dataset.md @@ -67,8 +67,9 @@ A collection named `x` is read from `INPUTS/x`, matched case-insensitively, or f of the event description in order (`1` = signal in both examples). * **Weights.** `EVENT/weight` is the event weight (1 when absent). Each class is normalised during training, so only the relative weights within a class matter. Negative weights count as 0. -* **Folds.** The cross-validation fold of an event is `EVENT/fold` when the file has that column. - Without it, pass `--n-folds K` to `corenet-train` (and to `corenet-validate`): the fold is then +* **Folds.** The cross-validation fold of an event is `EVENT/fold` when the file has that column; the + ids of `K` folds must be exactly `0 .. K-1`. Without the column, pass `--n-folds K` to + `corenet-train` (and to `corenet-validate`): the fold is then `squares(event) % K`, a hash of the event number in `EVENT/event` (integers ≥ 0) with the Squares counter-based generator (`corenet.data.folds.fold_of`; `--fold-seed`, default 0, selects the hash key), so any subset of a sample gets the same folds and no random split is ever made. With @@ -76,7 +77,9 @@ A collection named `x` is read from `INPUTS/x`, matched case-insensitively, or f `squares(event) % n_folds`), the number of folds and the seed are stored in the checkpoint, so `corenet-evaluate` applies the same rule to any file. `corenet-train --fold k` holds out fold `k` for testing and `--exclude-folds j` uses fold `j` for validation (`--exclude-folds next`: fold - `(k + 1) % n_folds`). + `(k + 1) % n_folds`). Only selecting a fold needs the folds: `corenet-evaluate --fold -1` and + `corenet-export` also work on a file with neither `EVENT/fold` nor `--n-folds` (the events then + carry fold -1 in the prediction file). * **Event identifier.** `EVENT/event` is copied to the prediction files and is the input of the `squares(event) % n_folds` rule. diff --git a/docs/EventInfo.md b/docs/EventInfo.md index 9487f3b..c7c7ec2 100644 --- a/docs/EventInfo.md +++ b/docs/EventInfo.md @@ -37,7 +37,8 @@ objects: * `jets`: the jet collection, with up to 12 jets per event. The listed features are the inputs of the network: exactly these columns of the input file ([Dataset.md](./Dataset.md)) are fed to it, `pt`, - `eta`, `phi` and `mass` as the kinematics and any other feature as an additional input. + `eta`, `phi` and `mass` as the kinematics and any other feature as an additional input. This holds + for every line taken from the collection, also one used only as `context:` (`other` below). * `b: jets btag 4`: the four jets most likely to be b jets, taken from the jet collection. The `btag` rule reads the `btag` column of the file for this choice only; it need not be a feature. (Listing `btag` under `features` would make it one more input of the network.) @@ -118,7 +119,7 @@ written as: | `W_lep: [lepton, met]` | the leptonic W, formed from the lepton and the missing momentum | `context: ` inside a hypothesis lets its first particles (the Higgs candidates above) also see -the jets of that name, that is, the whole event. +the jets of that name, that is, the whole event, with all the `features` of their collection. Nothing has to be declared about symmetries or combinatorics: products of the same kind are interchangeable (the two b of a Higgs, the two Higgs bosons of HH) and a jet is never used twice in one diff --git a/docs/HH4b.md b/docs/HH4b.md index c9b233e..bc5b8fb 100644 --- a/docs/HH4b.md +++ b/docs/HH4b.md @@ -9,7 +9,7 @@ and export. Every command is run from the repository root in the pixi environmen The repository ships `examples/data/hh4b_small.h5` (20,000 events, 8.1 MB), a random subsample of the full file `examples/data/hh4b_all.h5` (454,261 events: 417,978 multijet, 36,283 HH; 184 MB). The full -file is distributed separately ([Zenodo DOI to be added]); download it into `examples/data/`. Both +file is distributed separately on Zenodo ([doi:10.5281/zenodo.23567532](https://doi.org/10.5281/zenodo.23567532)); download it into `examples/data/`. Both files have the same layout ([Dataset.md](./Dataset.md)): | path | shape | dtype | meaning | @@ -174,6 +174,20 @@ exported model takes the jets already assigned (`b`, `other`) and the event quan of the b jets has to be reproduced wherever the model is deployed; see Exporting in the [README](../README.md). +## Report + +```bash +corenet-evaluate --checkpoint runs/hh4b_full/model_fold0.pkl --input examples/data/hh4b_all.h5 \ + --output runs/hh4b_full/eval_fold0.h5 --fold -1 +corenet-report --run-dir runs/hh4b_full --input examples/data/hh4b_all.h5 --output runs/hh4b_full/report.html +``` + +writes one HTML page with the training curves, the ROC curves and AUC per fold, the signal-score +distributions on the test and training folds, the probabilities of the three pairings and their +accuracy against `TARGETS`, the inputs as the network embeds them, the raw inputs and the event weights +([Report.md](./Report.md)). The first command is optional: it gives the training and validation folds +their own curves. + ## Changing things | to change | where | example | diff --git a/docs/Options.md b/docs/Options.md index 87e4423..e9fc4c2 100644 --- a/docs/Options.md +++ b/docs/Options.md @@ -34,10 +34,10 @@ Changing one of these gives a different network: validate the file with `corenet | `freeze_epoch` | `--freeze-epoch` | 15 | start of the fine-tuning stage: from this epoch on only the layers named by `finetune_layers` are trained (a value ≥ `epochs` disables this) | | `finetune_layers` | `--finetune-layers PREFIX [PREFIX ...]` | `head_main.`, `head_reduce.`, `aux_heads.` (the output layers) | parameter-name prefixes that stay trainable from `freeze_epoch` on; `none` (the word, alone; `[]` in the file) freezes nothing | | `finetune_lr` | `--finetune-lr` | unset | learning rate set at `freeze_epoch`; the `lr_milestones` after it multiply this value by `lr_gamma`; unset keeps the schedule | -| `finetune_batch_size` | `--finetune-batch-size` | unset | batch size from `freeze_epoch` on (the ghost batches of the stage are kept, capped so that none has fewer than 8 events); unset keeps the schedule | +| `finetune_batch_size` | `--finetune-batch-size` | unset | batch size from `freeze_epoch` on; the ghost batches of the stage are kept when they divide it, else the largest smaller number that does, so that every ghost batch has at least 8 events; unset keeps the schedule | | `class_norm` | `--class-norm` / `--no-class-norm` | true | give every class the same total weight in training | | `seed` | `--seed` | 42 | random seed; fold `k` uses `seed + k` | -| `early_stop_patience` | `--early-stop-patience` | 0 | stop when the validation AUC (on the `--exclude-folds` fold) has not improved for this many epochs, and keep the weights of the best epoch; 0 trains every epoch | +| `early_stop_patience` | `--early-stop-patience` | 0 | stop when the validation AUC (on the `--exclude-folds` fold) has not improved for this many epochs, and keep the weights of the best epoch; 0 trains every epoch. The validation fold must differ from the held-out fold and hold events of class 1 and of another class, otherwise the run stops with an error | | `rest_order` | `--rest-order` | `pt` | order of the jets taken by a `rest` line: `pt`, or `selector` to keep the order of the previous candidate line on the same collection (e.g. `b: jets btag 4`) | Write floating-point numbers with a decimal point (`0.005` or `5.0e-3`): YAML reads `5e-3` as a string, diff --git a/docs/Report.md b/docs/Report.md new file mode 100644 index 0000000..755591b --- /dev/null +++ b/docs/Report.md @@ -0,0 +1,90 @@ +# Results Report + +`corenet-report` writes one HTML file with the results of a training and the checks of its inputs. The +page holds every figure and table itself and opens in any browser, offline. + +## Running + +1. Train (`corenet-train`) and, to also see the training and validation folds in the report, evaluate every + event of the file into the same directory: + + ```bash + corenet-evaluate --checkpoint runs/hh4b/model_fold0.pkl --input examples/data/hh4b_all.h5 \ + --output runs/hh4b/eval_fold0.h5 --fold -1 + ``` + +2. Write the report: + + ```bash + corenet-report --run-dir runs/hh4b --input examples/data/hh4b_all.h5 --output runs/hh4b/report.html + ``` + +* `--run-dir runs/hh4b`: the directory written by `corenet-train` (and `corenet-evaluate`). +* `--input examples/data/hh4b_all.h5`: the file the run was trained on. +* `--output runs/hh4b/report.html`: the page to write (default: `report.html` in the run directory). +* `--folds 0 1`: report these folds only (default: every fold found in the directory). +* `--config my_config.yaml`: use this event description instead of the one stored in the checkpoint. +* `--max-events 500000`: at most this many events of the input file for the input and weight + distributions, a random subsample above it; `0` uses every event. +* `--n-folds 5`, `--fold-seed 0`, `--rest-order pt`: the fold rule and the ordering of `rest` jets + (default: as recorded in the run). + +On a machine without a GPU, prefix the command with `CONDA_OVERRIDE_CUDA=12.0`; the report runs on the CPU. + +## Input files + +Per fold `k`, in `--run-dir`: + +| file | content | needed for | +|---|---|---| +| `history_fold{k}.json` | training settings and per-epoch history | run summary, training curves | +| `model_fold{k}.pkl` | checkpoint | event description, parameter count, the inputs as the network sees them | +| `predictions_fold{k}.h5` | predictions on the held-out fold | ROC, AUC, scores, view probabilities, pairing accuracy | +| `eval_fold{k}.h5` | `corenet-evaluate` output; with `--fold -1` every event of the file | the training and validation folds in the ROC, AUC and score figures (overtraining check) | + +A fold with some of these files missing is reported with what is there. + +## Output + +| file | content | +|---|---| +| `report.html` | the page, figures embedded as PNG | + +## Sections of the page + +1. **Run summary**: the event description (collapsed), the input file, the classes and their event counts + and weights, the training settings, and per fold the validation folds, the events of the training, + validation and test sets, the epochs run, the best epoch, the parameter count and which prediction + files were found. +2. **Training curves**: training loss, validation loss and validation AUC per epoch, one line per fold, + the best epoch marked; the values of the last epoch in a table. +3. **Classification**: the ROC curve of the signal score (signal against every other class) per fold and + the mean over folds with its spread, on a linear and a logarithmic background-efficiency axis; the AUC + per fold and split with mean ± sd; the signal-score distributions of every class on the test folds, + overlaid with the training folds when `eval_fold{k}.h5` holds them; for more than two classes the + confusion matrix at the most probable class, the ROC curve of every class against the rest, and the + ROC curve of every auxiliary score (`heads/`). +4. **Hypotheses and pairing**: for every hypothesis with several candidate assignments (for HH → bb̄bb̄ the + three pairings of the four b jets into two Higgs bosons), the distribution of the highest view + probability per class and the share of each assignment; when the input file has `TARGETS` naming the + generator-matched jets of a particle of the decay chain (`TARGETS/h1/{b1,b2}`, `TARGETS/h2/{b1,b2}`), + the fraction of events in which the most probable assignment is the generator one, per fold, for the + events whose matched jets are all among the candidates, and the same fraction as a function of the + view probability. +5. **Inputs as the network sees them**: the quantities the network embeds, per class: for every decay + product of every particle its log(1+pT), η, Δφ to the parent candidate and log(1+m) (plus any further + feature, e.g. `isE`, `isM`), for every particle candidate its own kinematics, the jets of the `context`, + the `relations` pairs (log(1+M), ΔR, Δφ) and the event scalars (`nSelJets` as log(nSelJets − (min_jets − 1))). +6. **Raw inputs**: the columns of the input file for the jets each role selected, slot by slot (including a + column read for the selection only, such as `btag`), the single objects and the event scalars, per class. +7. **Weights**: the event weights per class as stored and as the training uses them (clipped at 0, every + class summing to 1), weight statistics with the effective number of events and the negative weights, the + events and weights per fold and class, and the training set of every fold. +8. **Other checks**: the mean event weight as a function of the leading and the last candidate jet pT and + of the jet multiplicity, per class. + +A figure whose inputs the run did not provide (no `TARGETS`, no predictions on the training folds, ...) is +left out. + +Weighted distributions use the event weights of the file clipped at 0, as the training does; each class is +normalised to unit area. Mean ± sd are over folds. diff --git a/pixi.lock b/pixi.lock index b4bde6c..67d99ab 100644 --- a/pixi.lock +++ b/pixi.lock @@ -26,8 +26,12 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-c-s3-0.13.1-h7508075_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-c-sdkutils-0.2.7-h01ee8f8_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-checksums-0.2.10-h01ee8f8_5.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-1.2.1-h505cf86_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-bin-1.2.1-h9908984_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/bzip2-1.0.8-hda65f42_10.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/c-ares-1.34.8-hebe6cf0_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/cairo-1.18.6-h3c89d7e_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/contourpy-1.4.0-py313h2551ef2_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-crt-tools-12.9.86-ha770c72_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-cudart-12.9.79-h5888daf_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-cuobjdump-12.9.82-hffce074_1.conda @@ -39,16 +43,28 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-nvvm-tools-12.9.86-h4bc722e_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cudnn-9.10.2.21-hbcb9cd8_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/fmt-12.1.0-h76c4fd7_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fontconfig-2.18.3-h4db4eae_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fonttools-4.67.0-py313h6cd9899_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/freetype-2.14.3-ha770c72_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fribidi-1.0.17-h7cc23a3_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/gmp-6.3.0-hfd2156b_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/gmpy2-2.3.2-py313ha4a6282_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/graphite2-1.3.15-h54a6638_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/h5py-3.16.0-nompi_py313hce4e9db_105.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/hdf5-2.2.0-nompi_hc529623_100.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/icu-78.3-py310h44b86e0_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/keyutils-1.6.3-h7cc23a3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/kiwisolver-1.5.1-py313h2551ef2_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/krb5-1.22.2-hbc21106_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/lcms2-2.19.1-h9073bf1_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/ld_impl_linux-64-2.46.1-default_hbd61a6d_102.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/lerc-4.2.0-hdb68285_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libabseil-20260526.0-cxx17_h0dc7533_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libaec-1.1.5-h088129d_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libblas-3.11.0-11_h5875eb1_mkl.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlicommon-1.2.1-h39a168f_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlidec-1.2.1-ha411449_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlienc-1.2.1-h018ffa1_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcap-2.78-h084b8d7_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcblas-3.11.0-11_hfef963f_mkl.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcublas-12.9.2.10-h676940d_0.conda @@ -61,15 +77,22 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/libcurl-8.21.0-hcf29cc6_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcusolver-11.7.5.82-h676940d_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcusparse-12.5.10.65-hecca717_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libdeflate-1.25-hd45a770_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libedit-3.1.20260512-pl5321h373387f_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libev-4.33-h280c20c_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libexpat-2.8.5-hd2095e1_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libffi-3.7.0-h81df57d_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype-2.14.3-ha770c72_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype6-2.14.3-h5e6c136_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-16.2.0-ha9f2e26_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-ng-16.2.0-h69a702a_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgfortran-16.2.0-h69a702a_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgfortran5-16.2.0-h6b99dfc_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libglib-2.90.0-h569388d_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libharfbuzz-14.6.0-h23af247_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libhwloc-2.15.0-default_hd7c4d0c_1001.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libiconv-1.18-h0cb94f2_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libjpeg-turbo-3.2.0-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/liblapack-3.11.0-11_h5e43f62_mkl.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/liblzma-5.8.3-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libmagma-2.10.0-ha7672b3_0.conda @@ -78,21 +101,28 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/libnl-3.11.0-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libnvjitlink-12.9.86-hecca717_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libonnx-1.23.2-hd8ccbc4_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libpng-1.6.59-h922cc85_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libprotobuf-7.35.1-h622638d_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libpython-3.13.16-hdc7f604_101_cp313.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.53.4-h0737f62_103.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libraqm-0.11.0-h6406941_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.54.1-h13e7031_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libssh2-1.11.1-h6154650_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libstdcxx-16.2.0-h934c35e_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libstdcxx-ng-16.2.0-hdf11a46_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libsystemd0-262-h26e0ef7_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libtiff-4.7.2-hcc2c06a_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libtorch-2.13.0-cuda129_mkl_h546cc0b_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libudev1-262-h26e0ef7_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libuuid-2.42.4-hcfc3c73_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libuv-1.52.1-h280c20c_1.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hbdfff7e_0.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-ha7a5771_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libwebp-base-1.6.0-hd42ef1d_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxcb-1.17.0-hb83e432_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hf3af7cc_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-h7df9aa5_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libzlib-1.3.2-h25fd6f3_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/llvm-openmp-23.1.3-h7148c6a_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/markupsafe-3.0.4-py313h6cd9899_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/matplotlib-base-3.11.2-py313he85fb2c_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/mkl-2026.1.0-hd2095e1_245.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/ml_dtypes-0.5.4-np2py313h38606b7_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/mpc-1.4.0-ha2cb11d_1.conda @@ -103,13 +133,19 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/onednn-3.13.3-omp_h64ed1e9_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/onnx-1.23.2-py313hf19c63c_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/onnxruntime-1.30.0-py313h77ceba0_202_cuda129.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/openjpeg-2.5.4-heb1ab33_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/openssl-3.6.5-h781a0a9_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/optree-0.20.0-py313h6bdb861_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pcre2-10.47-h8b3dc9c_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pillow-13.0.0-py313h3b26573_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pixman-0.46.4-h54a6638_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/protobuf-7.35.1-py313hf4667e8_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pthread-stubs-0.4-h7cc23a3_1004.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/python-3.13.16-hf47f18c_101_cp313.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pytorch-2.13.0-cuda129_mkl_py313_h571fb29_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pytorch-gpu-2.13.0-cuda129_mkl_h0d04637_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pyyaml-6.0.3-py313h3dea7bd_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/qhull-2020.2-h434a139_5.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/rdma-core-65.0-h7b4f2e6_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/readline-8.3-hd6e31c0_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/s2n-1.7.6-h7e3ee7f_0.conda @@ -119,7 +155,15 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/tbb-2023.1.0-haa8c3ee_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/tk-8.6.13-noxft_h1df4ec4_4.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/triton-3.7.1-cuda129py313h246eb7c_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libice-1.1.2-h280c20c_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libsm-1.2.6-h0d788c3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libx11-1.8.14-hb83e432_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxau-1.0.12-h7cc23a3_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxdmcp-1.1.5-h7cc23a3_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxext-1.3.7-h7cc23a3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxrender-0.9.13-h7cc23a3_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/yaml-0.2.5-hebe6cf0_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/zlib-ng-2.3.3-hce19668_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/zstd-1.5.7-hb78ec9c_7.conda - conda: https://conda.anaconda.org/conda-forge/noarch/ca-certificates-2026.7.22-hbd8a1cb_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cached-property-2.0.1-pyhcf101f3_0.conda @@ -129,13 +173,21 @@ environments: - conda: https://conda.anaconda.org/conda-forge/noarch/cpython-3.13.16-py313hd8ed1ab_101.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cuda-cudart_linux-64-12.9.79-h3f2d84a_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cuda-version-12.9-h4f385c5_3.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/cycler-0.12.1-pyhcf101f3_2.conda - conda: https://conda.anaconda.org/conda-forge/noarch/exceptiongroup-1.3.1-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/filelock-4.0.12-pyhd8ed1ab_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-dejavu-sans-mono-2.37-hab24e00_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-inconsolata-3.000-h77eed37_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-source-code-pro-2.038-h77eed37_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-ubuntu-0.83-h77eed37_3.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-ecosystem-1-0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-forge-1-hc364b38_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/fsspec-2026.9.0-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/iniconfig-2.3.0-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/jinja2-3.1.6-pyhcf101f3_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/joblib-1.6.0-pyhcf101f3_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/mpmath-1.4.1-pyhd8ed1ab_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/munkres-1.1.4-pyhd8ed1ab_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/narwhals-2.27.0-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/networkx-3.7-pyhc364b38_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/packaging-26.3-pyhc364b38_0.conda @@ -145,10 +197,14 @@ environments: - conda: https://conda.anaconda.org/conda-forge/noarch/pybind11-abi-11-hc364b38_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pybind11-global-3.0.4-pyh648e204_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pygments-2.21.0-pyhcf101f3_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/pyparsing-3.3.3-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pytest-9.1.1-pyhc364b38_2.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/python-dateutil-3.0.0-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/python-flatbuffers-25.12.19-pyh9d96877_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/python-tzdata-2026.5-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/python_abi-3.13-9_cp313.conda - conda: https://conda.anaconda.org/conda-forge/noarch/setuptools-81.0.0-pyh332efcf_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/six-1.17.0-pyhe01879c_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/sympy-1.14.0-pyh2585a3b_106.conda - conda: https://conda.anaconda.org/conda-forge/noarch/threadpoolctl-3.7.0-pyhc455866_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/tomli-2.4.1-pyhcf101f3_0.conda @@ -173,8 +229,12 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-c-s3-0.13.1-h7508075_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-c-sdkutils-0.2.7-h01ee8f8_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/aws-checksums-0.2.10-h01ee8f8_5.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-1.2.1-h505cf86_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-bin-1.2.1-h9908984_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/bzip2-1.0.8-hda65f42_10.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/c-ares-1.34.8-hebe6cf0_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/cairo-1.18.6-h3c89d7e_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/contourpy-1.4.0-py313h2551ef2_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-crt-tools-12.9.86-ha770c72_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-cudart-12.9.79-h5888daf_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-cuobjdump-12.9.82-hffce074_1.conda @@ -186,16 +246,28 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-nvvm-tools-12.9.86-h4bc722e_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/cudnn-9.10.2.21-hbcb9cd8_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/fmt-12.1.0-h76c4fd7_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fontconfig-2.18.3-h4db4eae_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fonttools-4.67.0-py313h6cd9899_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/freetype-2.14.3-ha770c72_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/fribidi-1.0.17-h7cc23a3_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/gmp-6.3.0-hfd2156b_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/gmpy2-2.3.2-py313ha4a6282_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/graphite2-1.3.15-h54a6638_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/h5py-3.16.0-nompi_py313hce4e9db_105.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/hdf5-2.2.0-nompi_hc529623_100.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/icu-78.3-py310h44b86e0_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/keyutils-1.6.3-h7cc23a3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/kiwisolver-1.5.1-py313h2551ef2_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/krb5-1.22.2-hbc21106_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/lcms2-2.19.1-h9073bf1_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/ld_impl_linux-64-2.46.1-default_hbd61a6d_102.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/lerc-4.2.0-hdb68285_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libabseil-20260526.0-cxx17_h0dc7533_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libaec-1.1.5-h088129d_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libblas-3.11.0-11_h5875eb1_mkl.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlicommon-1.2.1-h39a168f_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlidec-1.2.1-ha411449_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlienc-1.2.1-h018ffa1_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcap-2.78-h084b8d7_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcblas-3.11.0-11_hfef963f_mkl.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcublas-12.9.2.10-h676940d_0.conda @@ -208,15 +280,22 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/libcurl-8.21.0-hcf29cc6_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcusolver-11.7.5.82-h676940d_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libcusparse-12.5.10.65-hecca717_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libdeflate-1.25-hd45a770_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libedit-3.1.20260512-pl5321h373387f_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libev-4.33-h280c20c_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libexpat-2.8.5-hd2095e1_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libffi-3.7.0-h81df57d_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype-2.14.3-ha770c72_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype6-2.14.3-h5e6c136_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-16.2.0-ha9f2e26_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-ng-16.2.0-h69a702a_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgfortran-16.2.0-h69a702a_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libgfortran5-16.2.0-h6b99dfc_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libglib-2.90.0-h569388d_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libharfbuzz-14.6.0-h23af247_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libhwloc-2.15.0-default_hd7c4d0c_1001.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libiconv-1.18-h0cb94f2_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libjpeg-turbo-3.2.0-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/liblapack-3.11.0-11_h5e43f62_mkl.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/liblzma-5.8.3-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libmagma-2.10.0-ha7672b3_0.conda @@ -225,21 +304,28 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/libnl-3.11.0-hb03c661_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libnvjitlink-12.9.86-hecca717_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libonnx-1.23.2-hd8ccbc4_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libpng-1.6.59-h922cc85_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libprotobuf-7.35.1-h622638d_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libpython-3.13.16-hdc7f604_101_cp313.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.53.4-h0737f62_103.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libraqm-0.11.0-h6406941_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.54.1-h13e7031_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libssh2-1.11.1-h6154650_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libstdcxx-16.2.0-h934c35e_7.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libstdcxx-ng-16.2.0-hdf11a46_7.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libsystemd0-262-h26e0ef7_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libtiff-4.7.2-hcc2c06a_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libtorch-2.13.0-cuda129_mkl_h546cc0b_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libudev1-262-h26e0ef7_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libuuid-2.42.4-hcfc3c73_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libuv-1.52.1-h280c20c_1.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hbdfff7e_0.conda - - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-ha7a5771_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libwebp-base-1.6.0-hd42ef1d_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxcb-1.17.0-hb83e432_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hf3af7cc_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-h7df9aa5_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/libzlib-1.3.2-h25fd6f3_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/llvm-openmp-23.1.3-h7148c6a_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/markupsafe-3.0.4-py313h6cd9899_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/matplotlib-base-3.11.2-py313he85fb2c_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/mkl-2026.1.0-hd2095e1_245.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/ml_dtypes-0.5.4-np2py313h38606b7_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/mpc-1.4.0-ha2cb11d_1.conda @@ -250,13 +336,19 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/onednn-3.13.3-omp_h64ed1e9_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/onnx-1.23.2-py313hf19c63c_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/onnxruntime-1.30.0-py313h77ceba0_202_cuda129.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/openjpeg-2.5.4-heb1ab33_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/openssl-3.6.5-h781a0a9_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/optree-0.20.0-py313h6bdb861_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pcre2-10.47-h8b3dc9c_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pillow-13.0.0-py313h3b26573_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pixman-0.46.4-h54a6638_3.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/protobuf-7.35.1-py313hf4667e8_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/pthread-stubs-0.4-h7cc23a3_1004.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/python-3.13.16-hf47f18c_101_cp313.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pytorch-2.13.0-cuda129_mkl_py313_h571fb29_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pytorch-gpu-2.13.0-cuda129_mkl_h0d04637_304.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/pyyaml-6.0.3-py313h3dea7bd_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/qhull-2020.2-h434a139_5.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/rdma-core-65.0-h7b4f2e6_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/readline-8.3-hd6e31c0_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/s2n-1.7.6-h7e3ee7f_0.conda @@ -266,7 +358,15 @@ environments: - conda: https://conda.anaconda.org/conda-forge/linux-64/tbb-2023.1.0-haa8c3ee_2.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/tk-8.6.13-noxft_h1df4ec4_4.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/triton-3.7.1-cuda129py313h246eb7c_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libice-1.1.2-h280c20c_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libsm-1.2.6-h0d788c3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libx11-1.8.14-hb83e432_0.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxau-1.0.12-h7cc23a3_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxdmcp-1.1.5-h7cc23a3_2.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxext-1.3.7-h7cc23a3_1.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxrender-0.9.13-h7cc23a3_0.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/yaml-0.2.5-hebe6cf0_3.conda + - conda: https://conda.anaconda.org/conda-forge/linux-64/zlib-ng-2.3.3-hce19668_1.conda - conda: https://conda.anaconda.org/conda-forge/linux-64/zstd-1.5.7-hb78ec9c_7.conda - conda: https://conda.anaconda.org/conda-forge/noarch/ca-certificates-2026.7.22-hbd8a1cb_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cached-property-2.0.1-pyhcf101f3_0.conda @@ -276,13 +376,21 @@ environments: - conda: https://conda.anaconda.org/conda-forge/noarch/cpython-3.13.16-py313hd8ed1ab_101.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cuda-cudart_linux-64-12.9.79-h3f2d84a_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/cuda-version-12.9-h4f385c5_3.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/cycler-0.12.1-pyhcf101f3_2.conda - conda: https://conda.anaconda.org/conda-forge/noarch/exceptiongroup-1.3.1-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/filelock-4.0.12-pyhd8ed1ab_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-dejavu-sans-mono-2.37-hab24e00_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-inconsolata-3.000-h77eed37_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-source-code-pro-2.038-h77eed37_0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-ubuntu-0.83-h77eed37_3.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-ecosystem-1-0.tar.bz2 + - conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-forge-1-hc364b38_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/fsspec-2026.9.0-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/iniconfig-2.3.0-pyhd8ed1ab_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/jinja2-3.1.6-pyhcf101f3_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/joblib-1.6.0-pyhcf101f3_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/mpmath-1.4.1-pyhd8ed1ab_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/munkres-1.1.4-pyhd8ed1ab_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/narwhals-2.27.0-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/networkx-3.7-pyhc364b38_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/packaging-26.3-pyhc364b38_0.conda @@ -292,10 +400,14 @@ environments: - conda: https://conda.anaconda.org/conda-forge/noarch/pybind11-abi-11-hc364b38_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pybind11-global-3.0.4-pyh648e204_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pygments-2.21.0-pyhcf101f3_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/pyparsing-3.3.3-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/pytest-9.1.1-pyhc364b38_2.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/python-dateutil-3.0.0-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/python-flatbuffers-25.12.19-pyh9d96877_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/python-tzdata-2026.5-pyh5ded981_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/python_abi-3.13-9_cp313.conda - conda: https://conda.anaconda.org/conda-forge/noarch/setuptools-81.0.0-pyh332efcf_0.conda + - conda: https://conda.anaconda.org/conda-forge/noarch/six-1.17.0-pyhe01879c_1.conda - conda: https://conda.anaconda.org/conda-forge/noarch/sympy-1.14.0-pyh2585a3b_106.conda - conda: https://conda.anaconda.org/conda-forge/noarch/threadpoolctl-3.7.0-pyhc455866_0.conda - conda: https://conda.anaconda.org/conda-forge/noarch/tomli-2.4.1-pyhcf101f3_0.conda @@ -468,6 +580,37 @@ packages: - aws-checksums >=0.2.10,<0.2.11.0a0 size: 101887 timestamp: 1785159213897 +- conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-1.2.1-h505cf86_0.conda + sha256: 3efb8e6ac07f12773f02e876567700f554965e33a8e30ee05ddd891c303f7aa3 + md5: 65f1fc5f6a3a984a8a25641d0e60fc25 + depends: + - __glibc >=2.17,<3.0.a0 + - brotli-bin 1.2.1 h9908984_0 + - libbrotlidec 1.2.1 ha411449_0 + - libbrotlienc 1.2.1 h018ffa1_0 + - libgcc >=15 + license: MIT + purls: [] + run_exports: + weak: + - libbrotlicommon >=1.2.1,<1.3.0a0 + - libbrotlienc >=1.2.1,<1.3.0a0 + - libbrotlidec >=1.2.1,<1.3.0a0 + size: 20477 + timestamp: 1792653712185 +- conda: https://conda.anaconda.org/conda-forge/linux-64/brotli-bin-1.2.1-h9908984_0.conda + sha256: c64939f0ade0fea29525d2f549a700841a331c0886b3694083d08eb4769936ce + md5: 903eceb31743f0ce070abb9b9f879714 + depends: + - __glibc >=2.17,<3.0.a0 + - libbrotlidec 1.2.1 ha411449_0 + - libbrotlienc 1.2.1 h018ffa1_0 + - libgcc >=15 + license: MIT + purls: [] + run_exports: {} + size: 21494 + timestamp: 1792653706181 - conda: https://conda.anaconda.org/conda-forge/linux-64/bzip2-1.0.8-hda65f42_10.conda sha256: 1a0d382c515ebf55f8ee1f38c8b81bc95af5c2acc42ad53b66bc5df932032f96 md5: e675fabcf81499adc7edf58124fb1e01 @@ -496,6 +639,53 @@ packages: - c-ares >=1.34.8,<2.0a0 size: 228700 timestamp: 1787169971173 +- conda: https://conda.anaconda.org/conda-forge/linux-64/cairo-1.18.6-h3c89d7e_0.conda + sha256: 5d0efc6f9981468aca72897460d9b2d17b74dff72787f20a9d2330e214064515 + md5: 7920701f9abcb488411495742cf365a7 + depends: + - __glibc >=2.17,<3.0.a0 + - fontconfig >=2.18.3,<3.0a0 + - fonts-conda-ecosystem + - icu >=78.3,<79.0a0 + - libexpat >=2.8.1,<3.0a0 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - libgcc >=15 + - libglib >=2.90.0,<3.0a0 + - libpng >=1.6.58,<1.7.0a0 + - libstdcxx >=15 + - libxcb >=1.17.0,<2.0a0 + - libzlib >=1.3.2,<2.0a0 + - pixman >=0.46.4,<1.0a0 + - xorg-libice >=1.1.2,<2.0a0 + - xorg-libsm >=1.2.6,<2.0a0 + - xorg-libx11 >=1.8.13,<2.0a0 + - xorg-libxext >=1.3.7,<2.0a0 + - xorg-libxrender >=0.9.12,<0.10.0a0 + license: LGPL-2.1-only or MPL-1.1 + purls: [] + run_exports: + weak: + - cairo >=1.18.6,<2.0a0 + size: 1007110 + timestamp: 1790045358757 +- conda: https://conda.anaconda.org/conda-forge/linux-64/contourpy-1.4.0-py313h2551ef2_1.conda + sha256: 078f4a0f83ee7d0f88c110e15f9f94ff4080dc925711d1d22b52779c2e48cf87 + md5: 1883f23b947ffaa13f9511ea817fc4e6 + depends: + - numpy >=2.0 + - python + - __glibc >=2.17,<3.0.a0 + - libstdcxx >=15 + - libgcc >=15 + - python_abi 3.13.* *_cp313 + license: BSD-3-Clause + license_family: BSD + purls: + - pkg:pypi/contourpy?source=compressed-mapping + run_exports: {} + size: 329540 + timestamp: 1789746824998 - conda: https://conda.anaconda.org/conda-forge/linux-64/cuda-crt-tools-12.9.86-ha770c72_2.conda sha256: 2da9964591af14ba11b2379bed01d56e7185260ee0998d1a939add7fb752db45 md5: 503a94e20d2690d534d676a764a1852c @@ -648,6 +838,70 @@ packages: - fmt >=12.1.0,<12.2.0a0 size: 199072 timestamp: 1785915412102 +- conda: https://conda.anaconda.org/conda-forge/linux-64/fontconfig-2.18.3-h4db4eae_1.conda + sha256: 5a3eb10b18a97223ab06b3a7f0d7f56658db2f2800e2f2af95836fe3bf55ba63 + md5: 922776b528a470ab5afa81fd42abfa1d + depends: + - __glibc >=2.17,<3.0.a0 + - libexpat >=2.8.1,<3.0a0 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - libgcc >=15 + - libuuid >=2.42.2,<3.0a0 + - libzlib >=1.3.2,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - fontconfig >=2.18.3,<3.0a0 + - fonts-conda-ecosystem + size: 296288 + timestamp: 1786667377340 +- conda: https://conda.anaconda.org/conda-forge/linux-64/fonttools-4.67.0-py313h6cd9899_1.conda + sha256: b81de60784d902b27da442aacf9bf8f059b0aa778178a7d453bc654c43457ec1 + md5: dd719c40c297d87fe93085248d2ff07b + depends: + - __glibc >=2.17,<3.0.a0 + - brotli + - libgcc >=15 + - munkres + - python >=3.13,<3.14.0a0 + - python_abi 3.13.* *_cp313 + license: MIT + license_family: MIT + purls: + - pkg:pypi/fonttools?source=compressed-mapping + run_exports: {} + size: 3161654 + timestamp: 1792672552740 +- conda: https://conda.anaconda.org/conda-forge/linux-64/freetype-2.14.3-ha770c72_2.conda + sha256: 612a8e0c1a6ecae23da54d81ef395f6ebb5c8e4468ec2611593ebce75876a4e0 + md5: 2d0ea23b23603e07ca47f23f949f67b6 + depends: + - libfreetype 2.14.3 ha770c72_2 + - libfreetype6 2.14.3 h5e6c136_2 + license: GPL-2.0-only OR FTL + purls: [] + run_exports: + weak: + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + size: 175239 + timestamp: 1786641011029 +- conda: https://conda.anaconda.org/conda-forge/linux-64/fribidi-1.0.17-h7cc23a3_0.conda + sha256: 3ba79ebfbd14ca5db4affaf10ee8001064e907de5baabc2dbe5bcd9aec856c9f + md5: f9d5302bca196ee87e62f393f2626562 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + license: LGPL-2.1-or-later + purls: [] + run_exports: + weak: + - fribidi >=1.0.17,<2.0a0 + size: 64102 + timestamp: 1790011855888 - conda: https://conda.anaconda.org/conda-forge/linux-64/gmp-6.3.0-hfd2156b_3.conda sha256: f36c336efc874f346a53a9011f67e997f9faac505562cc18c13b6d399cfe61c7 md5: 576e32739f323438bf69ab006c21be7b @@ -680,6 +934,21 @@ packages: run_exports: {} size: 327168 timestamp: 1791166953052 +- conda: https://conda.anaconda.org/conda-forge/linux-64/graphite2-1.3.15-h54a6638_1.conda + sha256: 7fa3b6a9c081fa3e545573152a788d061a0a0ba57df7251cc0f4f75225fc93e7 + md5: f9fe2984587fa8235a6af6004760cd18 + depends: + - __glibc >=2.17,<3.0.a0 + - libstdcxx >=14 + - libgcc >=14 + license: LGPL-2.0-or-later + license_family: LGPL + purls: [] + run_exports: + weak: + - graphite2 >=1.3.15,<2.0a0 + size: 102835 + timestamp: 1786118485753 - conda: https://conda.anaconda.org/conda-forge/linux-64/h5py-3.16.0-nompi_py313hce4e9db_105.conda sha256: 9cbf3ffbf85a46abefed6b8127a738db9e7407fb3020aa7fdeeb75f66930d43d md5: 6b2d14128466dceabd35a3df2abe1aac @@ -725,6 +994,21 @@ packages: - hdf5 >=2.2.0,<3.0a0 size: 4250161 timestamp: 1786233343747 +- conda: https://conda.anaconda.org/conda-forge/linux-64/icu-78.3-py310h44b86e0_2.conda + sha256: 9f07834f0c546ab14d885ce0366285f61f44e326c0edd1fc63b8294e113ae432 + md5: 72a381cbad04f24b1c2a43ef707f45b4 + depends: + - __glibc >=2.17,<3.0.a0 + - libstdcxx >=14 + - libgcc >=14 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - icu >=78.3,<79.0a0 + size: 14459115 + timestamp: 1786545741408 - conda: https://conda.anaconda.org/conda-forge/linux-64/keyutils-1.6.3-h7cc23a3_1.conda sha256: dd053c96dcb0dcfd59422aefea9d2fe937a190167f34fba7893ec1e10a7e8963 md5: ba55d1b89fd7775e67de8291029b4059 @@ -738,6 +1022,22 @@ packages: - keyutils >=1.6.3,<2.0a0 size: 135295 timestamp: 1786739238128 +- conda: https://conda.anaconda.org/conda-forge/linux-64/kiwisolver-1.5.1-py313h2551ef2_3.conda + sha256: bb5beb208c8379eee9b3ce620684ac44ae73597ae0dab500c12f18e63e3c1020 + md5: 45c8698dd577ba504513001219cfd0fd + depends: + - python + - libstdcxx >=15 + - libgcc >=15 + - __glibc >=2.17,<3.0.a0 + - python_abi 3.13.* *_cp313 + license: BSD-3-Clause + license_family: BSD + purls: + - pkg:pypi/kiwisolver?source=compressed-mapping + run_exports: {} + size: 75111 + timestamp: 1790082004004 - conda: https://conda.anaconda.org/conda-forge/linux-64/krb5-1.22.2-hbc21106_2.conda sha256: 2a5c38c85e63df84c4e69ee71439841ce570d259ae3060627bb9a49a938d66f4 md5: 53318d715316929a574f83591308b1f8 @@ -757,6 +1057,22 @@ packages: - krb5 >=1.22.2,<1.23.0a0 size: 1394333 timestamp: 1786762112514 +- conda: https://conda.anaconda.org/conda-forge/linux-64/lcms2-2.19.1-h9073bf1_3.conda + sha256: 03e675b9adaf235aab880e4133c72f4830e87bc7da2df59d6194771612239b22 + md5: 596464faa06e0b656b35416188ef9bba + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libjpeg-turbo >=3.2.0,<4.0a0 + - libtiff >=4.7.2,<4.8.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - lcms2 >=2.19.1,<3.0a0 + size: 254292 + timestamp: 1789051947923 - conda: https://conda.anaconda.org/conda-forge/linux-64/ld_impl_linux-64-2.46.1-default_hbd61a6d_102.conda sha256: 27d83f1188cd19bcb7754a078b3fa7f4cfb8527f8eb2fde54dd01fc529d1adec md5: 449500f2c089da11c40f5c21312e3e07 @@ -771,6 +1087,21 @@ packages: run_exports: {} size: 745303 timestamp: 1784214507189 +- conda: https://conda.anaconda.org/conda-forge/linux-64/lerc-4.2.0-hdb68285_0.conda + sha256: bf9fdebf55d8bc99d83531cdffda00703f4dc5f93a1a956768c147362c72feda + md5: fb9d356b1a57d6d54768be7ebd5fce09 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=14 + - libstdcxx >=14 + license: Apache-2.0 + license_family: Apache + purls: [] + run_exports: + weak: + - lerc >=4.2.0,<5.0a0 + size: 271158 + timestamp: 1785036167977 - conda: https://conda.anaconda.org/conda-forge/linux-64/libabseil-20260526.0-cxx17_h0dc7533_2.conda sha256: 7bb3d495411b7059e85225aae500d84e7846a4bb02ebd77e4450200b20c180f0 md5: 6983bfe8e09992014b9cf393769c7a84 @@ -827,6 +1158,47 @@ packages: - libblas >=3.11.0,<4.0a0 size: 18622 timestamp: 1789061051585 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlicommon-1.2.1-h39a168f_0.conda + sha256: 8eccfbfdb8ecba32eabfb6ee92f70e818e5f3a35a01c507f7b6bd56e00e4c617 + md5: 9460bdde20bcdf9e6bbef0232b14ad2d + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + license: MIT + purls: [] + run_exports: + weak: + - libbrotlicommon >=1.2.1,<1.3.0a0 + size: 80370 + timestamp: 1792653685570 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlidec-1.2.1-ha411449_0.conda + sha256: 4b4bda6f70ad511bdff95fdd958dca74b6e0534ae335ddcb1d8c6569774d2429 + md5: 6d934e235f4d452ebec720ec7c3638a6 + depends: + - __glibc >=2.17,<3.0.a0 + - libbrotlicommon 1.2.1 h39a168f_0 + - libgcc >=15 + license: MIT + purls: [] + run_exports: + weak: + - libbrotlidec >=1.2.1,<1.3.0a0 + size: 34703 + timestamp: 1792653694124 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libbrotlienc-1.2.1-h018ffa1_0.conda + sha256: 61974cf9bb07115ed6725dacf83d0d985443d78dd123cc4f2758856fda5376ec + md5: 8f3078564b4a4cd9602ebaafe78e0b87 + depends: + - __glibc >=2.17,<3.0.a0 + - libbrotlicommon 1.2.1 h39a168f_0 + - libgcc >=15 + license: MIT + purls: [] + run_exports: + weak: + - libbrotlienc >=1.2.1,<1.3.0a0 + size: 298098 + timestamp: 1792653699767 - conda: https://conda.anaconda.org/conda-forge/linux-64/libcap-2.78-h084b8d7_1.conda sha256: 8cb25174d6b6fac95d31e86cfe41faffc8ee9dacbf2bfd22e6c23377e8f338c1 md5: 5db514adf5f843126ff846d1510f22a4 @@ -1020,6 +1392,20 @@ packages: run_exports: {} size: 208846028 timestamp: 1761069913328 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libdeflate-1.25-hd45a770_1.conda + sha256: 82e134c8a08b1eed9a2ed8ab578b89aa1730dcde3dea8dd87645ed0637878e54 + md5: 40f9b31aa9cf007789867df0decd0492 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=14 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - libdeflate >=1.25,<1.26.0a0 + size: 73710 + timestamp: 1785908694612 - conda: https://conda.anaconda.org/conda-forge/linux-64/libedit-3.1.20260512-pl5321h373387f_0.conda sha256: 45301f14f4eb66da57cfaf2d265f3eeee73f8b07e99614ee3afc5b26eea91126 md5: 27dbd7e2984b8fe48e069379636ccfdd @@ -1077,6 +1463,31 @@ packages: - libffi >=3.7.0,<3.8.0a0 size: 68004 timestamp: 1787753412298 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype-2.14.3-ha770c72_2.conda + sha256: fb12ecb46c30d18d928c0e8fc346ab13b5f6fc8a9cacae4f2f4bb188782586f5 + md5: f8054e759d0ddddaf34a5c8fedc900a1 + depends: + - libfreetype6 >=2.14.3 + license: GPL-2.0-only OR FTL + purls: [] + run_exports: {} + size: 8407 + timestamp: 1786641007099 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libfreetype6-2.14.3-h5e6c136_2.conda + sha256: ec607dd5445dd17ff6bf8b7fe6832e5504c226dd91d3aaea7ba83569808b0ce4 + md5: b72a266a9317036fd0464cb98b027cd0 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libpng >=1.6.58,<1.7.0a0 + - libzlib >=1.3.2,<2.0a0 + constrains: + - freetype >=2.14.3 + license: GPL-2.0-only OR FTL + purls: [] + run_exports: {} + size: 387671 + timestamp: 1786641006460 - conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-16.2.0-ha9f2e26_7.conda sha256: e031634c3a928f9594eba868bc46ef5230d6571a364a0c63215b06c4563bfcea md5: 735629b400b70eeab21a2ba8f5b1c2ce @@ -1092,6 +1503,19 @@ packages: run_exports: {} size: 1058444 timestamp: 1790303109576 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libgcc-ng-16.2.0-h69a702a_7.conda + sha256: 03b54d7008de080a17cc2174b0e4acd6d06e1c981b2eb19846f46d8f59966034 + md5: 2005515ac9340f816fb2ac7e83afe553 + depends: + - libgcc 16.2.0 ha9f2e26_7 + license: GPL-3.0-only WITH GCC-exception-3.1 + license_family: GPL + purls: [] + run_exports: + strong: + - libgcc + size: 28616 + timestamp: 1790303114248 - conda: https://conda.anaconda.org/conda-forge/linux-64/libgfortran-16.2.0-h69a702a_7.conda sha256: 9682ae4dd11487e10b904b3cb3ab6379cad25233d45fdfdf90d54349df244494 md5: 3a8d285f5f22138ef826ad54cf267e73 @@ -1119,6 +1543,46 @@ packages: run_exports: {} size: 2528180 timestamp: 1790303122004 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libglib-2.90.0-h569388d_0.conda + sha256: 9e688e046550415c030c3d08ed9a3fd13a51d7bc56295283b858b6c3070bb9f8 + md5: c85e9539551f1fd21104e518d940c244 + depends: + - libgcc >=15 + - __glibc >=2.17,<3.0.a0 + - libiconv >=1.18,<2.0a0 + - pcre2 >=10.47,<10.48.0a0 + - libffi >=3.7.0,<3.8.0a0 + - libzlib >=1.3.2,<2.0a0 + constrains: + - glib >2.66 + license: LGPL-2.1-or-later + purls: [] + run_exports: + weak: + - libglib >=2.90.0,<3.0a0 + size: 4819130 + timestamp: 1789478025196 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libharfbuzz-14.6.0-h23af247_2.conda + sha256: 3dea71a07100fd523ff173538577fd2e06a42beb5cb25e4a18ac9484773753ba + md5: 6a93147cab011f36d1659745c3f579e2 + depends: + - __glibc >=2.17,<3.0.a0 + - cairo >=1.18.6,<2.0a0 + - graphite2 >=1.3.15,<2.0a0 + - icu >=78.3,<79.0a0 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - libgcc >=15 + - libglib >=2.90.0,<3.0a0 + - libpng >=1.6.59,<1.7.0a0 + - libstdcxx >=15 + - libzlib >=1.3.2,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: {} + size: 1463302 + timestamp: 1791900600568 - conda: https://conda.anaconda.org/conda-forge/linux-64/libhwloc-2.15.0-default_hd7c4d0c_1001.conda sha256: a140f8b94caad6ea27e9deba66a74ae2566976591db8a867677c652add6ee5a0 md5: 66c755b3bd9ec82308fa6ec7674f473f @@ -1149,6 +1613,21 @@ packages: - libiconv >=1.18,<2.0a0 size: 789471 timestamp: 1787033836207 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libjpeg-turbo-3.2.0-hb03c661_1.conda + sha256: bba8538e6538ed58a8479b332337b96986561f975d06cfa2039a016c2d246ee4 + md5: 898d1c9793eaa52efc4727bd84d2e39a + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=14 + constrains: + - jpeg <0.0.0a + license: IJG AND BSD-3-Clause AND Zlib + purls: [] + run_exports: + weak: + - libjpeg-turbo >=3.2.0,<4.0a0 + size: 650434 + timestamp: 1785896381946 - conda: https://conda.anaconda.org/conda-forge/linux-64/liblapack-3.11.0-11_h5e43f62_mkl.conda build_number: 11 sha256: 7b80fc195419463533262b9d327c75559dd820251902e3fa41cce02a3b23bb5b @@ -1280,6 +1759,20 @@ packages: - libonnx >=1.23.2,<1.23.3.0a0 size: 1676952 timestamp: 1791291889441 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libpng-1.6.59-h922cc85_0.conda + sha256: c7381be8ef713501f6cd2f5ad473e401f2edeaec3c46788296b8dbf0de165e15 + md5: 8c9caf742a485ac688ffa88bbadbf0be + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libzlib >=1.3.2,<2.0a0 + license: zlib-acknowledgement + purls: [] + run_exports: + weak: + - libpng >=1.6.59,<1.7.0a0 + size: 316872 + timestamp: 1790805159503 - conda: https://conda.anaconda.org/conda-forge/linux-64/libprotobuf-7.35.1-h622638d_3.conda sha256: 37c51fbc936a6bb8bf5f67c8f056edacaaa41e8603738bc8a4dee186292bb114 md5: fd307b61cceb36fdca38e1718bdb2893 @@ -1313,22 +1806,39 @@ packages: - libpython >=3.13,<4.0a0 size: 9718885 timestamp: 1791334276485 -- conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.53.4-h0737f62_103.conda - sha256: 66f4831f4ee3055f0ffe94e01901198cf3601e2c1b4f6b51d05c3f2d8557ce1d - md5: cc1ad2520049c35344239454a17560ac +- conda: https://conda.anaconda.org/conda-forge/linux-64/libraqm-0.11.0-h6406941_0.conda + sha256: ab3ccb9fd5816f76b18ee8974d359cd2605ca87d8ddef55369dc461b9986f95c + md5: 3ac89a48d224409739dbf6200e524373 + depends: + - libgcc >=14 + - __glibc >=2.28,<3.0.a0 + - fribidi >=1.0.16,<2.0a0 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - libharfbuzz >=14.2.1 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - libraqm >=0.11.0,<0.12.0a0 + size: 33983 + timestamp: 1784850267262 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libsqlite-3.54.1-h13e7031_0.conda + sha256: 07b3c98282022f05781122d073f84e944bfb4299946e8457ece3ce6c384ff78e + md5: 8f15f0ec9b69344df5c3bb5cce1751a1 depends: - __glibc >=2.17,<3.0.a0 + - icu >=78.3,<79.0a0 - libgcc >=15 - libzlib >=1.3.2,<2.0a0 - constrains: - - icu <0.0a0 license: blessing purls: [] run_exports: weak: - - libsqlite >=3.53.4,<4.0a0 - size: 967489 - timestamp: 1791288185694 + - libsqlite >=3.54.1,<4.0a0 + size: 983668 + timestamp: 1792175587601 - conda: https://conda.anaconda.org/conda-forge/linux-64/libssh2-1.11.1-h6154650_1.conda sha256: 7e463000fa1cba3f3a6597b776f6ab301d425a96e3bc20d0d9d76a3b9f237aa3 md5: 5c526561e1d2a2e071941174eae84557 @@ -1359,6 +1869,19 @@ packages: run_exports: {} size: 6615436 timestamp: 1790303132935 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libstdcxx-ng-16.2.0-hdf11a46_7.conda + sha256: 337f481db36cc357884f05f404ccd1b3a7e0fc1ee48ec5c884e4281d213bc863 + md5: 25a23bde5e26d548374f3e02b4669670 + depends: + - libstdcxx 16.2.0 h934c35e_7 + license: GPL-3.0-only WITH GCC-exception-3.1 + license_family: GPL + purls: [] + run_exports: + strong: + - libstdcxx + size: 28665 + timestamp: 1790303164510 - conda: https://conda.anaconda.org/conda-forge/linux-64/libsystemd0-262-h26e0ef7_2.conda sha256: 265ba1e0bf848b9da2a33f7bec7745bc8d69bc2452e2d9fc75f340492a3b010d md5: e76269d6aff0b947103cbe37a96b5737 @@ -1371,6 +1894,27 @@ packages: run_exports: {} size: 572864 timestamp: 1790777898360 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libtiff-4.7.2-hcc2c06a_1.conda + sha256: 10e125da82ca93e09191e66e5a94d566b2cf8d4024e2a0db3a9114297447e0d9 + md5: 0907d876f460ebc941ae590338b6102f + depends: + - __glibc >=2.17,<3.0.a0 + - lerc >=4.2.0,<5.0a0 + - libdeflate >=1.25,<1.26.0a0 + - libgcc >=15 + - libjpeg-turbo >=3.2.0,<4.0a0 + - liblzma >=5.8.3,<6.0a0 + - libstdcxx >=15 + - libwebp-base >=1.6.0,<2.0a0 + - libzlib >=1.3.2,<2.0a0 + - zstd >=1.5.7,<1.6.0a0 + license: HPND + purls: [] + run_exports: + weak: + - libtiff >=4.7.2,<4.8.0a0 + size: 459753 + timestamp: 1787755584172 - conda: https://conda.anaconda.org/conda-forge/linux-64/libtorch-2.13.0-cuda129_mkl_h546cc0b_304.conda sha256: dbbcc77563fa466db03327374ee096bfe7e8b88d5fd285023c6fea64fc15c277 md5: dce74602df0dadf0c21b861f1e33b62c @@ -1460,36 +2004,68 @@ packages: - libuv >=1.52.1,<2.0a0 size: 420040 timestamp: 1785914567661 -- conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hbdfff7e_0.conda - sha256: b6f96287c408269f5067bb4f7a61693ff5c856152902362964d9807152b9c146 - md5: 176fa262c40e1141e9f364d3e4783795 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libwebp-base-1.6.0-hd42ef1d_1.conda + sha256: 8415001414f488c85b72b9d8cc2071dfb3981a47bc3c8eb56ef91a57d12eae7f + md5: 9332b53d0ea93c5d39e33be03a0c611a depends: - __glibc >=2.17,<3.0.a0 + - libgcc >=14 + constrains: + - libwebp 1.6.0 + license: BSD-3-Clause + license_family: BSD + purls: [] + run_exports: + weak: + - libwebp-base >=1.6.0,<2.0a0 + size: 428430 + timestamp: 1785954557217 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libxcb-1.17.0-hb83e432_2.conda + sha256: 7b49e6fd2a584b7f072943e6adf4149677af5121246975278804782d37752837 + md5: 61f0ffba9161cbb3a77722869b21e4bb + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - pthread-stubs + - xorg-libxau >=1.0.12,<2.0a0 + - xorg-libxdmcp >=1.1.5,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - libxcb >=1.17.0,<2.0a0 + size: 395120 + timestamp: 1787885788278 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-16-2.15.4-hf3af7cc_1.conda + sha256: d1906463070c21f3215cd5de93164f31c44d5bac6ae9b5d8f5adf42b1e304039 + md5: 046d9eada488e1a1657c80b55c23da52 + depends: + - __glibc >=2.17,<3.0.a0 + - icu >=78.3,<79.0a0 - libgcc >=15 - libiconv >=1.18,<2.0a0 - liblzma >=5.8.3,<6.0a0 - libzlib >=1.3.2,<2.0a0 constrains: - - icu <0.0a0 - libxml2 2.15.4 license: MIT license_family: MIT purls: [] run_exports: {} - size: 565717 - timestamp: 1788635209625 -- conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-ha7a5771_0.conda - sha256: db52c038430a508656530fc7b5650ae3e2629875693a6a256cc48dec203f8bfe - md5: c322ae4591dfe866b2a9b1e8537fe805 + size: 569062 + timestamp: 1791465631497 +- conda: https://conda.anaconda.org/conda-forge/linux-64/libxml2-2.15.4-h7df9aa5_1.conda + sha256: 8c32be9e8720bc413d2b4267517459d7d71fad4a5f398a2aa8319b6d7b8a00e6 + md5: 51e5162f95bd6840f6b2b3169125fff5 depends: - __glibc >=2.17,<3.0.a0 + - icu >=78.3,<79.0a0 - libgcc >=15 - libiconv >=1.18,<2.0a0 - liblzma >=5.8.3,<6.0a0 - - libxml2-16 2.15.4 hbdfff7e_0 + - libxml2-16 2.15.4 hf3af7cc_1 - libzlib >=1.3.2,<2.0a0 - constrains: - - icu <0.0a0 license: MIT license_family: MIT purls: [] @@ -1497,8 +2073,8 @@ packages: weak: - libxml2 - libxml2-16 >=2.15.4 - size: 46952 - timestamp: 1788635214834 + size: 46570 + timestamp: 1791465635923 - conda: https://conda.anaconda.org/conda-forge/linux-64/libzlib-1.3.2-h25fd6f3_3.conda sha256: eb8a0db0aa570124f7d2a93d7c7f596e3390df5e047818d873baad32985fc736 md5: 0de0122d9570a8ab637c6b73db268389 @@ -1548,6 +2124,38 @@ packages: run_exports: {} size: 26579 timestamp: 1791014766127 +- conda: https://conda.anaconda.org/conda-forge/linux-64/matplotlib-base-3.11.2-py313he85fb2c_1.conda + sha256: 0cdfcbf58ca1cab8a941c2dbb1c958ea211fbb92cd262626fb766c06615d5d5e + md5: 85e795dd9e44b84961fafb4e7dd589aa + depends: + - __glibc >=2.17,<3.0.a0 + - contourpy >=1.0.1 + - cycler >=0.10 + - fonttools >=4.28.2 + - freetype + - kiwisolver >=1.3.1 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - libgcc >=15 + - libraqm >=0.11.0,<0.12.0a0 + - libstdcxx >=15 + - numpy >=1.25 + - numpy >=1.25,<3 + - packaging >=20.0 + - pillow >=9 + - pyparsing >=3 + - python >=3.13,<3.14.0a0 + - python-dateutil >=2.7 + - python_abi 3.13.* *_cp313 + - qhull >=2020.2,<2020.3.0a0 + - tk >=8.6.13,<8.7.0a0 + license: PSF-2.0 + license_family: PSF + purls: + - pkg:pypi/matplotlib?source=compressed-mapping + run_exports: {} + size: 9026765 + timestamp: 1789824128731 - conda: https://conda.anaconda.org/conda-forge/linux-64/mkl-2026.1.0-hd2095e1_245.conda sha256: a2699b4db5f93ae7211249f5b8df4981c22765c392cc6d34c18d76994fe3b327 md5: 5c8b5643ac9b96cd1c5b65a17337ce5c @@ -1728,6 +2336,24 @@ packages: run_exports: {} size: 294012437 timestamp: 1790746169696 +- conda: https://conda.anaconda.org/conda-forge/linux-64/openjpeg-2.5.4-heb1ab33_2.conda + sha256: 131688a88bea8da92fe418c4558c5073435e2848911bc4c836c3a9dd5994b4aa + md5: a3f3ac7a68d01c198e276dbb99e62032 + depends: + - libgcc >=15 + - __glibc >=2.17,<3.0.a0 + - libstdcxx >=15 + - libtiff >=4.7.2,<4.8.0a0 + - libzlib >=1.3.2,<2.0a0 + - libpng >=1.6.58,<1.7.0a0 + license: BSD-2-Clause + license_family: BSD + purls: [] + run_exports: + weak: + - openjpeg >=2.5.4,<3.0a0 + size: 391242 + timestamp: 1788425045145 - conda: https://conda.anaconda.org/conda-forge/linux-64/openssl-3.6.5-h781a0a9_0.conda sha256: 9c1533f1cf784f0e250919c0c9762d0b0e6bc727a79290ffb7cea1ceca1d038c md5: 5a51a2ea653f120240b1a9bf76e0aae9 @@ -1760,6 +2386,61 @@ packages: run_exports: {} size: 586381 timestamp: 1788504569784 +- conda: https://conda.anaconda.org/conda-forge/linux-64/pcre2-10.47-h8b3dc9c_1.conda + sha256: 9ebb10a4eb3be51ae67d85c679f5a7a50cb040ed25c783e6331a225ea09991d4 + md5: 06f6113cf1ff4a54b65f87ead132a5f1 + depends: + - __glibc >=2.17,<3.0.a0 + - bzip2 >=1.0.8,<2.0a0 + - libgcc >=15 + - libzlib >=1.3.2,<2.0a0 + license: BSD-3-Clause + license_family: BSD + purls: [] + run_exports: + weak: + - pcre2 >=10.47,<10.48.0a0 + size: 1218833 + timestamp: 1787294571916 +- conda: https://conda.anaconda.org/conda-forge/linux-64/pillow-13.0.0-py313h3b26573_0.conda + sha256: e9efd1a2015165aafd118ea1bef35b70b4cd8bb0a7629bc6a2d6d92ae1d01aae + md5: f31c982db861132e73aa9bde2ebdc0b1 + depends: + - python + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libxcb >=1.17.0,<2.0a0 + - libfreetype >=2.14.3 + - libfreetype6 >=2.14.3 + - lcms2 >=2.19.1,<3.0a0 + - libwebp-base >=1.6.0,<2.0a0 + - openjpeg >=2.5.4,<3.0a0 + - python_abi 3.13.* *_cp313 + - zlib-ng >=2.3.3,<2.4.0a0 + - tk >=8.6.13,<8.7.0a0 + - libtiff >=4.7.2,<4.8.0a0 + - libjpeg-turbo >=3.2.0,<4.0a0 + license: HPND + purls: + - pkg:pypi/pillow?source=compressed-mapping + run_exports: {} + size: 1083173 + timestamp: 1792416406105 +- conda: https://conda.anaconda.org/conda-forge/linux-64/pixman-0.46.4-h54a6638_3.conda + sha256: 829d8288764282de5a9f7b9169acb75cc7dc0b6c3fe2535cfe87dea3436bbc5d + md5: 0ee5bb30034b081a1386c1e2c98ab0a7 + depends: + - libstdcxx >=14 + - libgcc >=14 + - __glibc >=2.17,<3.0.a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - pixman >=0.46.4,<1.0a0 + size: 376704 + timestamp: 1786106621354 - conda: https://conda.anaconda.org/conda-forge/linux-64/protobuf-7.35.1-py313hf4667e8_3.conda sha256: 78296e9a95fb21402682043e79a8bbc829fc9cc317d7ff818fcb07bf4182eb2a md5: e835bdf081aab510728d2c4994492f10 @@ -1781,6 +2462,18 @@ packages: run_exports: {} size: 492698 timestamp: 1788687323172 +- conda: https://conda.anaconda.org/conda-forge/linux-64/pthread-stubs-0.4-h7cc23a3_1004.conda + sha256: 4a44fd00ea73b79ca2c89b0727b9ccf61c506ead71e67a9abfa4c590042b5a4a + md5: bb66b610707b811e3c65181a6431e7a8 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + license: MIT + license_family: MIT + purls: [] + run_exports: {} + size: 9630 + timestamp: 1788381877753 - conda: https://conda.anaconda.org/conda-forge/linux-64/python-3.13.16-hf47f18c_101_cp313.conda build_number: 101 sha256: 914ce0022ee08951329ce85d8a35f55fa11e74239568bca6102375d630ca9e61 @@ -1872,7 +2565,7 @@ packages: license: BSD-3-Clause license_family: BSD purls: - - pkg:pypi/torch?source=compressed-mapping + - pkg:pypi/torch?source=hash-mapping run_exports: weak: - pytorch >=2.13.0,<2.14.0a0 @@ -1906,6 +2599,20 @@ packages: run_exports: {} size: 201616 timestamp: 1770223543730 +- conda: https://conda.anaconda.org/conda-forge/linux-64/qhull-2020.2-h434a139_5.conda + sha256: 776363493bad83308ba30bcb88c2552632581b143e8ee25b1982c8c743e73abc + md5: 353823361b1d27eb3960efb076dfcaf6 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc-ng >=12 + - libstdcxx-ng >=12 + license: LicenseRef-Qhull + purls: [] + run_exports: + weak: + - qhull >=2020.2,<2020.3.0a0 + size: 552937 + timestamp: 1720813982144 - conda: https://conda.anaconda.org/conda-forge/linux-64/rdma-core-65.0-h7b4f2e6_0.conda sha256: 2da73957fc57eb46efc0d42bb8d73fa3bf36ea98c02a31ce17de06a40d2a1884 md5: b47757f1512801cd064f0f973896d724 @@ -2072,6 +2779,109 @@ packages: run_exports: {} size: 40349627 timestamp: 1781881959760 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libice-1.1.2-h280c20c_0.conda + sha256: 49b532d1df875c6749d9078b56a76f3f5db49a5abe0ca620b593ed474ef0ebf1 + md5: 85c9442aec283b4e464fa9ecc484a2f3 + depends: + - libgcc >=14 + - __glibc >=2.17,<3.0.a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libice >=1.1.2,<2.0a0 + size: 62517 + timestamp: 1786474410404 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libsm-1.2.6-h0d788c3_1.conda + sha256: ef907caee0665cf3b0f775602cc50e388e3bade8ce22ecade0873ff26604b2fd + md5: aa7459ed9ad086ba11dda843d764b33d + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=14 + - xorg-libice >=1.1.2,<2.0a0 + - libuuid >=2.42.2,<3.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libsm >=1.2.6,<2.0a0 + size: 30739 + timestamp: 1786545374265 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libx11-1.8.14-hb83e432_0.conda + sha256: 1815dc9900bf3a4678caeb343bccbbc74cd5e5a5ff3f6e0a2c01c5e19cd9d30b + md5: e7752792cb214013a34e265cc0838352 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libxcb >=1.17.0,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libx11 >=1.8.14,<2.0a0 + size: 846774 + timestamp: 1792545407029 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxau-1.0.12-h7cc23a3_2.conda + sha256: 3ec065b94554dc48a4ca582a960a5484bc166b26e83ce0954653e08d17e8bd53 + md5: da33efee1a93603d92e06d4a2b68def6 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libxau >=1.0.12,<2.0a0 + size: 18793 + timestamp: 1788960512381 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxdmcp-1.1.5-h7cc23a3_2.conda + sha256: 8a095df9bc1d2e50f2504ce8858a047ae29ea634aad6cc0b43298c63361da6d9 + md5: eabfc45786244de252090cf8d55128a8 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libxdmcp >=1.1.5,<2.0a0 + size: 22606 + timestamp: 1789676943335 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxext-1.3.7-h7cc23a3_1.conda + sha256: aa9bbe8b278aacc194e280ff5037f9f9a1f2c5b33ed97de8e7f01cfbe90dda43 + md5: e5b6b28536b81b3f4cb20db4668a4642 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - xorg-libx11 >=1.8.13,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libxext >=1.3.7,<2.0a0 + size: 53124 + timestamp: 1787100841900 +- conda: https://conda.anaconda.org/conda-forge/linux-64/xorg-libxrender-0.9.13-h7cc23a3_0.conda + sha256: 008d610d6ec1d021ad678024afd5dcf534ff84d9c1e9d2877ff36d112c89896f + md5: 7249ce52789c3680e889e7335a2e1c3e + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - xorg-libx11 >=1.8.13,<2.0a0 + license: MIT + license_family: MIT + purls: [] + run_exports: + weak: + - xorg-libxrender >=0.9.13,<0.10.0a0 + size: 34604 + timestamp: 1792530399214 - conda: https://conda.anaconda.org/conda-forge/linux-64/yaml-0.2.5-hebe6cf0_3.conda sha256: d164dfa75ecd538f6fd68765defcc06aa875bc697b9b215362d79a2a73125dd0 md5: e741576fb8f89821ac7c1c537322a33d @@ -2086,6 +2896,21 @@ packages: - yaml >=0.2.5,<0.3.0a0 size: 84936 timestamp: 1787228426393 +- conda: https://conda.anaconda.org/conda-forge/linux-64/zlib-ng-2.3.3-hce19668_1.conda + sha256: 8b786bb4380fa69a718b08a400040b865bc9207eda337e0a1955717c3c3a9403 + md5: 1726acec19beeed1c764385cd8622405 + depends: + - __glibc >=2.17,<3.0.a0 + - libgcc >=15 + - libstdcxx >=15 + license: Zlib + license_family: Other + purls: [] + run_exports: + weak: + - zlib-ng >=2.3.3,<2.4.0a0 + size: 123959 + timestamp: 1786736890973 - conda: https://conda.anaconda.org/conda-forge/linux-64/zstd-1.5.7-hb78ec9c_7.conda sha256: 47d682b9f6d6ec9eb1a6e6c3e75ea6273e899e78fb7fc59f81d39745009fbc60 md5: aa459086047c0e5e27023ab19f8cb86a @@ -2193,6 +3018,19 @@ packages: run_exports: {} size: 21578 timestamp: 1746134436166 +- conda: https://conda.anaconda.org/conda-forge/noarch/cycler-0.12.1-pyhcf101f3_2.conda + sha256: bb47aec5338695ff8efbddbc669064a3b10fe34ad881fb8ad5d64fbfa6910ed1 + md5: 4c2a8fef270f6c69591889b93f9f55c1 + depends: + - python >=3.10 + - python + license: BSD-3-Clause + license_family: BSD + purls: + - pkg:pypi/cycler?source=hash-mapping + run_exports: {} + size: 14778 + timestamp: 1764466758386 - conda: https://conda.anaconda.org/conda-forge/noarch/exceptiongroup-1.3.1-pyhd8ed1ab_0.conda sha256: ee6cf346d017d954255bbcbdb424cddea4d14e4ed7e9813e429db1d795d01144 md5: 8e662bd460bda79b1ea39194e3c4c9ab @@ -2217,6 +3055,67 @@ packages: run_exports: {} size: 85969 timestamp: 1791210321663 +- conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-dejavu-sans-mono-2.37-hab24e00_0.tar.bz2 + sha256: 58d7f40d2940dd0a8aa28651239adbf5613254df0f75789919c4e6762054403b + md5: 0c96522c6bdaed4b1566d11387caaf45 + license: BSD-3-Clause + license_family: BSD + purls: [] + run_exports: {} + size: 397370 + timestamp: 1566932522327 +- conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-inconsolata-3.000-h77eed37_0.tar.bz2 + sha256: c52a29fdac682c20d252facc50f01e7c2e7ceac52aa9817aaf0bb83f7559ec5c + md5: 34893075a5c9e55cdafac56607368fc6 + license: OFL-1.1 + license_family: Other + purls: [] + run_exports: {} + size: 96530 + timestamp: 1620479909603 +- conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-source-code-pro-2.038-h77eed37_0.tar.bz2 + sha256: 00925c8c055a2275614b4d983e1df637245e19058d79fc7dd1a93b8d9fb4b139 + md5: 4d59c254e01d9cde7957100457e2d5fb + license: OFL-1.1 + license_family: Other + purls: [] + run_exports: {} + size: 700814 + timestamp: 1620479612257 +- conda: https://conda.anaconda.org/conda-forge/noarch/font-ttf-ubuntu-0.83-h77eed37_3.conda + sha256: 2821ec1dc454bd8b9a31d0ed22a7ce22422c0aef163c59f49dfdf915d0f0ca14 + md5: 49023d73832ef61042f6a237cb2687e7 + license: LicenseRef-Ubuntu-Font-Licence-Version-1.0 + license_family: Other + purls: [] + run_exports: {} + size: 1620504 + timestamp: 1727511233259 +- conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-ecosystem-1-0.tar.bz2 + sha256: a997f2f1921bb9c9d76e6fa2f6b408b7fa549edd349a77639c9fe7a23ea93e61 + md5: fee5683a3f04bd15cbd8318b096a27ab + depends: + - fonts-conda-forge + license: BSD-3-Clause + license_family: BSD + purls: [] + run_exports: {} + size: 3667 + timestamp: 1566974674465 +- conda: https://conda.anaconda.org/conda-forge/noarch/fonts-conda-forge-1-hc364b38_1.conda + sha256: 54eea8469786bc2291cc40bca5f46438d3e062a399e8f53f013b6a9f50e98333 + md5: a7970cd949a077b7cb9696379d338681 + depends: + - font-ttf-ubuntu + - font-ttf-inconsolata + - font-ttf-dejavu-sans-mono + - font-ttf-source-code-pro + license: BSD-3-Clause + license_family: BSD + purls: [] + run_exports: {} + size: 4059 + timestamp: 1762351264405 - conda: https://conda.anaconda.org/conda-forge/noarch/fsspec-2026.9.0-pyhd8ed1ab_0.conda sha256: 7d17272012e7479ba58f8a72b522d3e88e6837a04a111a278ff417aa82364255 md5: 42de5cd170647599a84f468a42fabb89 @@ -2281,6 +3180,18 @@ packages: run_exports: {} size: 464918 timestamp: 1773662068273 +- conda: https://conda.anaconda.org/conda-forge/noarch/munkres-1.1.4-pyhd8ed1ab_1.conda + sha256: d09c47c2cf456de5c09fa66d2c3c5035aa1fa228a1983a433c47b876aa16ce90 + md5: 37293a85a0f4f77bbd9cf7aaefc62609 + depends: + - python >=3.9 + license: Apache-2.0 + license_family: Apache + purls: + - pkg:pypi/munkres?source=hash-mapping + run_exports: {} + size: 15851 + timestamp: 1749895533014 - conda: https://conda.anaconda.org/conda-forge/noarch/narwhals-2.27.0-pyh5ded981_0.conda sha256: 93be72d3cfd946e08295f9c77e0e3896caad7b1f818e17388d2e682f60b8925d md5: 9e4512fe1aa400d3c08d1ced8eab3139 @@ -2406,6 +3317,19 @@ packages: run_exports: {} size: 959376 timestamp: 1786995678795 +- conda: https://conda.anaconda.org/conda-forge/noarch/pyparsing-3.3.3-pyh5ded981_0.conda + sha256: dc10d64d4b9a1310083b2f563416ab45560ee2a04fd6baab40f223f2b1ab37bc + md5: 8bf43ae1312e9b966063459e04c20fa7 + depends: + - python >=3.11 + - python + license: MIT + license_family: MIT + purls: + - pkg:pypi/pyparsing?source=compressed-mapping + run_exports: {} + size: 114046 + timestamp: 1790063282695 - conda: https://conda.anaconda.org/conda-forge/noarch/pytest-9.1.1-pyhc364b38_2.conda sha256: 430051d80765207a7d782b2b188230ba1489d35c6e75fd9903f76cb9fda4af16 md5: 64c98a12c4e23eb238bf66bbecafdf3c @@ -2428,6 +3352,20 @@ packages: run_exports: {} size: 306724 timestamp: 1782127176429 +- conda: https://conda.anaconda.org/conda-forge/noarch/python-dateutil-3.0.0-pyh5ded981_0.conda + sha256: 8ad3e3fd71fa3881081a38d660cd9cc7bd18fd4eb99cee96225903e5e2eb9ae9 + md5: d75c60076c9f2cfa4a459e9f17cecd93 + depends: + - python >=3.11 + - six >=1.6 + - python-tzdata + - python + license: Apache-2.0 + purls: + - pkg:pypi/python-dateutil?source=compressed-mapping + run_exports: {} + size: 89526 + timestamp: 1792674886408 - conda: https://conda.anaconda.org/conda-forge/noarch/python-flatbuffers-25.12.19-pyh9d96877_0.conda sha256: dff2f2fc462996f71686c05a0fba78a0dc919034ef8d879dd8366534ad64c0f1 md5: 3a28cf4d2187c1dfb0c31aedbf275b8c @@ -2440,6 +3378,19 @@ packages: run_exports: {} size: 35556 timestamp: 1784053026634 +- conda: https://conda.anaconda.org/conda-forge/noarch/python-tzdata-2026.5-pyh5ded981_0.conda + sha256: b66b783be882069a9f6c00bc44ee561fc3dfba21fe4bab6de31599caf3c63154 + md5: 4a0ae2a06b469e6f8ca95d0655e31c86 + depends: + - python >=3.11 + - python + license: Apache-2.0 + license_family: APACHE + purls: + - pkg:pypi/tzdata?source=compressed-mapping + run_exports: {} + size: 141379 + timestamp: 1791025923003 - conda: https://conda.anaconda.org/conda-forge/noarch/python_abi-3.13-9_cp313.conda build_number: 9 sha256: 9dc0cb6a20ebc7e9bcd8c6868f2bec083c114b0ba20b7118405c036533d5c522 @@ -2464,6 +3415,19 @@ packages: run_exports: {} size: 676888 timestamp: 1770456470072 +- conda: https://conda.anaconda.org/conda-forge/noarch/six-1.17.0-pyhe01879c_1.conda + sha256: 458227f759d5e3fcec5d9b7acce54e10c9e1f4f4b7ec978f3bfd54ce4ee9853d + md5: 3339e3b65d58accf4ca4fb8748ab16b3 + depends: + - python >=3.9 + - python + license: MIT + license_family: MIT + purls: + - pkg:pypi/six?source=hash-mapping + run_exports: {} + size: 18455 + timestamp: 1753199211006 - conda: https://conda.anaconda.org/conda-forge/noarch/sympy-1.14.0-pyh2585a3b_106.conda sha256: 1c8057e6875eba958aa8b3c1a072dc9a75d013f209c26fd8125a5ebd3abbec0c md5: 32d866e43b25275f61566b9391ccb7b5 @@ -2545,6 +3509,7 @@ packages: - h5py>=3.0 - pyyaml>=5.4 - scikit-learn>=1.0 + - matplotlib>=3.5 - onnx>=1.14 ; extra == 'export' - onnxruntime>=1.15 ; extra == 'export' - pytest>=7 ; extra == 'test' diff --git a/pixi.toml b/pixi.toml index adcf7e5..d3429e2 100644 --- a/pixi.toml +++ b/pixi.toml @@ -17,6 +17,7 @@ numpy = "*" h5py = "*" pyyaml = "*" scikit-learn = "*" +matplotlib-base = "*" pytest = "*" onnx = "*" onnxruntime = "*" diff --git a/pyproject.toml b/pyproject.toml index 9bfa8b8..0bb615a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -17,6 +17,7 @@ dependencies = [ "h5py>=3.0", "pyyaml>=5.4", "scikit-learn>=1.0", + "matplotlib>=3.5", ] keywords = [ "particle-physics", @@ -49,6 +50,7 @@ corenet-train = "corenet.train:main" corenet-evaluate = "corenet.evaluate:cli" corenet-export = "corenet.export:cli" corenet-validate = "corenet.validate:cli" +corenet-report = "corenet.report:cli" [tool.setuptools.packages.find] where = ["src"] diff --git a/src/corenet/data/folds.py b/src/corenet/data/folds.py index 885862a..a975ede 100644 --- a/src/corenet/data/folds.py +++ b/src/corenet/data/folds.py @@ -19,12 +19,14 @@ import hashlib from collections.abc import Iterable +from typing import Union import numpy as np _UINT64_32 = np.uint64(32) -SeedLike = int | str | Iterable[int | str] +# a runtime alias (not an annotation): typing.Union, so that the module imports on Python 3.9 +SeedLike = Union[int, str, Iterable[Union[int, str]]] # noqa: UP007 def _str_to_entropy(text: str) -> list[np.uint64]: diff --git a/src/corenet/data/h5.py b/src/corenet/data/h5.py index 56b5d46..9c38dfe 100644 --- a/src/corenet/data/h5.py +++ b/src/corenet/data/h5.py @@ -11,16 +11,21 @@ inputs["event"] (N, len(spec.event)) event scalars in the configured order plus ``labels`` (``EVENT/label``, or ``CLASSIFICATIONS/EVENT/signal`` when absent), ``weights`` -(``EVENT/weight``), ``folds`` and ``events`` (``EVENT/event``). The cross-validation fold of an event -is ``EVENT/fold`` by default; with ``n_folds`` it is ``fold_of(event, n_folds, fold_seed)`` = -``squares(event) % n_folds`` (``corenet.data.folds``) computed from the event numbers, and the file -needs no ``EVENT/fold`` column. No generator-level +(``EVENT/weight``), ``folds`` and ``events`` (``EVENT/event``), and ``selected`` (role -> (N, n) int64: +the slot of the file each role slot was taken from, -1 for a padded slot; ``corenet-report`` uses it to +compare the candidates with ``TARGETS``). The cross-validation fold of an event +is ``EVENT/fold`` by default, whose ids must be exactly 0..K-1 for K folds; with ``n_folds`` it is +``fold_of(event, n_folds, fold_seed)`` = ``squares(event) % n_folds`` (``corenet.data.folds``) +computed from the event numbers, and the file needs no ``EVENT/fold`` column. A file with neither +loads without folds (``has_folds`` False, every fold -1): evaluating or exporting on every event works, +selecting a fold (``fold_indices``: training, evaluating one fold) raises. No generator-level information is read: the roles are chosen from reconstructed quantities only. A collection's ``features`` are exactly the channels of its tensors. A selector reads the column it needs (its column and ``pt``) from the file on its own, whether or not it is a feature; such a column is used for the selection only (``columns_to_read``). A column listed under ``features`` is always a -network input, ``btag`` included. +network input, ``btag`` included. Columns are read at the file's dtype and sorted at that precision; +they are cast to float32 only when the input tensors are built. ds = H5Dataset("examples/data/bbww_small.h5", load_config("examples/bbww/bbww.yaml")) out = model(ds.inputs) @@ -64,10 +69,12 @@ "CLASSIFICATIONS/EVENT/signal", ) # first one present is used WEIGHT_KEY, FOLD_KEY, EVENT_KEY = "EVENT/weight", "EVENT/fold", "EVENT/event" -FOLD_RULE_EVENT, FOLD_RULE_COLUMN = ( +FOLD_RULE_EVENT, FOLD_RULE_COLUMN, FOLD_RULE_NONE = ( "squares(event) % n_folds", "EVENT/fold", -) # H5Dataset.fold_rule + "none", +) # H5Dataset.fold_rule; "none": the file has no EVENT/fold and no n_folds was given +NO_FOLD = -1 # the fold of every event of a file loaded without folds # ------------------------------------------------------------------ what the loader reads (shared with corenet-validate) @@ -111,11 +118,22 @@ def resolve_group(f: h5py.File, name: str) -> str: ) +def _sortable(a: np.ndarray) -> np.ndarray: + """A column as read from the file, in a dtype that sorts exactly and can be negated: floats and + signed integers unchanged, bool -> int8, unsigned integers -> int64.""" + if a.dtype.kind == "b": + return a.astype(np.int8) + if a.dtype.kind == "u": + return a.astype(np.int64) + return a + + def _read_collection( f: h5py.File, coll: Collection, columns: list[str] ) -> tuple[dict[str, np.ndarray], np.ndarray]: """(columns, present) of one collection, both shaped (N, slots in the file); ``columns`` are the - features plus the selector columns (``columns_to_read``).""" + features plus the selector columns (``columns_to_read``), at the file's dtype (``_sortable``) so that + the selectors sort at full precision; ``_stack`` casts to float32.""" group = f["INPUTS"][resolve_group(f, coll.name)] missing = [k for k in columns if k not in group] if missing: @@ -123,7 +141,7 @@ def _read_collection( f"collection {coll.name!r} (INPUTS/{group.name.split('/')[-1]}): missing columns {missing} " f"(features {coll.features}, selector columns {[c for c in columns if c not in coll.features]})" ) - feats = {k: np.atleast_2d(group[k][...].astype(np.float32).T).T for k in columns} + feats = {k: np.atleast_2d(_sortable(group[k][...]).T).T for k in columns} n, slots = feats[columns[0]].shape if coll.max > 1 and slots < coll.max: raise ValueError( @@ -168,7 +186,7 @@ def __init__( fold_seed: int = FOLD_SEED, ): """``n_folds``: use fold = squares(event) % n_folds (``fold_of``, seed ``fold_seed``) instead of the - ``EVENT/fold`` column.""" + ``EVENT/fold`` column. Without both, the file loads without folds (``has_folds`` is False).""" if rest_order not in ("pt", "selector"): raise ValueError( f"rest_order must be 'pt' or 'selector', got {rest_order!r}" @@ -180,6 +198,9 @@ def __init__( self._last_keys: dict[ str, tuple[np.ndarray, ...] ] = {} # per collection: sort keys of the last explicit selector + self.selected: dict[ + str, np.ndarray + ] = {} # per role: (N, n) file slot of every role slot, -1 when padded inputs: dict[str, torch.Tensor] = {} with h5py.File(path, "r") as f: @@ -213,14 +234,28 @@ def __init__( else torch.arange(n, dtype=torch.int64) ) self.fold_seed = int(fold_seed) - self.folds, self.n_folds, self.fold_rule = self._folds(f, n_folds) + self.folds, self.n_folds, self.fold_rule = self._folds(f, n_folds, n) self.inputs = inputs + @property + def has_folds(self) -> bool: + """False when the file has no ``EVENT/fold`` and no ``n_folds`` was given: every fold is NO_FOLD.""" + return self.fold_rule != FOLD_RULE_NONE + + def require_folds(self, what: str = "selecting a fold") -> None: + """Raise KeyError with the remedy when the file was loaded without folds.""" + if not self.has_folds: + raise KeyError( + f"{self.path.name}: no {FOLD_KEY} dataset; {what} needs the fold of every event: pass --n-folds K " + f"(H5Dataset(n_folds=K)) to use fold = squares(event) % K from {EVENT_KEY}" + ) + def _folds( - self, f: h5py.File, n_folds: int | None + self, f: h5py.File, n_folds: int | None, n: int ) -> tuple[torch.Tensor, int, str]: - """(fold of every event, number of folds, rule): ``EVENT/fold`` (n_folds = the number of distinct - values) or, with ``n_folds``, ``fold_of(EVENT/event, n_folds, self.fold_seed)``.""" + """(fold of every event, number of folds K, rule): with ``n_folds``, ``fold_of(EVENT/event, n_folds, + self.fold_seed)``; else ``EVENT/fold``, whose ids must be exactly 0..K-1; else (no column) every fold + is NO_FOLD, K = 0 and the rule is FOLD_RULE_NONE.""" if n_folds is not None: rule = fold_rule(n_folds, self.fold_seed) if EVENT_KEY not in f: @@ -234,12 +269,16 @@ def _folds( folds = fold_of(f[EVENT_KEY][...], n_folds, self.fold_seed) return torch.from_numpy(folds), n_folds, FOLD_RULE_EVENT if FOLD_KEY not in f: - raise KeyError( - f"{self.path.name}: no {FOLD_KEY} dataset; pass --n-folds K (H5Dataset(n_folds=K)) " - f"to use fold = squares(event) % K from {EVENT_KEY}" - ) + return torch.full((n,), NO_FOLD, dtype=torch.int64), 0, FOLD_RULE_NONE folds = f[FOLD_KEY][...].astype(np.int64) - return torch.from_numpy(folds), len(np.unique(folds)), FOLD_RULE_COLUMN + ids = np.unique(folds) + if not np.array_equal(ids, np.arange(len(ids))): + shown = ids[:12].tolist() + (["..."] if len(ids) > 12 else []) + raise ValueError( + f"{self.path.name}: {FOLD_KEY} holds the fold ids {shown}; the ids of K folds must be exactly " + f"0..K-1 (--exclude-folds next = fold (k + 1) % K); renumber the column or pass --n-folds K" + ) + return torch.from_numpy(folds), len(ids), FOLD_RULE_COLUMN # ---------------------------------------------------------------- construction def _role_tensor( @@ -264,6 +303,7 @@ def _role_tensor( rows = np.arange(len(present))[:, None] for j in range(role.n): # mark the chosen members as consumed taken[rows[:, 0], idx[:, j]] |= valid[:, j] + self.selected[role.name] = np.where(valid, idx, -1).astype(np.int64) return self._stack(coll, feats, idx, valid) @staticmethod @@ -273,13 +313,13 @@ def _stack( idx: np.ndarray, valid: np.ndarray, ) -> torch.Tensor: - """(N, len(features) + 1, n): features in configuration order, presence flag last, padding -1. - Columns read only for a selector are not stacked.""" + """(N, len(features) + 1, n): features in configuration order, presence flag last, padding -1; + float32 (the columns are cast here). Columns read only for a selector are not stacked.""" n, slots = idx.shape rows = np.arange(n)[:, None] x = np.full((n, len(coll.features) + 1, slots), -1.0, dtype=np.float32) for i, k in enumerate(coll.features): - x[:, i] = np.where(valid, feats[k][rows, idx], -1.0) + x[:, i] = np.where(valid, feats[k][rows, idx], -1.0).astype(np.float32) x[:, -1] = np.where(valid, 1.0, -1.0) return torch.from_numpy(x) @@ -337,7 +377,8 @@ def __getitem__(self, idx): } def fold_indices(self, fold: int) -> tuple[np.ndarray, np.ndarray]: - """(train_idx, val_idx) holding out ``fold``.""" + """(train_idx, val_idx) holding out ``fold``; raises KeyError on a file loaded without folds.""" + self.require_folds(f"holding out fold {fold}") val = (self.folds == fold).numpy() return np.flatnonzero(~val), np.flatnonzero(val) @@ -348,8 +389,10 @@ def fold_indices(self, fold: int) -> tuple[np.ndarray, np.ndarray]: "FOLD_KEY", "FOLD_RULE_COLUMN", "FOLD_RULE_EVENT", + "FOLD_RULE_NONE", "GROUP_ALIASES", "LABEL_KEYS", + "NO_FOLD", "WEIGHT_KEY", "H5Dataset", "columns_to_read", diff --git a/src/corenet/evaluate.py b/src/corenet/evaluate.py index d327fc9..92a5427 100644 --- a/src/corenet/evaluate.py +++ b/src/corenet/evaluate.py @@ -7,7 +7,8 @@ Events: by default the fold held out when the checkpoint was trained (``--fold -1`` = every event of the file), with the checkpoint's fold rule: ``EVENT/fold`` of the file, or fold = squares(event) % n_folds with the checkpoint's seed when the model was trained with ``--n-folds`` (``--n-folds K`` and -``--fold-seed`` override). The output file (layout in +``--fold-seed`` override). A file without ``EVENT/fold`` is evaluated with ``--fold -1`` (or +``--n-folds K``); its events then carry fold -1 in the output. The output file (layout in docs/Dataset.md) holds the class probabilities, the view (candidate) probabilities of every multi-candidate hypothesis, the auxiliary head probabilities, the event labels / weights and, for the main output, the ROC curve; the AUCs are printed and written as JSON next to it (``.json``). @@ -30,7 +31,7 @@ import torch import torch.nn.functional as F -from .data.h5 import H5Dataset +from .data.h5 import FOLD_KEY, H5Dataset from .models.config import NetworkSpec from .models.network import CoReNet @@ -246,6 +247,11 @@ def main(argv=None): args.input, spec, rest_order=rest_order, n_folds=n_folds, fold_seed=fold_seed ) fold = args.fold if args.fold is not None else meta.get("fold", -1) + if fold is not None and fold >= 0 and not ds.has_folds: + raise SystemExit( + f"{args.input}: no {FOLD_KEY} dataset, so fold {fold} cannot be selected: pass --fold -1 to evaluate " + "every event, or --n-folds K to use fold = squares(event) % K from EVENT/event" + ) idx = ( np.arange(len(ds)) if fold is None or fold < 0 diff --git a/src/corenet/models/network.py b/src/corenet/models/network.py index 32b17ba..85cb4d7 100644 --- a/src/corenet/models/network.py +++ b/src/corenet/models/network.py @@ -10,7 +10,9 @@ 2. features log(1+pt), eta, dphi-to-parent, log(1+m) for every child slot; the top node drops phi; event scalars whose name contains ``nSelJets`` enter as log(nSelJets - (min_jets - 1)) - 3. embedding GBN embed (+ event scalars for leaves) -> conv, per role / node + 3. embedding GBN embed (+ event scalars for leaves) -> conv, per role / node; the ``context`` + objects are embedded from their kinematics, a flag and the context role's extra + (non-kinematic) features 4. reinforce bottom-up ResNetBlock per composite node (a node reused by several hypotheses is computed once); ``context`` attention after the first level; ``attention`` entries as soon as both nodes exist, with the pair @@ -115,13 +117,15 @@ def setGhostBatches(self, n): class _Context(nn.Module): - """``context: ``: the objects self-attend, then the first-level candidates attend to them.""" + """``context: ``: the objects self-attend, then the first-level candidates attend to them. + ``n_extra``: the number of extra (non-kinematic) features of the context role, which every object + carries after its flag (``CoReNet._prepare_context``); 0 for a role with kinematic features only.""" - def __init__(self, d, heads, phase_symmetric, gate, kv, device): + def __init__(self, d, heads, phase_symmetric, gate, kv, device, n_extra=0): super().__init__() self.embed = _Embed( - 4, d, phase_symmetric, "attention jet", device - ) # log pt, eta, log m, flag + 4 + n_extra, d, phase_symmetric, "attention jet", device + ) # log pt, eta, log m, flag (+ the context role's extra features) self.pair = _Embed( 3, d, phase_symmetric, "M(a,b), dPhi(a,b)", device ) # log(1+M), dphi, pair-type flag @@ -258,6 +262,7 @@ def __init__(self, spec: NetworkSpec, device="cpu"): tr.attention_gate, tr.kv_projection, device, + n_extra=len(self._extras(hs.context)), ) for h, hs in spec.hypotheses.items() if hs.context @@ -615,15 +620,21 @@ def _prepare_hypothesis( hs = self.spec.hypotheses[h] if hs.context: prep["context"] = self._prepare_context( - h, T, hs.context, leaf_kin, leaf_present, kin, kinP + h, T, hs.context, leaf_kin, leaf_extra, leaf_present, kin, kinP ) return prep - def _prepare_context(self, h, T, crole, leaf_kin, leaf_present, kin, kinP): + def _prepare_context( + self, h, T, crole, leaf_kin, leaf_extra, leaf_present, kin, kinP + ): """Objects = this hypothesis' leaves (jet roles, lepton, MET; flag 2) + the context role (its own flag); pair matrices object-object (flag 0) and first-level-candidate-object (flag 1), with masks. ``crole`` is the context role (e.g. ``other``: the jets not assigned to any decay product); ``o`` (n, 5, S) the - S objects the candidates may look at; ``d`` (n, 4, P) the P candidates of the first-level node.""" + S objects the candidates may look at; ``d`` (n, 4, P) the P candidates of the first-level node. + ``feats`` (n, 4 + E, S) is the embedding input of the objects: log pt, eta, log m, flag, then the E extra + (non-kinematic) features of the context role in its feature order. An object of another leaf carries the + same E channels: an extra its own collection also has is taken from its tensor, one it lacks is -1 (the + padding value); the pair features are kinematic only and unchanged.""" first = self.levels[h][0] # every leaf of the hypothesis (jet roles and single-object collections such as lepton / MET), flag 2 leaf_roles = [ @@ -674,6 +685,25 @@ def _prepare_context(self, h, T, crole, leaf_kin, leaf_present, kin, kinP): torch.isinf(of) | torch.isnan(of), torch.full_like(of, -1.0), of ) of = torch.cat((of[:, :2], of[:, 3:]), 1) # log pt, eta, log m, flag + extras = self._extras(crole) + if extras: # (n, E, S): the context role's extra features for every object, -1 where a leaf lacks one + cols = [] + for r in [*leaf_roles, crole]: + have = self._extras(r) + for e in extras: + cols.append( + leaf_extra[r][:, have.index(e) : have.index(e) + 1, :] + if e in have + else torch.full_like(leaf_kin[r][:, 0:1, :], -1.0) + ) + ex = torch.cat( + [ + torch.cat(cols[i * len(extras) : (i + 1) * len(extras)], 1) + for i in range(len(leaf_roles) + 1) + ], + 2, + ) + of = torch.cat([of, ex], 1) return { "feats": of, "mask": mask, diff --git a/src/corenet/report.py b/src/corenet/report.py new file mode 100644 index 0000000..45d1e53 --- /dev/null +++ b/src/corenet/report.py @@ -0,0 +1,1982 @@ +""" +corenet-report: one self-contained HTML page with the results and the input checks of a run directory. + + corenet-report --run-dir runs/hh4b --input examples/data/hh4b_all.h5 --output runs/hh4b/report.html + +Per fold k the run directory is read for ``history_fold{k}.json``, ``model_fold{k}.pkl``, +``predictions_fold{k}.h5`` (the held-out fold) and, when present, ``eval_fold{k}.h5`` written by +``corenet-evaluate`` (with ``--fold -1`` it holds every event of the file, which gives the training and +validation folds their own curves). The input file provides the events, labels, weights, folds and, when +present, the generator-level ``TARGETS``. + +Sections: run summary, training curves, classification (ROC, AUC, scores, confusion matrix, auxiliary +heads), hypotheses and pairing (view probabilities, pairing accuracy against ``TARGETS``), the inputs +as the network embeds them, the raw inputs, the event weights and further checks. Every figure is embedded as PNG; the page needs no other file. +""" + +from __future__ import annotations + +import argparse +import base64 +import datetime +import html +import io +import json +import logging +import re +import subprocess +import sys +import traceback +from dataclasses import dataclass, field +from pathlib import Path + +import h5py +import matplotlib +import numpy as np +import torch +import yaml + +matplotlib.use("Agg") +import matplotlib.pyplot as plt + +from . import __version__ +from .checkpoint import fold_rule_of, load_checkpoint +from .data.h5 import ( + FOLD_RULE_EVENT, + H5Dataset, + event_source, + resolve_group, +) +from .models.config import NetworkSpec, parse_config, read_config_dict +from .models.network import CoReNet + +SECTIONS = [ + ("summary", "Run summary"), + ("training", "Training curves"), + ("classification", "Classification"), + ("hypotheses", "Hypotheses and pairing"), + ("inputs", "Inputs as the network sees them"), + ("raw", "Raw inputs"), + ("weights", "Weights"), + ("other", "Other checks"), +] +SPLITS = ("test", "validation", "train") +SPLIT_STYLE = {"test": "-", "train": "--", "validation": "-."} +ROC_GRID = np.unique(np.concatenate([np.logspace(-4, 0, 300), np.linspace(0, 1, 300)])) +VIEW_THRESHOLDS = (0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95) +NBINS = 40 + +plt.rcParams.update( + { + "font.size": 8, + "axes.labelsize": 9, + "axes.titlesize": 9, + "legend.fontsize": 7, + "figure.dpi": 100, + "axes.spines.top": False, + "axes.spines.right": False, + } +) + + +# ----------------------------------------------------------------------------- small helpers +def esc(x) -> str: + return html.escape(str(x)) + + +def fmt(x, nd: int = 4) -> str: + """A number for a table cell; '-' for None / NaN.""" + if x is None: + return "-" + if isinstance(x, (bool, np.bool_)): + return str(bool(x)) + if isinstance(x, (int, np.integer)): + return f"{int(x):,}" + try: + xf = float(x) + except (TypeError, ValueError): + return esc(x) + if not np.isfinite(xf): + return "-" + return f"{xf:.{nd}f}" + + +def mean_sd(values) -> str: + """'mean ± sd' over the finite entries (sd with ddof=1; '-' for fewer than two).""" + v = np.array([float(x) for x in values if x is not None], dtype=float) + v = v[np.isfinite(v)] + if len(v) == 0: + return "-" + if len(v) == 1: + return f"{v[0]:.4f}" + return f"{v.mean():.4f} ± {v.std(ddof=1):.4f}" + + +def table(header: list[str], rows: list[list], caption: str = "") -> str: + head = "".join(f"{esc(h)}" for h in header) + body = "".join( + "" + + "".join(f"{c if isinstance(c, str) else fmt(c)}" for c in r) + + "" + for r in rows + ) + cap = f"{esc(caption)}" if caption else "" + return f"{cap}{head}{body}
" + + +def figure(fig, caption: str = "", dpi: int = 100) -> str: + """The figure as an inline PNG (the figure is closed).""" + buf = io.BytesIO() + try: + fig.tight_layout() + except Exception: # a figure that cannot be laid out is still saved + pass + fig.savefig(buf, format="png", dpi=dpi) + plt.close(fig) + data = base64.b64encode(buf.getvalue()).decode("ascii") + return ( + f'
{esc(caption)}' + f"
{esc(caption)}
" + ) + + +def note(text: str) -> str: + return f'

{esc(text)}

' + + +def class_color(i: int) -> str: + return "0.45" if i == 0 else f"C{(i - 1) % 10}" + + +def shared_bins(arrays, n: int = NBINS) -> np.ndarray: + """Bin edges over the 0.1-99.9 percentiles of every array; unit bins for integer-valued data.""" + vals = [np.asarray(a, dtype=float) for a in arrays if a is not None and len(a)] + vals = [v[np.isfinite(v)] for v in vals] + vals = [v for v in vals if len(v)] + if not vals: + return np.linspace(0, 1, n + 1) + allv = np.concatenate(vals) + lo, hi = np.percentile(allv, [0.1, 99.9]) + if np.all(np.mod(allv, 1) == 0) and hi - lo <= 60: + return np.arange(np.floor(lo) - 0.5, np.ceil(hi) + 1.5, 1.0) + if lo == hi: + lo, hi = lo - 1, hi + 1 + return np.linspace(lo, hi, n + 1) + + +def hist_panel(ax, series, bins, xlabel: str, logy: bool = False) -> None: + """``series``: (label, values, weights, line style dict); weighted densities as step histograms.""" + for lab, v, w, st in series: + v, w = np.asarray(v, dtype=float), np.asarray(w, dtype=float) + ok = np.isfinite(v) & np.isfinite(w) + if not ok.any() or w[ok].sum() <= 0: + continue + ax.hist( + v[ok], + bins=bins, + weights=w[ok], + density=True, + histtype="step", + label=lab, + **st, + ) + ax.set_xlabel(xlabel) + ax.set_ylabel("density") + if logy: + ax.set_yscale("log") + + +def grid_figure(n_panels: int, per_row: int = 4, w: float = 2.6, h: float = 2.1): + rows = max(1, int(np.ceil(n_panels / per_row))) + cols = min(per_row, max(1, n_panels)) + fig, axes = plt.subplots(rows, cols, figsize=(w * cols, h * rows), squeeze=False) + flat = axes.ravel() + for ax in flat[n_panels:]: + ax.axis("off") + return fig, flat + + +def mean_curve(curves, grid=ROC_GRID): + """(mean, sd) of the TPR of several (fpr, tpr) curves on ``grid`` (sd None for one curve).""" + tprs = np.array([np.interp(grid, f, t) for f, t in curves]) + return tprs.mean(0), (tprs.std(0, ddof=1) if len(tprs) > 1 else None) + + +def roc_points(y: np.ndarray, s: np.ndarray, w: np.ndarray): + from sklearn.metrics import roc_curve + + f, t, _ = roc_curve(y, s, sample_weight=w) + return f, t + + +def auc_of(y: np.ndarray, s: np.ndarray, w: np.ndarray): + from sklearn.metrics import roc_auc_score + + if y.all() or not y.any() or w.sum() <= 0: + return None + return float(roc_auc_score(y, s, sample_weight=w)) + + +def norm_name(s: str) -> str: + return re.sub(r"[^a-z0-9]", "", s.lower()) + + +# ----------------------------------------------------------------------------- run files +@dataclass +class Split: + index: np.ndarray # rows of the input file + label: np.ndarray + weight: np.ndarray + probs: np.ndarray + views: dict[str, np.ndarray] + heads: dict[str, np.ndarray] + source: str + + +@dataclass +class FoldData: + fold: int + history: dict | None = None + checkpoint: Path | None = None + checkpoint_meta: dict = field(default_factory=dict) + splits: dict[str, Split] = field(default_factory=dict) + n_parameters: int | None = None + + @property + def exclude_folds(self) -> list[int] | None: + for src in (self.history, self.checkpoint_meta): + if src and src.get("exclude_folds") is not None: + return [int(e) for e in src["exclude_folds"]] + return None + + @property + def best_epoch(self) -> int | None: + if not self.history: + return None + if self.history.get("best_epoch") is not None: + return int(self.history["best_epoch"]) + aucs = [(e.get("val_auc"), e["epoch"]) for e in self.history.get("history", [])] + aucs = [(a, ep) for a, ep in aucs if a is not None and np.isfinite(a)] + return max(aucs)[1] if aucs else None + + +def read_prediction_file(path: Path): + """(fold attribute, Split without a source) of a prediction file.""" + with h5py.File(path, "r") as f: + views = {h: f["views"][h][...] for h in f["views"]} if "views" in f else {} + heads = {k: f["heads"][k][...] for k in f["heads"]} if "heads" in f else {} + split = Split( + f["EVENT/index"][...].astype(np.int64), + f["EVENT/label"][...].astype(np.int64), + f["EVENT/weight"][...].astype(np.float64), + f["class_probabilities"][...].astype(np.float64), + views, + heads, + path.name, + ) + fold = int(f.attrs.get("fold", -1)) + n_par = f.attrs.get("n_parameters") + return fold, split, (int(n_par) if n_par is not None else None) + + +def sub_split(s: Split, sel: np.ndarray, source: str) -> Split: + return Split( + s.index[sel], + s.label[sel], + s.weight[sel], + s.probs[sel], + {k: v[sel] for k, v in s.views.items()}, + {k: v[sel] for k, v in s.heads.items()}, + source, + ) + + +def discover_folds(run_dir: Path, wanted: list[int] | None) -> list[int]: + folds = set() + for pattern in ( + "history_fold*.json", + "model_fold*.pkl", + "predictions_fold*.h5", + "eval_fold*.h5", + ): + for p in run_dir.glob(pattern): + m = re.fullmatch( + r"(?:history|model|predictions|eval)_fold(\d+)\.(?:json|pkl|h5)", p.name + ) + if m: + folds.add(int(m.group(1))) + if wanted is not None: + return sorted(wanted) + return sorted(folds) + + +# ----------------------------------------------------------------------------- the report +class Report: + def __init__(self, args): + self.args = args + self.run_dir = Path(args.run_dir) + self.input = Path(args.input) + self.gaps: list[str] = [] + self.generated = datetime.datetime.now().strftime("%Y-%m-%d %H:%M") + self.folds: list[FoldData] = [] + self.config: dict | None = None + self.config_source = "" + self.spec: NetworkSpec | None = None + self.model: CoReNet | None = None + self.ds: H5Dataset | None = None + self.rng = np.random.default_rng(0) + + # ---------------------------------------------------------------- loading + def load(self) -> None: + a = self.args + if not self.run_dir.is_dir(): + raise SystemExit(f"{self.run_dir}: not a directory") + fold_ids = discover_folds(self.run_dir, a.folds) + if not fold_ids: + raise SystemExit( + f"{self.run_dir}: no history_fold*.json, model_fold*.pkl or predictions_fold*.h5" + ) + for k in fold_ids: + fd = FoldData(k) + hp = self.run_dir / f"history_fold{k}.json" + if hp.is_file(): + fd.history = json.loads(hp.read_text()) + cp = self.run_dir / f"model_fold{k}.pkl" + if cp.is_file(): + fd.checkpoint = cp + self.folds.append(fd) + + # configuration and network: --config, else the first checkpoint, else the history's config path + first_ckpt = next((fd for fd in self.folds if fd.checkpoint), None) + if first_ckpt is not None: + self.model, self.spec, ckpt = load_checkpoint( + first_ckpt.checkpoint, a.config, "cpu" + ) + self.config = read_config_dict(a.config) if a.config else ckpt["config"] + self.config_source = ( + a.config or f"{first_ckpt.checkpoint.name} (stored configuration)" + ) + else: + cfg_path = a.config + if cfg_path is None: + for fd in self.folds: + if ( + fd.history + and fd.history.get("config") + and Path(fd.history["config"]).is_file() + ): + cfg_path = fd.history["config"] + break + if cfg_path is None: + raise SystemExit( + f"{self.run_dir}: no checkpoint and no readable configuration; pass --config" + ) + self.config = read_config_dict(cfg_path) + self.spec = parse_config(self.config) + self.model = CoReNet(self.spec, device="cpu") + self.model.eval() + self.config_source = str(cfg_path) + self.gaps.append( + "no checkpoint in the run directory: the network was built from the configuration without trained weights (parameter count only)" + ) + for fd in self.folds: + if fd.checkpoint: + meta = torch.load( + fd.checkpoint, map_location="cpu", weights_only=False + ).get("metadata", {}) + fd.checkpoint_meta = dict(meta) + + # the dataset, with the fold rule of the run + meta = first_ckpt.checkpoint_meta if first_ckpt else {} + hist = next((fd.history for fd in self.folds if fd.history), None) or {} + n_folds, fold_seed = fold_rule_of(meta) if meta else (None, 0) + if not meta and hist.get("fold_rule") == FOLD_RULE_EVENT: + n_folds, fold_seed = hist.get("n_folds"), int(hist.get("fold_seed", 0)) + if a.n_folds is not None: + n_folds = a.n_folds + if a.fold_seed is not None: + fold_seed = a.fold_seed + rest_order = ( + a.rest_order + or meta.get("rest_order") + or (hist.get("recipe") or {}).get("rest_order", "pt") + ) + self.ds = H5Dataset( + self.input, + self.spec, + rest_order=rest_order, + n_folds=n_folds, + fold_seed=fold_seed, + ) + self.labels = self.ds.labels.numpy() + self.weights = self.ds.weights.numpy().astype(np.float64) + self.fold_of = self.ds.folds.numpy() + logging.info( + f"{self.input}: {len(self.ds):,} events, folds {self.ds.fold_rule}, classes {self.spec.classes}" + ) + + # predictions: predictions_fold{k}.h5 (held-out fold) and eval_fold{k}.h5 (any fold, or every event) + for fd in self.folds: + k = fd.fold + for name in (f"predictions_fold{k}.h5", f"eval_fold{k}.h5"): + p = self.run_dir / name + if not p.is_file(): + continue + fold_attr, split, n_par = read_prediction_file(p) + if n_par is not None and fd.n_parameters is None: + fd.n_parameters = n_par + if split.index.max(initial=-1) >= len(self.ds): + self.gaps.append( + f"{name}: event indices exceed the input file ({len(self.ds):,} events); the file does not belong to this input" + ) + continue + if fold_attr == k: + fd.splits.setdefault( + "test", sub_split(split, np.ones(len(split.index), bool), name) + ) + continue + if fold_attr >= 0: + self.gaps.append( + f"{name}: evaluated fold {fold_attr}, not the held-out fold {k} nor every event; ignored" + ) + continue + folds_of_rows = self.fold_of[split.index] + test = folds_of_rows == k + val = np.isin(folds_of_rows, fd.exclude_folds or []) + train = ~test & ~val + fd.splits.setdefault("test", sub_split(split, test, name)) + if fd.exclude_folds: + if val.any(): + fd.splits["validation"] = sub_split(split, val, name) + else: + self.gaps.append( + f"fold {k}: the validation folds are unknown (no history / checkpoint metadata); every event outside fold {k} in {name} counts as training" + ) + if train.any(): + fd.splits["train"] = sub_split(split, train, name) + if "test" not in fd.splits: + self.gaps.append( + f"fold {k}: no predictions_fold{k}.h5 or eval_fold{k}.h5: no ROC, scores or view probabilities for this fold" + ) + if "train" not in fd.splits: + self.gaps.append( + f"fold {k}: no predictions on the training folds (run corenet-evaluate with --fold -1 and keep eval_fold{k}.h5 in the run directory): no overtraining check for this fold" + ) + if fd.history is None: + self.gaps.append( + f"fold {k}: no history_fold{k}.json: no training curves for this fold" + ) + if fd.n_parameters is None and self.model is not None: + fd.n_parameters = int(sum(p.numel() for p in self.model.parameters())) + + # ---------------------------------------------------------------- shared data + @property + def classes(self) -> list[str]: + return self.spec.classes + + def class_series(self, values, labels, weights, style=None): + """(class name, values, clipped weights, style) per class present in ``labels``.""" + out = [] + for i, c in enumerate(self.classes): + sel = labels == i + if not sel.any(): + continue + st = {"color": class_color(i), "lw": 1.3} + st.update(style or {}) + out.append((c, values[sel], np.clip(weights[sel], 0, None), st)) + return out + + def splits_of(self, name: str): + return [(fd, fd.splits[name]) for fd in self.folds if name in fd.splits] + + def selected_subset(self, n_max: int) -> np.ndarray: + n = len(self.ds) + if n_max and n > n_max: + return np.sort(self.rng.choice(n, n_max, replace=False)) + return np.arange(n) + + # ---------------------------------------------------------------- sections + def section_summary(self) -> str: + parts = [] + rows = [ + ["input file", esc(self.input)], + ["events", f"{len(self.ds):,}"], + [ + "classes", + esc(", ".join(f"{i} = {c}" for i, c in enumerate(self.classes))), + ], + ["folds", f"{self.ds.n_folds} ({esc(self.ds.fold_rule)})"], + ["configuration", esc(self.config_source)], + ["run directory", esc(self.run_dir)], + [ + "package", + esc( + f"corenet {__version__}" + + (f", git {git_hash()}" if git_hash() else "") + ), + ], + ["generated", esc(self.generated)], + ] + parts.append(table(["", ""], rows, "Run")) + cfg_text = yaml.safe_dump(self.config, sort_keys=False, allow_unicode=True) + parts.append( + f"
Configuration ({esc(self.config_source)})
{esc(cfg_text)}
" + ) + + # events per class + rows = [] + for i, c in enumerate(self.classes): + sel = self.labels == i + rows.append( + [ + c, + int(sel.sum()), + float(self.weights[sel].sum()), + int((self.weights[sel] < 0).sum()), + ] + ) + parts.append( + table( + ["class", "events", "sum of weights", "negative weights"], + rows, + "Events per class in the input file", + ) + ) + + # recipe + hist = next((fd.history for fd in self.folds if fd.history), None) + meta = next( + (fd.checkpoint_meta for fd in self.folds if fd.checkpoint_meta), None + ) + recipe = (hist or {}).get("recipe") or { + k: v + for k, v in (meta or {}).items() + if k + not in ( + "fold", + "epoch", + "exclude_folds", + "input", + "n_folds", + "fold_rule", + "fold_seed", + "finetune_trainable_parameters", + ) + } + if recipe: + parts.append( + table( + ["setting", "value"], + [[k, esc(v)] for k, v in recipe.items()], + "Training recipe", + ) + ) + else: + self.gaps.append( + "training recipe: no history_fold*.json or model_fold*.pkl" + ) + + # per-fold table + rows = [] + wall_recorded = False + for fd in self.folds: + k = fd.fold + h = fd.history or {} + entries = h.get("history", []) + n_test = int((self.fold_of == k).sum()) + n_train = h.get("n_train") + n_val = h.get("n_val") + if n_train is None: + excl = fd.exclude_folds or [] + n_val = int(np.isin(self.fold_of, excl).sum()) + n_train = len(self.ds) - n_test - n_val + wall = None + for key in ("wall_time_s", "seconds", "time"): + if entries and key in entries[-1]: + wall = entries[-1][key] + wall_recorded = True + rows.append( + [ + k, + esc(fd.exclude_folds if fd.exclude_folds is not None else "-"), + n_train, + n_val, + n_test, + len(entries) if entries else "-", + fmt(fd.best_epoch) if fd.best_epoch is not None else "-", + fmt(h.get("stopped_after_epoch")) + if h.get("stopped_after_epoch") is not None + else "-", + fd.n_parameters if fd.n_parameters is not None else "-", + fmt(h.get("finetune_trainable_parameters")) + if h.get("finetune_trainable_parameters") is not None + else "-", + fmt(wall, 0) if wall is not None else "-", + esc(", ".join(sorted(fd.splits))) or "-", + ] + ) + header = [ + "fold", + "validation folds", + "training events", + "validation events", + "test events", + "epochs run", + "best epoch", + "stopped after", + "parameters", + "fine-tuned parameters", + "wall time [s]", + "predictions", + ] + if ( + not wall_recorded + ): # the column exists only when the history files carry a wall time + header.pop(10) + rows = [r[:10] + r[11:] for r in rows] + parts.append(table(header, rows, "Folds")) + if any( + fd.history and fd.history.get("early_stop_patience") for fd in self.folds + ): + parts.append( + note( + "Best epoch: the epoch with the highest validation AUC; its weights are the ones in the checkpoint (early stopping). 'Stopped after' is the epoch at which training stopped, '-' when every epoch was run." + ) + ) + else: + parts.append( + note( + "Best epoch: the epoch with the highest validation AUC in the history; without early stopping the checkpoint holds the weights after the last epoch." + ) + ) + return "".join(parts) + + def section_training(self) -> str: + with_hist = [ + fd for fd in self.folds if fd.history and fd.history.get("history") + ] + if not with_hist: + self.gaps.append("training curves: no history_fold*.json with epochs") + return "" + keys = [ + ("train_loss", "training loss"), + ("val_loss", "validation loss"), + ("val_auc", "validation AUC"), + ] + fig, axes = plt.subplots(1, 3, figsize=(9.6, 2.9)) + any_val = False + for i, (key, lab) in enumerate(keys): + ax = axes[i] + for j, fd in enumerate(with_hist): + entries = fd.history["history"] + ep = [e["epoch"] for e in entries if e.get(key) is not None] + val = [e[key] for e in entries if e.get(key) is not None] + if not ep: + continue + any_val |= key != "train_loss" + ax.plot( + ep, val, "-", color=f"C{j % 10}", lw=1.2, label=f"fold {fd.fold}" + ) + best = fd.best_epoch + if best is not None and best in ep: + ax.plot( + [best], [val[ep.index(best)]], "*", color=f"C{j % 10}", ms=9 + ) + ax.set_xlabel("epoch") + ax.set_ylabel(lab) + ax.grid(alpha=0.25, lw=0.5) + axes[0].legend(frameon=False) + out = figure( + fig, + "Training loss, validation loss and validation AUC per epoch, one line per fold; the star marks the best epoch (highest validation AUC).", + ) + if not any_val: + self.gaps.append( + "validation curves: the histories have no validation entries (train with --exclude-folds)" + ) + rows = [] + for fd in with_hist: + e = fd.history["history"][-1] + rows.append( + [ + fd.fold, + e["epoch"], + e.get("batch_size"), + fmt(e.get("lr"), 6), + e.get("train_loss"), + e.get("val_loss"), + e.get("val_auc"), + ] + ) + out += table( + [ + "fold", + "last epoch", + "batch size", + "learning rate", + "training loss", + "validation loss", + "validation AUC", + ], + rows, + "Last epoch of every fold", + ) + return out + + # ---- classification + def _main_curves(self): + """{split: [(fold, fpr, tpr, auc)]} of the main score (class 1 against the rest).""" + out = {} + for name in SPLITS: + for fd, s in self.splits_of(name): + y, sc, w = s.label == 1, s.probs[:, 1], np.clip(s.weight, 0, None) + if y.all() or not y.any(): + continue + f, t = roc_points(y, sc, w) + out.setdefault(name, []).append((fd.fold, f, t, auc_of(y, sc, w))) + return out + + def section_classification(self) -> str: + if not any("test" in fd.splits for fd in self.folds): + self.gaps.append("classification: no prediction file for any fold") + return "" + parts = [] + curves = self._main_curves() + if not curves: + self.gaps.append( + "ROC curves: the predicted events have only one class (no signal or no background)" + ) + else: + fig, (a, b) = plt.subplots(1, 2, figsize=(7.8, 3.4)) + colors = {"test": "C0", "train": "C3", "validation": "C2"} + for name in SPLITS: + if name not in curves: + continue + cs = curves[name] + mu, sd = mean_curve([(f, t) for _, f, t, _ in cs]) + aucs = [c[3] for c in cs] + lab = f"{name} ({len(cs)} fold{'s' if len(cs) > 1 else ''}) AUC {mean_sd(aucs)}" + for ax in (a, b): + if name == "test": + for _, f, t, _ in cs: + ax.plot(f, t, color=colors[name], lw=0.6, alpha=0.35) + ax.plot( + ROC_GRID, + mu, + color=colors[name], + ls=SPLIT_STYLE[name], + lw=1.5, + label=lab if ax is a else None, + ) + if sd is not None: + ax.fill_between( + ROC_GRID, + mu - sd, + mu + sd, + color=colors[name], + alpha=0.2, + lw=0, + ) + a.plot([0, 1], [0, 1], "k:", lw=0.8) + a.set_xlabel("background efficiency") + a.set_ylabel("signal efficiency") + a.set_xlim(0, 1) + a.set_ylim(0, 1.02) + a.legend(loc="lower right", frameon=False) + a.set_title("(a)", loc="left") + b.set_xscale("log") + b.set_xlim(1e-4, 1) + b.set_ylim(0, 1.02) + b.set_xlabel("background efficiency (log)") + b.set_title("(b)", loc="left") + b.grid(alpha=0.25, which="both", lw=0.5) + parts.append( + figure( + fig, + f"ROC of the main score P({self.classes[1]}) (class 1 against every other class, weighted): thin lines are the test folds, thick lines the mean over folds with the ± 1 sd band; dashed = training folds, dash-dotted = validation fold, where predictions exist.", + ) + ) + # AUC table + rows = [] + folds = sorted({c[0] for cs in curves.values() for c in cs}) + for k in folds: + row = [k] + for name in SPLITS: + v = next((c[3] for c in curves.get(name, []) if c[0] == k), None) + row.append(fmt(v) if v is not None else "-") + rows.append(row) + rows.append( + ["mean ± sd"] + + [mean_sd([c[3] for c in curves.get(name, [])]) for name in SPLITS] + ) + parts.append( + table( + ["fold", "test", "validation", "train"], + rows, + "AUC of the main score (sd over folds, ddof = 1)", + ) + ) + for name in ("validation", "train"): + if name not in curves: + self.gaps.append( + f"ROC curves: no predictions on the {name} folds (corenet-evaluate --fold -1 provides them)" + ) + + # score distributions + parts.append(self._score_figure()) + # multi-class + if len(self.classes) > 2: + parts.append(self._confusion_figure()) + parts.append(self._ovr_and_heads_figure()) + elif any(s.heads for fd in self.folds for s in fd.splits.values()): + parts.append(self._ovr_and_heads_figure()) + return "".join(parts) + + def _score_figure(self) -> str: + test = self.splits_of("test") + if not test: + return "" + bins = np.linspace(0, 1, 41) + fig, (a, b) = plt.subplots(1, 2, figsize=(7.8, 3.2)) + for name in ("test", "train"): + sp = self.splits_of(name) + if not sp: + continue + lab = np.concatenate([s.label for _, s in sp]) + sc = np.concatenate([s.probs[:, 1] for _, s in sp]) + w = np.concatenate([s.weight for _, s in sp]) + series = self.class_series(sc, lab, w, {"ls": SPLIT_STYLE[name]}) + series = [(f"{c} ({name})", v, ww, st) for c, v, ww, st in series] + hist_panel(a, series, bins, f"P({self.classes[1]})") + hist_panel(b, series, bins, f"P({self.classes[1]})", logy=True) + a.legend(frameon=False) + a.set_title("(a) linear", loc="left") + b.set_title("(b) log", loc="left") + cap = "Main score per class on the test folds (solid), weighted densities" + if self.splits_of("train"): + cap += "; dashed: the same on the training folds (overtraining check)" + return figure(fig, cap + ".") + + def _confusion_figure(self) -> str: + test = self.splits_of("test") + n = len(self.classes) + cm = np.zeros((n, n)) + for _, s in test: + pred = s.probs.argmax(1) + w = np.clip(s.weight, 0, None) + for i in range(n): + for j in range(n): + cm[i, j] += w[(s.label == i) & (pred == j)].sum() + present = [i for i in range(n) if cm[i].sum() > 0] + if not present: + return "" + norm = cm / np.where(cm.sum(1, keepdims=True) > 0, cm.sum(1, keepdims=True), 1) + fig, ax = plt.subplots(figsize=(1.2 * n + 2.2, 1.0 * n + 1.6)) + im = ax.imshow(norm[present], cmap="Blues", vmin=0, vmax=1) + for a, i in enumerate(present): + for j in range(n): + ax.text( + j, + a, + f"{norm[i, j]:.3f}", + ha="center", + va="center", + color="white" if norm[i, j] > 0.6 else "black", + fontsize=8, + ) + ax.set_xticks(range(n)) + ax.set_xticklabels(self.classes) + ax.set_yticks(range(len(present))) + ax.set_yticklabels([self.classes[i] for i in present]) + ax.set_xlabel("predicted class (argmax)") + ax.set_ylabel("true class") + fig.colorbar(im, ax=ax, fraction=0.046) + return figure( + fig, + "Confusion matrix at the argmax of the class probabilities on the test folds, weighted, each true class normalised to 1.", + ) + + def _ovr_and_heads_figure(self) -> str: + test = self.splits_of("test") + parts = [] + has_heads = any(s.heads for _, s in test) + fig, axes = plt.subplots( + 1, + 2 if has_heads else 1, + figsize=(7.8 if has_heads else 4.0, 3.4), + squeeze=False, + ) + ax = axes[0, 0] + rows = [] + for i, c in enumerate(self.classes): + cs = [ + ( + fd.fold, + *roc_points( + s.label == i, s.probs[:, i], np.clip(s.weight, 0, None) + ), + auc_of(s.label == i, s.probs[:, i], np.clip(s.weight, 0, None)), + ) + for fd, s in test + if (s.label == i).any() and not (s.label == i).all() + ] + if not cs: + continue + mu, sd = mean_curve([(f, t) for _, f, t, _ in cs]) + ax.plot( + ROC_GRID, + mu, + color=class_color(i), + lw=1.4, + label=f"{c} vs rest AUC {mean_sd([x[3] for x in cs])}", + ) + if sd is not None: + ax.fill_between( + ROC_GRID, mu - sd, mu + sd, color=class_color(i), alpha=0.2, lw=0 + ) + rows.append([f"P({c}) vs label = {c}", mean_sd([x[3] for x in cs])]) + ax.plot([0, 1], [0, 1], "k:", lw=0.8) + ax.set_xlabel("efficiency of the other classes") + ax.set_ylabel("efficiency of the class") + ax.set_title("(a) one class against the rest", loc="left") + ax.legend(loc="lower right", frameon=False) + if has_heads: + ax = axes[0, 1] + for name in sorted({k for _, s in test for k in s.heads}): + i = self.classes.index(name) if name in self.classes else None + if i is None: + continue + cs = [ + ( + fd.fold, + *roc_points( + s.label == i, s.heads[name], np.clip(s.weight, 0, None) + ), + auc_of(s.label == i, s.heads[name], np.clip(s.weight, 0, None)), + ) + for fd, s in test + if name in s.heads + and (s.label == i).any() + and not (s.label == i).all() + ] + if not cs: + continue + mu, sd = mean_curve([(f, t) for _, f, t, _ in cs]) + ax.plot( + ROC_GRID, + mu, + color=class_color(i), + lw=1.4, + label=f"head {name} AUC {mean_sd([x[3] for x in cs])}", + ) + if sd is not None: + ax.fill_between( + ROC_GRID, + mu - sd, + mu + sd, + color=class_color(i), + alpha=0.2, + lw=0, + ) + rows.append( + [f"head {name} vs label = {name}", mean_sd([x[3] for x in cs])] + ) + ax.plot([0, 1], [0, 1], "k:", lw=0.8) + ax.set_xlabel("efficiency of the other classes") + ax.set_ylabel("efficiency of the class") + ax.set_title("(b) auxiliary heads", loc="left") + ax.legend(loc="lower right", frameon=False) + parts.append( + figure( + fig, + "Test folds, mean over folds with the ± 1 sd band: (a) every class scored by its own probability against all other classes" + + ( + "; (b) the auxiliary head of every further output against its class." + if has_heads + else "." + ), + ) + ) + parts.append( + table( + ["score", "AUC (mean ± sd over test folds)"], + rows, + "One-against-the-rest and auxiliary-head AUC", + ) + ) + return "".join(parts) + + # ---- hypotheses and pairing + def section_hypotheses(self) -> str: + test = self.splits_of("test") + hyps = sorted({h for _, s in test for h in s.views}) + parts = [] + if not hyps: + if test: + self.gaps.append( + "hypotheses: every hypothesis has a single candidate (no view probabilities)" + ) + else: + self.gaps.append("hypotheses: no prediction file") + return "".join(parts) + fig, axes = grid_figure(len(hyps), per_row=3, w=3.2, h=2.6) + for ax, h in zip(axes, hyps): + V = next(s.views[h].shape[1] for _, s in test if h in s.views) + bins = np.linspace(1.0 / V, 1.0, 31) + for name in ("test", "train"): + sp = [(fd, s) for fd, s in self.splits_of(name) if h in s.views] + if not sp: + continue + mx = np.concatenate([s.views[h].max(1) for _, s in sp]) + lab = np.concatenate([s.label for _, s in sp]) + w = np.concatenate([s.weight for _, s in sp]) + series = [ + (f"{c} ({name})", v, ww, st) + for c, v, ww, st in self.class_series( + mx, lab, w, {"ls": SPLIT_STYLE[name]} + ) + ] + hist_panel(ax, series, bins, f"max view probability, {h} ({V} views)") + ax.legend(frameon=False) + parts.append( + figure( + fig, + "Highest view probability per event and class on the test folds (dashed: training folds when available). The views of a hypothesis are its candidate assignments, e.g. the pairings of the b jets into two Higgs bosons.", + ) + ) + # argmax view shares + rows = [] + for h in hyps: + V = next(s.views[h].shape[1] for _, s in test if h in s.views) + names = self._view_names(h, V) + for i, c in enumerate(self.classes): + counts = np.zeros(V) + for _, s in test: + if h not in s.views: + continue + sel = s.label == i + if not sel.any(): + continue + am = s.views[h][sel].argmax(1) + w = np.clip(s.weight[sel], 0, None) + counts += np.bincount(am, weights=w, minlength=V)[:V] + if counts.sum() > 0: + rows.append([h, c] + [fmt(x) for x in counts / counts.sum()]) + if rows: + parts.append( + table( + ["hypothesis", "class", *names], + [r for r in rows if r[0] == h], + f"Weighted share of the argmax view, test folds: {h}", + ) + ) + parts.append(self._pairing_accuracy(test, hyps)) + return "".join(parts) + + def _view_names(self, h: str, V: int) -> list[str]: + T = self.model.tables.get(h) + if T is None or len(T.views) != V: + return [f"view {v}" for v in range(V)] + names = [] + for view in T.views: + names.append( + "|".join("".join(str(slot) for _, slot in child) for child in view) + ) + return names + + def _targets(self) -> dict[str, dict[str, np.ndarray]]: + out = {} + with h5py.File(self.input, "r") as f: + if "TARGETS" not in f: + return out + for g in f["TARGETS"]: + grp = f["TARGETS"][g] + if not isinstance(grp, h5py.Group): + continue + leaves = { + k: grp[k][...].astype(np.int64) + for k in grp + if isinstance(grp[k], h5py.Dataset) + and grp[k].dtype.kind in "iu" + and grp[k].ndim == 1 + and len(grp[k]) == len(self.ds) + } + if leaves: + out[g] = leaves + return out + + def _view_assignments(self, h: str, v: int) -> dict[str, tuple]: + """{declared node: assignment (tuples of (role, slot))} of view ``v`` of hypothesis ``h``.""" + T = self.model.tables[h] + out = {} + + def rec(declared, canonical, cand): + nt = T.nodes[canonical] + out[declared] = nt.candidates[cand] + for j, c in enumerate(self.spec.nodes[declared].children): + if c in self.model.leaves: + continue + rec(c, T.aliases.get(c, c), int(nt.child_index[cand, j])) + + rec(h, h, int(T.view_index[v])) + return out + + def _pairing_accuracy(self, test, hyps) -> str: + targets = self._targets() + if not targets: + self.gaps.append( + "pairing accuracy: the input file has no TARGETS group (generator-matched jets)" + ) + return "" + parts = [] + sel = self.ds.selected + for h in hyps: + V = next(s.views[h].shape[1] for _, s in test if h in s.views) + T = self.model.tables.get(h) + if T is None or len(T.views) != V: + continue + declared = self.spec.nodes_in(h) + matched = { + g: n for g in targets for n in declared if norm_name(g) == norm_name(n) + } + if not matched: + self.gaps.append( + f"pairing accuracy of {h}: no TARGETS group named like one of its nodes ({sorted(declared)}; TARGETS has {sorted(targets)})" + ) + continue + # per matched node: the sorted file slots of its generator-matched jets (truth) and of its + # candidate under every view; sibling nodes that are interchangeable (H1 / H2) are compared as + # a set, so that swapping them does not count as a wrong assignment + n = len(self.ds) + correct = np.ones((V, n), bool) + complete = np.ones(n, bool) + contained = np.ones(n, bool) + usable = [] + truth_of, views_of = {}, {} + for g, node in matched.items(): + truth = np.stack(list(targets[g].values()), 1) # (n, k) file slots + assignments = [self._view_assignments(h, v).get(node) for v in range(V)] + pairs = [ + [(r, s) for r, s in a if r in sel] if a is not None else None + for a in assignments + ] + if any(p is None or len(p) != truth.shape[1] for p in pairs): + continue # the group lists a different number of jets than the node has candidates + usable.append((g, node)) + complete &= (truth >= 0).all(1) + roles = sorted({r for p in pairs for r, _ in p}) + pool = np.concatenate( + [sel[r] for r in roles], 1 + ) # (n, slots) file slots of the roles + for col in range(truth.shape[1]): + contained &= (pool == truth[:, col : col + 1]).any(1) + truth_of[node] = np.sort(truth, 1) + views_of[node] = [ + np.sort(np.stack([sel[r][:, slot] for r, slot in p], 1), 1) + for p in pairs + ] + # groups of interchangeable siblings that are all matched + grouped, groups = set(), [] + for parent in declared: + nt = T.nodes.get(T.aliases.get(parent, parent)) + kids = ( + self.spec.nodes[parent].children + if parent in self.spec.nodes + else [] + ) + if ( + nt is not None + and nt.symmetric_children + and len(kids) > 1 + and all(k in truth_of for k in kids) + ): + groups.append(list(kids)) + grouped.update(kids) + + def key(slots): # one integer per sorted slot tuple + base = 1 << 16 + return sum( + (slots[:, j].astype(np.int64) + 1) * base**j + for j in range(slots.shape[1]) + ) + + for node in truth_of: + if node in grouped: + continue + for v in range(V): + correct[v] &= (views_of[node][v] == truth_of[node]).all(1) + for kids in groups: + t = np.sort(np.stack([key(truth_of[k]) for k in kids], 1), 1) + for v in range(V): + correct[v] &= ( + np.sort(np.stack([key(views_of[k][v]) for k in kids], 1), 1) + == t + ).all(1) + if not usable: + self.gaps.append( + f"pairing accuracy of {h}: the matched TARGETS groups ({', '.join(matched)}) list a different number of jets than the nodes they name" + ) + continue + if all(np.array_equal(correct[0], correct[v]) for v in range(1, V)): + self.gaps.append( + f"pairing accuracy of {h}: the TARGETS groups do not distinguish its views" + ) + continue + rows = [] + per_fold = [] + thr_rows = {} + for fd, s in test: + if h not in s.views: + continue + idx = s.index + comp, cont = complete[idx], contained[idx] & complete[idx] + am = s.views[h].argmax(1) + mx = s.views[h].max(1) + ok = correct[am, idx] + acc_m = ok[comp].mean() if comp.any() else np.nan + acc_c = ok[cont].mean() if cont.any() else np.nan + per_fold.append((acc_m, acc_c)) + rows.append( + [fd.fold, int(comp.sum()), int(cont.sum()), fmt(acc_m), fmt(acc_c)] + ) + thr = [1.0 / V] + [t for t in VIEW_THRESHOLDS if t > 1.0 / V + 1e-9] + thr_rows[fd.fold] = [ + ( + t, + ok[cont & (mx > t)].mean() + if (cont & (mx > t)).any() + else np.nan, + int((cont & (mx > t)).sum()), + ) + for t in thr + ] + if not rows: + continue + rows.append( + [ + "mean ± sd", + "", + "", + mean_sd([a for a, _ in per_fold]), + mean_sd([c for _, c in per_fold]), + ] + ) + parts.append( + table( + [ + "fold", + "events with complete truth", + "of which contained", + "accuracy (complete)", + "accuracy (contained)", + ], + rows, + f"Pairing accuracy of the argmax view of {h} against TARGETS {', '.join(g for g, _ in usable)} (random: 1/{V} = {1.0 / V:.3f})", + ) + ) + fig, (a, b) = plt.subplots(1, 2, figsize=(7.8, 3.2)) + thr = [t for t, _, _ in next(iter(thr_rows.values()))] + accs = np.array([[acc for _, acc, _ in thr_rows[k]] for k in thr_rows]) + ns = np.array([[nn for _, _, nn in thr_rows[k]] for k in thr_rows]).sum(0) + x = np.arange(len(thr)) + for j, k in enumerate(thr_rows): + a.plot( + x, + 100 * accs[j], + "-", + color=f"C{j % 10}", + lw=0.8, + alpha=0.5, + label=f"fold {k}", + ) + cnt = np.isfinite(accs).sum(0) + mu = np.where(cnt > 0, np.nansum(accs, 0) / np.maximum(cnt, 1), np.nan) + a.plot(x, 100 * mu, "o-", color="k", lw=1.4, ms=4, label="mean over folds") + for xi, (m, nn) in enumerate(zip(mu, ns)): + a.annotate( + f"N = {nn:,}", + (xi, 100 * m), + textcoords="offset points", + xytext=(0, 7 + 8 * (xi % 2)), + ha="center", + fontsize=6, + ) + a.axhline(100.0 / V, color="k", ls="--", lw=0.8) + a.set_xticks(x) + a.set_xticklabels([f"> {t:.2f}" for t in thr]) + a.set_xlabel(f"max view probability of {h}") + a.set_ylabel("pairing accuracy [%]") + a.set_ylim(0, 105) + a.legend(frameon=False) + a.grid(alpha=0.25, lw=0.5) + a.set_title("(a) accuracy above a confidence threshold", loc="left") + # binned accuracy + edges = np.linspace(1.0 / V, 1.0, 11) + for j, (fd, s) in enumerate([(fd, s) for fd, s in test if h in s.views]): + idx = s.index + cont = contained[idx] & complete[idx] + am, mx = s.views[h].argmax(1), s.views[h].max(1) + ok = correct[am, idx] + which = np.clip(np.digitize(mx, edges) - 1, 0, len(edges) - 2) + acc = [ + ok[cont & (which == b)].mean() + if (cont & (which == b)).any() + else np.nan + for b in range(len(edges) - 1) + ] + b.plot( + 0.5 * (edges[1:] + edges[:-1]), + 100 * np.array(acc), + "o-", + color=f"C{j % 10}", + lw=1.0, + ms=3, + label=f"fold {fd.fold}", + ) + b.plot( + [1.0 / V, 1], + [100.0 / V, 100], + "k:", + lw=0.8, + label="perfect calibration", + ) + b.set_xlabel(f"max view probability of {h}") + b.set_ylabel("pairing accuracy [%]") + b.set_ylim(0, 105) + b.legend(frameon=False) + b.grid(alpha=0.25, lw=0.5) + b.set_title("(b) accuracy per probability bin", loc="left") + parts.append( + figure( + fig, + f"Pairing accuracy of {h} on contained events of the test folds (all generator-matched jets among the candidates): (a) for events whose highest view probability exceeds a threshold, N summed over folds, dashed = random; (b) in bins of the highest view probability, dotted = accuracy equal to the probability.", + ) + ) + return "".join(parts) + + # ---- inputs as the network sees them + def _node_phi_label(self, h: str, node: str) -> str: + T = self.model.tables[h] + parents = [ + (pn, j) + for pn in self.model.levels[h] + for j, c in enumerate(T.nodes[pn].children) + if c == node + ] + pnodes = {pn for pn, _ in parents} + if len(pnodes) == 1: + pnode = next(iter(pnodes)) + P = T.nodes[node].n + allidx = torch.cat([T.nodes[pnode].child_index[:, j] for _, j in parents]) + if allidx.numel() == P and allidx.unique().numel() == P: + return f"Δφ to {pnode}" + return "φ" + + def collect_network_inputs(self, idx: np.ndarray, chunk: int = 16384) -> dict: + """{key: (feature names, values (m, F), labels (m,), weights (m,))} of every tensor the network + embeds, over the events ``idx`` and all present candidates.""" + spec, model = self.spec, self.model + acc: dict[str, tuple] = {} + + def add(key, names, x, present, rows): + x = x.detach().cpu().numpy() + n, _, P = x.shape + pres = ( + np.ones((n, P), bool) + if present is None + else present.detach().cpu().numpy().astype(bool) + ) + vals = x.transpose(0, 2, 1)[pres] + lab = np.broadcast_to(self.labels[rows][:, None], (n, P))[pres] + w = np.broadcast_to(self.weights[rows][:, None], (n, P))[pres] + e = acc.setdefault(key, (list(names), [], [], [])) + e[1].append(vals) + e[2].append(lab) + e[3].append(w) + + def event_name(name): + return f"log({name} - {spec.min_jets - 1})" if "nSelJets" in name else name + + with torch.no_grad(): + for start in range(0, len(idx), chunk): + rows = idx[start : start + chunk] + sel = torch.as_tensor(rows) + inputs = {k: v[sel] for k, v in self.ds.inputs.items()} + prep = model._prepare(inputs) + if prep["event"] is not None: + add( + "event scalars", + [event_name(n) for n in spec.event], + prep["event"], + None, + rows, + ) + seen = set() + for h, T in model.tables.items(): + hp = prep["hyp"][h] + for node in model.levels[h]: + if node in seen: + continue + seen.add(node) + nt = T.nodes[node] + for j, child in enumerate(nt.children): + names = ["log(1+pT)", "η", f"Δφ to {node}", "log(1+m)"] + if child in model.leaves: + names += model._extras(child) + add( + f"{node} ← {child}", + names, + hp["slot_feats"][node][j], + hp["slot_present"][node][j], + rows, + ) + if node in model.tops: + names = ["log(1+pT)", "η", "log(1+m)"] + else: + owner = spec.owner.get(node, h) + owner = owner if node in model.levels.get(owner, []) else h + names = [ + "log(1+pT)", + "η", + self._node_phi_label(owner, node), + "log(1+m)", + ] + add( + f"{node} candidates", + names, + prep["embed_in"][node], + prep["present"][node], + rows, + ) + if "context" in hp: + c = hp["context"] + add( + f"{h} context: {spec.hypotheses[h].context} and the candidates' jets", + [ + "log(1+pT)", + "η", + "log(1+m)", + "flag (2 = candidate jet, 1 = context jet)", + ], + c["feats"], + ~c["mask"], + rows, + ) + for key, r in prep["relations"].items(): + add( + f"relation {key.replace('|', ' - ')}", + ["log(1+M)", "ΔR", "Δφ"], + r, + None, + rows, + ) + return { + k: (names, np.concatenate(v), np.concatenate(lab), np.concatenate(w)) + for k, (names, v, lab, w) in acc.items() + } + + def section_inputs(self) -> str: + idx = self.selected_subset(self.args.max_events) + data = self.collect_network_inputs(idx) + parts = [ + note( + f"Distributions of the quantities the network embeds, computed from the input file by the loader and the network's own preparation step, over {len(idx):,} events" + + ( + " (a random subsample; --max-events)" + if len(idx) < len(self.ds) + else "" + ) + + " and all present candidates; weighted by the event weight (clipped at 0), each class normalised to unit area. Jet features: log(1+pT), η" + + ( + " (reflected so that the leading jet of the first role has η ≥ 0)" + if self.spec.training.reflect_eta + else "" + ) + + ", Δφ to the parent candidate, log(1+m)." + ) + ] + for key, (names, vals, lab, w) in data.items(): + fig, axes = grid_figure(len(names), per_row=6) + for ax, j, name in zip(axes, range(len(names)), names): + series = self.class_series(vals[:, j], lab, w) + hist_panel(ax, series, shared_bins([v for _, v, _, _ in series]), name) + axes[0].legend(frameon=False) + parts.append(figure(fig, f"{key}: {', '.join(names)}.")) + return "".join(parts) + + # ---- raw inputs + def section_raw(self) -> str: + ds, spec = self.ds, self.spec + idx = self.selected_subset(self.args.max_events) + lab, w = self.labels[idx], self.weights[idx] + parts = [ + note( + f"The features of the input file for the members each role selected (per slot), the single objects and the event scalars, before any transformation; {len(idx):,} events" + + (" (random subsample)" if len(idx) < len(ds) else "") + + ", weighted, each class normalised to unit area. A column read for the selection only (e.g. btag) is shown for the selected jets as well." + ) + ] + extra_cols = {} + with h5py.File(self.input, "r") as f: + for role in spec.roles.values(): + coll = spec.collections[role.collection] + cols = [c for c in role.columns if c not in coll.features] + if cols: + group = f["INPUTS"][resolve_group(f, coll.name)] + extra_cols[role.name] = { + c: np.atleast_2d(group[c][...].astype(np.float32).T).T + for c in cols + if c in group + } + for role in spec.roles.values(): + coll = spec.collections[role.collection] + x = ds.inputs[role.name][idx].numpy() # (n, F+1, slots) + present = x[:, -1] > 0 + feats = list(coll.features) + extra = extra_cols.get(role.name, {}) + n_slots = x.shape[2] + names = feats + list(extra) + fig, axes = plt.subplots( + len(names), + n_slots, + figsize=(2.3 * n_slots, 1.9 * len(names)), + squeeze=False, + ) + for i, name in enumerate(names): + if name in feats: + col = x[:, feats.index(name)] + else: + slots = ds.selected[role.name][idx] + col = np.where( + slots >= 0, + extra[name][idx[:, None], np.clip(slots, 0, None)], + np.nan, + ) + bins = shared_bins([col[present]]) + for s in range(n_slots): + ax = axes[i, s] + ok = present[:, s] + series = self.class_series(col[ok, s], lab[ok], w[ok]) + hist_panel(ax, series, bins, f"{name} [{role.name} slot {s}]") + axes[0, 0].legend(frameon=False) + sel_text = f"{role.collection} {role.selector} {role.n}" + ( + " lowest" if role.ascending else "" + ) + parts.append( + figure( + fig, + f"Role {role.name} ({sel_text}): rows = features, columns = slots in selection order; empty slots are left out.", + dpi=90, + ) + ) + singles = [ + c for c in spec.collections.values() if c.max == 1 and c.name in ds.inputs + ] + for c in singles: + x = ds.inputs[c.name][idx].numpy() + present = x[:, -1, 0] > 0 + fig, axes = grid_figure(len(c.features), per_row=6) + for ax, j, name in zip(axes, range(len(c.features)), c.features): + series = self.class_series(x[present, j, 0], lab[present], w[present]) + hist_panel( + ax, + series, + shared_bins([v for _, v, _, _ in series]), + f"{name} [{c.name}]", + ) + axes[0].legend(frameon=False) + parts.append( + figure(fig, f"Single object {c.name}: {', '.join(c.features)}.") + ) + if spec.event: + ev = ds.inputs["event"][idx].numpy() + fig, axes = grid_figure(len(spec.event), per_row=6) + for ax, j, name in zip(axes, range(len(spec.event)), spec.event): + kind, where = event_source(spec, name) + series = self.class_series(ev[:, j], lab, w) + hist_panel( + ax, + series, + shared_bins([v for _, v, _, _ in series]), + f"{name} ({'computed from ' + where if kind == 'computed' else where})", + ) + axes[0].legend(frameon=False) + parts.append( + figure(fig, "Event scalars as read or computed by the loader.") + ) + return "".join(parts) + + # ---- weights + def section_weights(self) -> str: + lab, w = self.labels, self.weights + parts = [] + fig, (a, b) = plt.subplots(1, 2, figsize=(7.8, 3.2)) + series = [] + for i, c in enumerate(self.classes): + sel = lab == i + if sel.any(): + series.append( + ( + c, + w[sel], + np.ones(int(sel.sum())), + {"color": class_color(i), "lw": 1.3}, + ) + ) + bins = shared_bins([v for _, v, _, _ in series], 60) + for lab_c, v, _, st in series: + a.hist(v, bins=bins, histtype="step", label=lab_c, **st) + a.set_yscale("log") + a.set_xlabel("event weight (as in the file)") + a.set_ylabel("events") + a.legend(frameon=False) + a.set_title("(a) raw weights", loc="left") + normed = w.copy() + for i in range(len(self.classes)): + sel = lab == i + tot = np.clip(w[sel], 0, None).sum() + if tot > 0: + normed[sel] = np.clip(w[sel], 0, None) / tot + series = [ + ( + c, + normed[lab == i], + np.ones(int((lab == i).sum())), + {"color": class_color(i), "lw": 1.3}, + ) + for i, c in enumerate(self.classes) + if (lab == i).any() + ] + bins = shared_bins([v for _, v, _, _ in series], 60) + for lab_c, v, _, st in series: + b.hist(v, bins=bins, histtype="step", label=lab_c, **st) + b.set_yscale("log") + b.set_xlabel("class-normalised weight (clipped at 0, class sum = 1)") + b.set_ylabel("events") + b.set_title("(b) as training uses them", loc="left") + parts.append( + figure( + fig, + "Event weights per class over the whole input file: (a) as stored; (b) clipped at 0 and divided by the class sum, which is how the training loss weights an event (the training normalises over its own training events of every fold).", + ) + ) + rows = [] + for i, c in enumerate(self.classes): + sel = lab == i + if not sel.any(): + continue + ws = w[sel] + s1, s2 = ws.sum(), (ws**2).sum() + neff = s1**2 / s2 if s2 > 0 else 0.0 + rows.append( + [ + c, + int(sel.sum()), + float(s1), + f"{ws.min():.3g}", + f"{ws.max():.3g}", + f"{ws.mean():.3g}", + f"{ws.std():.3g}", + fmt(neff, 1), + fmt(neff / sel.sum(), 4), + int((ws < 0).sum()), + fmt(abs(ws[ws < 0].sum()) / s1 if s1 > 0 else np.nan), + ] + ) + parts.append( + table( + [ + "class", + "events", + "sum", + "min", + "max", + "mean", + "sd", + "effective events (sum w)^2 / sum w^2", + "effective / events", + "negative weights", + "|negative sum| / sum", + ], + rows, + "Weight statistics per class (whole file)", + ) + ) + folds = sorted(np.unique(self.fold_of).tolist()) + rows = [] + for k in folds: + fsel = self.fold_of == k + row = [k, int(fsel.sum())] + for i in range(len(self.classes)): + sel = fsel & (lab == i) + tot = w[lab == i].sum() + row += [ + int(sel.sum()), + float(w[sel].sum()), + fmt(w[sel].sum() / tot if tot else np.nan, 3), + int((w[sel] < 0).sum()), + ] + rows.append(row) + header = ["fold", "events"] + [ + f"{c}: {q}" + for c in self.classes + for q in ("events", "sum", "share", "negative") + ] + parts.append( + table( + header, + rows, + "Events and weights per fold and class (closure: every fold should carry about 1/n_folds of each class)", + ) + ) + rows = [] + for fd in self.folds: + k = fd.fold + excl = fd.exclude_folds or [] + train = ~np.isin(self.fold_of, [k, *excl]) + row = [k, esc(excl) if excl else "-"] + for i in range(len(self.classes)): + sel = train & (lab == i) + tot = np.clip(w[sel], 0, None).sum() + row += [int(sel.sum()), float(tot), int((w[sel] < 0).sum())] + rows.append(row) + header = ["held-out fold", "validation folds"] + [ + f"{c}: {q}" + for c in self.classes + for q in ("training events", "training sum", "negative (count as 0)") + ] + parts.append( + table( + header, + rows, + "Training set of every fold: the weights of every class are normalised to 1 over these events", + ) + ) + return "".join(parts) + + # ---- other checks + def section_other(self) -> str: + ds, spec = self.ds, self.spec + parts = [] + idx = self.selected_subset(self.args.max_events) + lab, w = self.labels[idx], self.weights[idx] + profiles = [] + first_role = next(iter(spec.roles.values()), None) + if first_role is not None: + x = ds.inputs[first_role.name][idx].numpy() + feats = list(spec.collections[first_role.collection].features) + if "pt" in feats: + pt = x[:, feats.index("pt")] + present = x[:, -1] > 0 + for s in (0, first_role.n - 1): + if s < pt.shape[1]: + profiles.append( + ( + f"pT of {first_role.name} slot {s} [GeV]", + np.where(present[:, s], pt[:, s], np.nan), + ) + ) + for j, name in enumerate(spec.event): + if "nSelJets" in name: + profiles.append((name, ds.inputs["event"][idx].numpy()[:, j])) + if not profiles: + self.gaps.append( + "weight profiles: no jet role or jet multiplicity in the configuration" + ) + return "".join(parts) + fig, axes = grid_figure(len(profiles), per_row=3, w=3.2, h=2.6) + for ax, (name, x) in zip(axes, profiles): + ok = np.isfinite(x) + bins = shared_bins([x[ok]], 20) + centres = 0.5 * (bins[1:] + bins[:-1]) + for i, c in enumerate(self.classes): + sel = ok & (lab == i) + if not sel.any(): + continue + which = np.clip(np.digitize(x[sel], bins) - 1, 0, len(bins) - 2) + means = np.array( + [ + w[sel][which == b].mean() if (which == b).any() else np.nan + for b in range(len(bins) - 1) + ] + ) + sems = np.array( + [ + w[sel][which == b].std() / np.sqrt((which == b).sum()) + if (which == b).sum() > 1 + else np.nan + for b in range(len(bins) - 1) + ] + ) + ax.errorbar( + centres, + means, + sems, + fmt="o-", + color=class_color(i), + ms=3, + lw=1.0, + capsize=2, + label=c, + ) + ax.set_xlabel(name) + ax.set_ylabel("mean event weight") + ax.grid(alpha=0.25, lw=0.5) + axes[0].legend(frameon=False) + parts.append( + figure( + fig, + "Mean event weight (as stored) per class in bins of the leading and the last candidate jet pT of the first role and of the jet multiplicity; error bars are the standard error of the mean.", + ) + ) + return "".join(parts) + + # ---------------------------------------------------------------- page + def render(self) -> str: + bodies = {} + for key, title in SECTIONS: + try: + bodies[key] = getattr(self, f"section_{key}")() + except Exception as e: # a section that fails is reported, not fatal + logging.error(f"section '{title}' failed:\n{traceback.format_exc()}") + self.gaps.append( + f"section '{title}' could not be produced: {type(e).__name__}: {e}" + ) + bodies[key] = ( + f'

This section could not be produced: {esc(type(e).__name__)}: {esc(e)}

' + ) + plt.close("all") + toc = "".join( + f'
  • {i + 1}. {esc(t)}
  • ' + for i, (k, t) in enumerate(SECTIONS) + ) + sections = "".join( + f'

    {i + 1}. {esc(t)}

    {bodies[k]}
    ' + for i, (k, t) in enumerate(SECTIONS) + ) + cmd = esc(" ".join(sys.argv)) + return f""" +CoReNet report: {esc(self.run_dir)} + +

    CoReNet report

    +

    run directory {esc(self.run_dir)} | input {esc(self.input)} | generated {esc(self.generated)} | corenet {esc(__version__)}
    command: {cmd}

    + +{sections} +

    Weighted quantities use the event weights of the input file clipped at 0, as the training does. Mean ± sd are over folds with ddof = 1.

    + +""" + + +_GIT_HASH: str | None = None + + +def git_hash() -> str: + """Short git hash of the package checkout, '' when not in a git repository.""" + global _GIT_HASH + if _GIT_HASH is None: + try: + out = subprocess.run( + [ + "git", + "-C", + str(Path(__file__).resolve().parent), + "rev-parse", + "--short", + "HEAD", + ], + capture_output=True, + text=True, + timeout=5, + ) + _GIT_HASH = out.stdout.strip() if out.returncode == 0 else "" + except Exception: + _GIT_HASH = "" + return _GIT_HASH + + +# ----------------------------------------------------------------------------- command line +def build_parser() -> argparse.ArgumentParser: + p = argparse.ArgumentParser( + prog="corenet-report", + description=__doc__, + formatter_class=argparse.RawDescriptionHelpFormatter, + ) + p.add_argument( + "--run-dir", + required=True, + help="directory written by corenet-train (and corenet-evaluate)", + ) + p.add_argument( + "--input", required=True, help="the HDF5 file the run was trained on" + ) + p.add_argument( + "--output", default=None, help="HTML file to write [/report.html]" + ) + p.add_argument( + "--config", + default=None, + help="configuration file [the one stored in the checkpoint]", + ) + p.add_argument( + "--folds", + type=int, + nargs="*", + default=None, + help="folds to include [every fold found in the run directory]", + ) + p.add_argument( + "--max-events", + type=int, + default=500000, + help="events of the input file used for the input and weight distributions (random subsample above this) [500000]; 0 = all", + ) + p.add_argument( + "--n-folds", + type=int, + default=None, + help="fold = squares(event) %% K instead of EVENT/fold [the run's rule]", + ) + p.add_argument( + "--fold-seed", type=int, default=None, help="seed of that rule [the run's]" + ) + p.add_argument( + "--rest-order", + choices=["pt", "selector"], + default=None, + help="ordering of the 'rest' selector [the run's]", + ) + return p + + +def main(argv=None) -> Path: + args = build_parser().parse_args(argv) + logging.basicConfig( + level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s" + ) + report = Report(args) + report.load() + page = report.render() + out = Path(args.output) if args.output else report.run_dir / "report.html" + out.parent.mkdir(parents=True, exist_ok=True) + out.write_text(page, encoding="utf-8") + logging.info(f"wrote {out} ({out.stat().st_size / 1e6:.2f} MB)") + for g in dict.fromkeys(report.gaps): # inputs the run did not provide; verbose only + logging.debug(f"not produced: {g}") + return out + + +def cli(): + main() + + +if __name__ == "__main__": + cli() diff --git a/src/corenet/train.py b/src/corenet/train.py index 0405754..93717b6 100644 --- a/src/corenet/train.py +++ b/src/corenet/train.py @@ -8,7 +8,8 @@ * Adam, learning rate ``lr`` (1e-2), MultiStepLR with ``lr_milestones`` (epochs 15..24) and ``lr_gamma`` (0.25): the learning rate is multiplied by 0.25 after each of those epochs; * ``epochs`` (20) epochs; batch size ``batch_size_init`` (1024) at epoch 0, doubled at epochs - 1, 3, 6 and 10, with (64, 16, 4, 1, 0) ghost batches; + 1, 3, 6 and 10, with (64, 16, 4, 1, 0) ghost batches (a batch is split into equal ghost batches, + so their number must divide the batch size); * the input normalisation of every embedder is frozen from the training set before epoch 0; * at epoch ``freeze_epoch`` (15) the fine-tuning stage starts: every parameter whose name does not start with one of the ``finetune_layers`` prefixes (the output heads) is frozen; ``finetune_lr`` @@ -22,9 +23,11 @@ binary cross entropy of its head against ``label == class index of o``. The view (candidate selection) scores are not supervised. -Folds: ``EVENT/fold`` of the file, or ``--n-folds K`` for fold = squares(event) % K from the event -numbers (``corenet.data.folds.fold_of``, seed ``--fold-seed``, 0 = the rule of the example files); -``--exclude-folds next`` is the fold (k + 1) % n_folds of held-out fold k. +Folds: ``EVENT/fold`` of the file (ids exactly 0..K-1), or ``--n-folds K`` for fold = squares(event) % K +from the event numbers (``corenet.data.folds.fold_of``, seed ``--fold-seed``, 0 = the rule of the +example files); ``--exclude-folds next`` is the fold (k + 1) % K of held-out fold k. Training needs one +of the two. ``--early-stop-patience`` needs a validation set with events of class 1 and of another +class; an empty validation set or a missing validation AUC stops the run with an error. For each held-out fold k this writes, in ``--output-dir``: model_fold{k}.pkl checkpoint: state dict, configuration, classes and metadata @@ -46,7 +49,7 @@ from .checkpoint import save_checkpoint from .data.folds import FOLD_SEED -from .data.h5 import H5Dataset +from .data.h5 import FOLD_KEY, H5Dataset from .evaluate import Batches, count_parameters, predict, write_predictions from .models.config import ConfigError, NetworkSpec, parse_config, read_config_dict from .models.network import CoReNet @@ -93,28 +96,63 @@ def batch_schedule(epoch: int, bs_init: int) -> tuple[int, int]: ) bs = bs_init << stage n_gb = (64, 16, 4, 1, 0)[stage] - return bs, min(n_gb, max(1, bs // MIN_GHOST_BATCH_SIZE)) # no-op for bs_init >= 512 + return bs, ghost_batches( + bs, n_gb + ) # equals the stage's count for any power-of-two batch size >= 512 + + +def ghost_batches(bs: int, cap: int) -> int: + """The largest number of ghost batches <= ``cap`` that divides the batch size ``bs`` into ghost batches of + at least MIN_GHOST_BATCH_SIZE events (the GBN layers split a batch into equal parts); 1 when no such + divisor exists, 0 when ``cap`` is 0 (running statistics).""" + if cap <= 0: + return 0 + for n in range(min(cap, bs // MIN_GHOST_BATCH_SIZE), 1, -1): + if bs % n == 0: + return n + return 1 def finetune_batch(epoch: int, bs_init: int, finetune_batch_size: int | None): - """(batch size, ghost batches) of the fine-tuning stage: ``finetune_batch_size`` when set (ghost batches - capped so that none has fewer than MIN_GHOST_BATCH_SIZE events, 0 stays 0), else the schedule.""" + """(batch size, ghost batches) of the fine-tuning stage: ``finetune_batch_size`` when set, with the + schedule's ghost batches reduced to the largest divisor of the batch size that keeps every ghost batch + at >= MIN_GHOST_BATCH_SIZE events (``ghost_batches``; 0 stays 0), else the schedule.""" bs, n_gb = batch_schedule(epoch, bs_init) if finetune_batch_size is None: return bs, n_gb bs = finetune_batch_size - return bs, min(n_gb, max(1, bs // MIN_GHOST_BATCH_SIZE)) + return bs, ghost_batches(bs, n_gb) + + +def check_finetune_layers(model: CoReNet, prefixes) -> None: + """At least one ``finetune_layers`` prefix must match a trainable parameter name; otherwise the fine-tuning + stage would freeze everything. Raises ConfigError naming the prefixes and the prefixes that exist. A single + prefix that matches nothing is allowed: the default ``head_reduce.`` and ``aux_heads.`` exist only in + networks with several combined pixels or auxiliary heads.""" + names = [n for n, p in model.named_parameters() if p.requires_grad] + if not any(n.startswith(p) for p in prefixes for n in names): + valid = sorted({n.split(".")[0] + "." for n in names}) + raise ConfigError( + f"finetune_layers: {list(prefixes)} match no trainable parameter (the fine-tuning stage would train " + f"nothing); parameter names start with {valid}, e.g. {names[:3]}; '{NO_FINETUNE_LAYERS}' freezes nothing" + ) def freeze_heads(model: CoReNet, prefixes=HEAD_PREFIXES) -> int: """Freeze every parameter whose name does not start with one of ``prefixes`` (default: the output heads head_main, head_reduce, aux_heads); parameters that are non-trainable by construction stay so. - Returns the number of trainable parameters.""" + Returns the number of trainable parameters; raises ConfigError when none is left.""" prefixes = tuple(prefixes) + check_finetune_layers(model, prefixes) # names the prefixes that match nothing for name, param in model.named_parameters(): if not name.startswith(prefixes): param.requires_grad = False - return count_trainable(model) + n = count_trainable(model) + if n == 0: + raise ConfigError( + f"finetune_layers {list(prefixes)}: no trainable parameter is left for the fine-tuning stage" + ) + return n def count_trainable(model: torch.nn.Module) -> int: @@ -248,11 +286,19 @@ def train_fold( logging.info( f"fold {fold}: {len(train_idx):,} training / {len(val_idx):,} validation / {len(test_idx):,} held-out events" ) + if args.early_stop_patience and not len(val_idx): + raise SystemExit( + f"fold {fold}: --early-stop-patience {args.early_stop_patience} needs validation events, but " + f"--exclude-folds {exclude_folds} leaves none (held-out fold {fold}; folds 0..{ds.n_folds - 1}): " + "exclude a fold other than the held-out one" + ) model = CoReNet(spec, device=str(device)).to(device) logging.info( f"{args.config}: classes {spec.classes}, {count_parameters(model):,} parameters" ) + if args.finetune_layers: + check_finetune_layers(model, args.finetune_layers) data = Batches(ds, device) w_dev = weights.to(device) aux, head_weights = aux_targets(spec), dict(spec.training.head_weights or {}) @@ -380,9 +426,19 @@ def metadata(epoch: int) -> dict: else "" ) ) - if args.early_stop_patience and len(val_idx): + if args.early_stop_patience: auc = entry.get("val_auc") - if auc is not None and (best_auc is None or auc > best_auc): + if auc is None or not np.isfinite(auc): + val_label = labels[val_idx] + counts = { + c: int((val_label == i).sum()) for i, c in enumerate(spec.classes) + } + raise SystemExit( + f"fold {fold}: epoch {epoch}: no validation AUC, so --early-stop-patience cannot be applied: " + f"the validation folds {exclude_folds} hold {counts} events (positive-weight events of " + f"class 1 `{spec.classes[1]}` and of another class are needed)" + ) + if best_auc is None or auc > best_auc: best_auc, best_epoch, bad = auc, epoch, 0 best_state = { k: v.detach().clone() for k, v in model.state_dict().items() @@ -663,6 +719,11 @@ def main(argv=None): n_folds=args.n_folds, fold_seed=args.fold_seed, ) + if not ds.has_folds: + raise SystemExit( + f"{args.input}: no {FOLD_KEY} dataset; training needs the fold of every event: pass --n-folds K " + "to use fold = squares(event) % K from EVENT/event" + ) logging.info( f"{args.input}: {len(ds):,} events, {ds.n_folds} folds ({ds.fold_rule}), device={device}, recipe " + ", ".join(f"{k}={getattr(args, k)}" for k in RECIPE_DEFAULTS) diff --git a/src/corenet/validate.py b/src/corenet/validate.py index d5f3a2c..fb1d01a 100644 --- a/src/corenet/validate.py +++ b/src/corenet/validate.py @@ -10,9 +10,9 @@ 2. Data (with ``--input``): every dataset the loader reads is looked up with the loader's own rules (``corenet.data.h5``: group aliases, used collections, computed event scalars, label keys) and its dtype and shape are checked; the labels must lie in 0..len(outputs), the weights must be finite, - the folds must exist for training (``EVENT/fold``, or ``--n-folds K`` for fold = squares(event) % K - from the event numbers, seed ``--fold-seed``; the report gives the fold source and the events per - fold); the present members + the folds must exist for training (``EVENT/fold`` with ids exactly 0..K-1, or ``--n-folds K`` for + fold = squares(event) % K from the event numbers, seed ``--fold-seed``; the report gives the fold + source and the events per fold); the present members must have finite features. Finally the actual loader (``H5Dataset``) reads the file and the network runs a forward pass on the first events, so a file that passes here also loads in corenet-train. @@ -411,6 +411,20 @@ def length(name, ds): ) elif length(loader.FOLD_KEY, fo): column = fo[...].astype(np.int64) + ids = np.unique(column) + if not np.array_equal(ids, np.arange(len(ids))): + shown = ids[:12].tolist() + (["..."] if len(ids) > 12 else []) + msg = ( + f"{loader.FOLD_KEY}: fold ids {shown} are not 0..K-1 (the ids of K folds must be " + "exactly 0..K-1; --exclude-folds next = fold (k + 1) % K): renumber the column" + ) + if n_folds is None: + report.fail(sec, msg + " or pass --n-folds K") + else: + report.warn( + sec, msg + f"; --n-folds {n_folds} ignores the column" + ) + column = None elif n_folds is None: report.fail( sec, diff --git a/tests/test_compat.py b/tests/test_compat.py new file mode 100644 index 0000000..bf6eedc --- /dev/null +++ b/tests/test_compat.py @@ -0,0 +1,100 @@ +"""The package imports on every Python of ``requires-python`` (>= 3.9): no ``X | Y`` type union is built at +run time (outside annotations, which ``from __future__ import annotations`` keeps as strings), and every +source file parses with the 3.9 grammar.""" + +from __future__ import annotations + +import ast +import typing +from pathlib import Path + +import pytest + +SRC = Path(__file__).resolve().parent.parent / "src" / "corenet" +SOURCES = sorted(SRC.rglob("*.py")) + + +def _statements(body): + """Statements evaluated at import time: module and class bodies, not function bodies; the value of an + annotated assignment, not its annotation.""" + for stmt in body: + if isinstance(stmt, (ast.FunctionDef, ast.AsyncFunctionDef)): + continue + if isinstance(stmt, ast.ClassDef): + yield from _statements(stmt.body) + elif isinstance(stmt, ast.AnnAssign): + if stmt.value is not None: + yield stmt.value + else: + yield stmt + + +def _type_like(node) -> bool: + """A name, attribute, subscript (``list[int]``), ``None`` or a ``|`` of those: a type expression.""" + if isinstance(node, (ast.Name, ast.Attribute, ast.Subscript)): + return True + if isinstance(node, ast.Constant) and node.value is None: + return True + return ( + isinstance(node, ast.BinOp) + and isinstance(node.op, ast.BitOr) + and _type_like(node.left) + and _type_like(node.right) + ) + + +def runtime_unions(path: Path) -> list[str]: + """``:`` of every ``A | B`` between type expressions evaluated at import time.""" + found = [] + for stmt in _statements(ast.parse(path.read_text()).body): + for node in ast.walk(stmt): + if ( + isinstance(node, ast.BinOp) + and isinstance(node.op, ast.BitOr) + and _type_like(node) + ): + found.append(f"{_label(path)}:{node.lineno}") + return sorted(set(found)) + + +def _label(path: Path) -> str: + return str(path.relative_to(SRC)) if SRC in path.parents else path.name + + +def test_detector_sees_a_runtime_union(tmp_path): + bad = tmp_path / "bad.py" + bad.write_text( + "from __future__ import annotations\nfrom collections.abc import Iterable\n" + "SeedLike = int | str | Iterable[int | str]\nx: int | None = None\n" + "def f(a: int | None) -> str | None:\n y: int | None = None\n return None\n" + ) + assert runtime_unions(bad) == [ + "bad.py:3" + ] # the annotations on lines 4-6 are not evaluated + + +@pytest.mark.parametrize( + "path", SOURCES, ids=[str(p.relative_to(SRC)) for p in SOURCES] +) +def test_no_runtime_type_union(path): + assert runtime_unions(path) == [], ( + "type union evaluated at import time (fails on Python 3.9); use typing.Union" + ) + + +@pytest.mark.parametrize( + "path", SOURCES, ids=[str(p.relative_to(SRC)) for p in SOURCES] +) +def test_parses_with_python39_grammar(path): + ast.parse(path.read_text(), filename=str(path), feature_version=(3, 9)) + + +def test_seed_alias_is_a_typing_union(): + from corenet.data.folds import SeedLike, Squares + + assert typing.get_origin(SeedLike) is typing.Union + assert set(typing.get_args(SeedLike)[:2]) == {int, str} + assert ( + Squares("delphes_h5", "fold", 0)._seed + == Squares(("delphes_h5", "fold", 0))._seed + ) diff --git a/tests/test_h5.py b/tests/test_h5.py index d2c8aec..03913c3 100644 --- a/tests/test_h5.py +++ b/tests/test_h5.py @@ -3,6 +3,7 @@ from __future__ import annotations +import re from pathlib import Path import h5py @@ -417,8 +418,10 @@ def test_fold_rule(): def test_missing_fold_column(tmp_path): - """Without EVENT/fold the loader asks for --n-folds; with it the folds come from EVENT/event.""" + """Without EVENT/fold (and without n_folds) the file loads without folds: every fold is -1 and selecting a + fold asks for --n-folds; with n_folds the folds come from EVENT/event.""" from corenet.data.folds import fold_of + from corenet.data.h5 import FOLD_RULE_NONE, NO_FOLD src = _need(DATA / "hh4b_small.h5") spec = load_config(repo / "examples/hh4b/hh4b.yaml") @@ -437,11 +440,146 @@ def test_missing_fold_column(tmp_path): ): g.create_dataset(key, data=f[key][:500]) event = f["EVENT/event"][:500] - with pytest.raises(KeyError, match="no EVENT/fold dataset; pass --n-folds"): - H5Dataset(dst, spec) + ds0 = H5Dataset(dst, spec) + assert not ds0.has_folds and ds0.n_folds == 0 and ds0.fold_rule == FOLD_RULE_NONE + assert len(ds0) == 500 and (ds0.folds.numpy() == NO_FOLD).all() + assert ds0[3]["fold"] == NO_FOLD and ds0.inputs["b"].shape == (500, 5, 4) + with pytest.raises( + KeyError, + match=r"no EVENT/fold dataset; holding out fold 0 needs .* pass --n-folds K", + ): + ds0.fold_indices(0) + with pytest.raises(KeyError, match="no EVENT/fold dataset; training needs"): + ds0.require_folds("training") ds = H5Dataset(dst, spec, n_folds=5) + assert ds.has_folds assert ds.n_folds == 5 and (ds.folds.numpy() == fold_of(event, 5)).all() with h5py.File(dst, "a") as g: del g["EVENT/event"] with pytest.raises(KeyError, match="no EVENT/event dataset"): H5Dataset(dst, spec, n_folds=5) + + +def test_fold_ids_must_be_0_to_k(tmp_path): + """EVENT/fold ids must be exactly 0..K-1: with a gap K would be miscounted and `next` would pick the wrong + fold, so the loader refuses the column (--n-folds K still ignores it); contiguous ids give K and + `next` = (k + 1) % K.""" + from synthetic import write_synthetic + + from corenet.train import resolve_exclude_folds + + path = write_synthetic(tmp_path / "f.h5", "hh4b", n=300) + spec = load_config(repo / "examples/hh4b/hh4b.yaml") + with h5py.File(path, "a") as f: + event = f["EVENT/event"][...] + del f["EVENT/fold"] + f.create_dataset("EVENT/fold", data=event % 3) + ds = H5Dataset(path, spec) + assert ds.n_folds == 3 and resolve_exclude_folds(["next"], 2, ds.n_folds) == [0] + assert resolve_exclude_folds(["next"], 0, ds.n_folds) == [1] + for bad in ([0, 1, 3], [1, 2, 3], [0, 2, 4]): + with h5py.File(path, "a") as f: + del f["EVENT/fold"] + f.create_dataset( + "EVENT/fold", data=np.array(bad, dtype=np.int64)[event % 3] + ) + with pytest.raises( + ValueError, + match=re.escape( + f"EVENT/fold holds the fold ids {bad}; the ids of K folds must be exactly 0..K-1" + ), + ): + H5Dataset(path, spec) + assert H5Dataset(path, spec, n_folds=3).n_folds == 3 # the column is ignored + + +def test_selector_sorts_at_file_precision(tmp_path): + """Selector columns are compared at the file's dtype: an int64 column with 16777216 and 16777217 (equal + once cast to float32) selects the higher one; unsigned and bool columns sort by value too (no wrap-around + on negation). The input tensors stay float32.""" + from synthetic import write_synthetic + + path = write_synthetic(tmp_path / "s.h5", "hh4b", n=300) + with h5py.File(path, "a") as f: + mask = f["INPUTS/Source/MASK"][...] + pt = f["INPUTS/Source/pt"][...] + n, slots = mask.shape + assert mask[:, :2].all() # jets 0 and 1 are present in every event (>= 4 jets) + rank = np.zeros((n, slots), dtype=np.int64) + rank[:, 0], rank[:, 1] = 16777216, 16777217 + f.create_dataset("INPUTS/Source/rank", data=rank) + low = np.full((n, slots), 2**40, dtype=np.int64) + low[:, 0], low[:, 1] = 16777216, 16777217 + f.create_dataset("INPUTS/Source/low", data=low) + urank = np.zeros((n, slots), dtype=np.uint8) + urank[:, 1] = ( + 7 # jet 0 has 0: -0 wraps to 0 and would be preferred by a wrapped negation + ) + f.create_dataset("INPUTS/Source/urank", data=urank) + flag = np.zeros((n, slots), dtype=bool) + flag[:, 1] = True + f.create_dataset("INPUTS/Source/flag", data=flag) + for column, kind in (("rank", "int64"), ("urank", "uint8"), ("flag", "bool")): + cfg = _hh4b_cfg(["pt", "eta", "phi", "mass"]) + cfg["objects"]["b"] = f"jets {column} 4" + x = H5Dataset(path, parse_config(cfg)).inputs + assert x["b"].dtype == torch.float32 and x["b"].shape == (n, 5, 4) + assert np.array_equal(x["b"][:, 0, 0].numpy(), pt[:, 1]), ( + f"{kind} column: jet 1 has the highest value" + ) + if column == "rank": # 16777216 > 0: jet 0 is second + assert np.array_equal(x["b"][:, 0, 1].numpy(), pt[:, 0]) + cfg = _hh4b_cfg(["pt", "eta", "phi", "mass"]) + cfg["objects"]["b"] = "jets low 4 lowest" + x = H5Dataset(path, parse_config(cfg)).inputs + assert np.array_equal( + x["b"][:, 0, 0].numpy(), pt[:, 0] + ) # the lowest: 16777216 is jet 0 + assert np.array_equal(x["b"][:, 0, 1].numpy(), pt[:, 1]) + # the b-jet selection of the example files is unchanged (float32 in the file): test_hh4b_small / test_bbww_small + + +def test_context_role_extras_reach_the_network(): + """An extra feature (btag) of the jets also enters the context objects: the context embedder has 4 + 1 + input channels, the output depends on the btag of the `other` jets (which the network sees as context only), + and without the extra the embedder has 4 channels and the parameter count is the reference 1,508.""" + from corenet.models.network import CoReNet + + path = _need(DATA / "hh4b_small.h5") + spec4, spec5 = ( + parse_config(_hh4b_cfg(["pt", "eta", "phi", "mass"])), + parse_config(_hh4b_cfg(["pt", "eta", "phi", "mass", "btag"])), + ) + m4, m5 = CoReNet(spec4, device="cpu"), CoReNet(spec5, device="cpu") + assert m4._extras("other") == [] and m5._extras("other") == ["btag"] + assert m4.context["HH"].embed.embed.features == 4 + assert m5.context["HH"].embed.embed.features == 5 + n4, n5 = (sum(p.numel() for p in m.parameters()) for m in (m4, m5)) + assert n4 == 1508 and n5 > n4, (n4, n5) + x5 = H5Dataset(path, spec5).inputs + batch = {k: v[:512].clone() for k, v in x5.items()} + m5.train() + m5.updateMeanStd(batch) + m5.initMeanStd() + m5.eval() + m5.setGhostBatches(0) + with torch.no_grad(): + ref = m5(batch) + flipped = {k: v.clone() for k, v in batch.items()} + present = flipped["other"][:, -1] > 0 + assert present.any() + flipped["other"][:, 4][present] = ( + 1.0 - flipped["other"][:, 4][present] + ) # btag of the other jets + assert torch.equal(flipped["b"], batch["b"]) and torch.equal( + flipped["event"], batch["event"] + ) + with torch.no_grad(): + out = m5(flipped) + again = m5(batch) + assert torch.isfinite(out["logits"]).all() + assert not torch.equal(out["logits"], ref["logits"]), ( + "the btag of the other jets is not used" + ) + assert not torch.equal(out["views"]["HH"], ref["views"]["HH"]) + assert torch.equal(again["logits"], ref["logits"]) diff --git a/tests/test_report.py b/tests/test_report.py new file mode 100644 index 0000000..9d4ffc5 --- /dev/null +++ b/tests/test_report.py @@ -0,0 +1,177 @@ +"""corenet-report on 2-epoch synthetic trainings (CPU): the HTML file exists, is small, holds every section, +renders the pairing accuracy when the file has TARGETS and the overtraining check when corenet-evaluate +predictions of every event exist, and leaves those figures out when they do not.""" + +from __future__ import annotations + +import re +import tempfile +from pathlib import Path + +import h5py +import numpy as np +from synthetic import write_synthetic +from test_train import BBWW, HH4B, _train + +from corenet import evaluate, report + +HEADINGS = [f"{i + 1}. {t}" for i, (_, t) in enumerate(report.SECTIONS)] +MAX_BYTES = 4_000_000 + + +def _add_targets(path: Path) -> None: + """TARGETS/h1/{b1,b2} and TARGETS/h2/{b1,b2}: four distinct present jet slots for signal events, -1 otherwise.""" + rng = np.random.default_rng(1) + with h5py.File(path, "a") as f: + label = f["EVENT/label"][...] + n_present = f["INPUTS/Source/MASK"][...].sum(1) + cols = { + k: np.full(len(label), -1, dtype=np.int64) + for k in ("h1/b1", "h1/b2", "h2/b1", "h2/b2") + } + for i in np.flatnonzero(label == 1): + slots = rng.choice(int(n_present[i]), 4, replace=False) + for k, s in zip(cols, slots): + cols[k][i] = s + for k, v in cols.items(): + f.create_dataset(f"TARGETS/{k}", data=v) + + +def _evaluate_all(out_dir: Path, data: Path) -> None: + evaluate.main( + [ + "--checkpoint", + str(out_dir / "model_fold0.pkl"), + "--input", + str(data), + "--output", + str(out_dir / "eval_fold0.h5"), + "--fold", + "-1", + ] + ) + + +def _report(out_dir: Path, data: Path, *extra: str) -> str: + out = report.main( + [ + "--run-dir", + str(out_dir), + "--input", + str(data), + "--output", + str(out_dir / "report.html"), + *extra, + ] + ) + assert out.is_file(), out + assert out.stat().st_size < MAX_BYTES, out.stat().st_size + page = out.read_text(encoding="utf-8") + assert len(HEADINGS) == 8, HEADINGS + assert re.findall(r"

    (.*?)

    ", page) == HEADINGS, ( + "the page must hold exactly the eight sections in order" + ) + low = page.lower() + assert "not produced" not in low, "the 'Not produced' list is back" + assert "wall time" not in low, "the wall-time column / note is back" + assert "could not be produced" not in page, "a section failed" + assert "data:image/png;base64," in page + return page + + +def test_report_with_targets_and_all_events() -> None: + with tempfile.TemporaryDirectory() as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + _add_targets(data) + out_dir = tmp / "run" + _train(data, out_dir) + _evaluate_all(out_dir, data) + page = _report(out_dir, data) + assert "Pairing accuracy of the argmax view of HH" in page + assert re.search( + r"mean ± sd[0-9.]+[0-9.]+[0-9.]+", page + ) + assert "test, train, validation" in page + assert "overtraining check" in page + assert "AUC of the main score" in page + assert "H1 ← b" in page and "HH context" in page and "event scalars" in page + assert "Role b (jets btag 4)" in page and "Single object" not in page + + +def test_report_without_targets_and_train_predictions() -> None: + with tempfile.TemporaryDirectory() as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + out_dir = tmp / "run" + _train(data, out_dir) + page = _report(out_dir, data, "--max-events", "500") + # hypotheses: view probabilities and argmax shares, but no pairing accuracy (no TARGETS) + assert "Highest view probability per event and class on the test folds" in page + assert "Weighted share of the argmax view, test folds: HH" in page + assert "Pairing accuracy" not in page and "against TARGETS" not in page + # classification: AUC on the test fold only, no train / validation curves, no overtraining overlay + assert re.search( + r"mean ± sd[0-9.]+--", page + ) + assert "overtraining check" not in page + assert "test" in page and "test, train, validation" not in page + low_page = page.lower() + assert "pairing accuracy" not in low_page + assert "No predictions on the" not in page + assert "random subsample" in page + + +def test_report_bbww_multiclass() -> None: + from corenet import train + + with tempfile.TemporaryDirectory() as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "bbww_all.h5", "bbww") + out_dir = tmp / "run" + train.main( + [ + "--config", + str(BBWW), + "--input", + str(data), + "--output-dir", + str(out_dir), + "--fold", + "0", + "--exclude-folds", + "1", + "--epochs", + "2", + "--batch-size-init", + "128", + "--freeze-epoch", + "1", + "--device", + "cpu", + ] + ) + _evaluate_all(out_dir, data) + page = _report(out_dir, data) + assert "Confusion matrix" in page + assert "head ttbar" in page + assert "Weighted share of the argmax view, test folds: tt" in page + assert "Single object lepton" in page and "Single object met" in page + assert "relation lepton - met" in page + + +def test_report_needs_a_run_directory() -> None: + import pytest + + with tempfile.TemporaryDirectory() as tmp: + with pytest.raises(SystemExit): + report.main( + [ + "--run-dir", + tmp, + "--input", + str(HH4B), + "--output", + str(Path(tmp) / "r.html"), + ] + ) diff --git a/tests/test_train.py b/tests/test_train.py index 8d02438..b4d8148 100644 --- a/tests/test_train.py +++ b/tests/test_train.py @@ -7,12 +7,15 @@ 4. the checkpoint is self-contained and corenet-evaluate reproduces the training-time predictions, 5. the `trainer:` section of the configuration sets the recipe and the command line overrides it, 6. the fine-tuning stage (finetune_layers / finetune_lr / finetune_batch_size) and its defaults, - 7. the fold rule (--n-folds, --exclude-folds next) and its record in checkpoint and predictions. + 7. the fold rule (--n-folds, --exclude-folds next) and its record in checkpoint and predictions, + 8. the guards: ghost batches divide the fine-tuning batch size, finetune_layers prefixes must match, + early stopping needs a validation AUC, a file without folds trains not but evaluates and exports. """ from __future__ import annotations import json +import shutil import tempfile from pathlib import Path @@ -176,6 +179,31 @@ def check_freeze(config: Path, expected_params: int) -> None: assert 0 < n_trainable < expected_params +def test_batch_schedule_divides_any_batch_size() -> None: + """The ghost-batch count of the regular schedule divides the batch size for any --batch-size-init.""" + assert [train.batch_schedule(e, 300) for e in (0, 1, 3, 6, 10)] == [ + (300, 30), + (600, 15), + (1200, 4), + (2400, 1), + (4800, 0), + ] + for bs_init in (64, 100, 300, 1000, 1500): + for e in range(12): + bs, n_gb = train.batch_schedule(e, bs_init) + assert n_gb == 0 or ( + bs % n_gb == 0 and bs // n_gb >= train.MIN_GHOST_BATCH_SIZE + ), (bs_init, e, bs, n_gb) + for k in range(6, 13): # powers of two: unchanged from the original schedule + bs_init = 1 << k + for e in range(12): + bs, n_gb = train.batch_schedule(e, bs_init) + stage = 0 if e == 0 else 1 if e < 3 else 2 if e < 6 else 3 if e < 10 else 4 + assert n_gb == min( + (64, 16, 4, 1, 0)[stage], max(1, bs // train.MIN_GHOST_BATCH_SIZE) + ) + + def test_batch_schedule() -> None: assert [train.batch_schedule(e, 1024) for e in (0, 1, 3, 6, 10)] == [ (1024, 64), @@ -445,6 +473,168 @@ def test_finetune_lr_and_batch_size() -> None: assert train.finetune_batch(0, 1024, 4) == (4, 1) +def test_finetune_batch_size_divisor() -> None: + """The ghost batches of the fine-tuning stage divide the batch size (the GBN layers split a batch into + equal ghost batches): 300 events get 15 (epochs 1-2, cap 16) or 30 (epoch 0, cap 64) ghost batches of + >= 8 events, powers of two keep the schedule, and training with --finetune-batch-size 300 runs.""" + assert train.ghost_batches(300, 16) == 15 and train.ghost_batches(300, 64) == 30 + assert train.ghost_batches(300, 4) == 4 and train.ghost_batches(300, 1) == 1 + assert train.ghost_batches(300, 0) == 0 and train.ghost_batches(7, 64) == 1 + assert train.ghost_batches(301, 16) == 7 # 301 = 7 * 43 + assert train.finetune_batch(1, 1024, 300) == (300, 15) + assert train.finetune_batch(0, 1024, 300) == (300, 30) + assert train.finetune_batch(3, 1024, 300) == (300, 4) + assert train.finetune_batch(10, 1024, 300) == (300, 0) + for epoch, cap in ((0, 64), (1, 16), (3, 4), (6, 1), (10, 0)): + for bs in (64, 128, 256, 512, 1024, 2048, 4096): + old = min(cap, max(1, bs // train.MIN_GHOST_BATCH_SIZE)) if cap else 0 + assert train.finetune_batch(epoch, 1024, bs) == (bs, old), (epoch, bs) + assert bs % max(old, 1) == 0 + with tempfile.TemporaryDirectory(prefix="corenet_ft_") as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + h = _train(data, tmp / "m", "--finetune-batch-size", "300") + e0, e1 = h["history"] + assert e0["batch_size"] == 128 and e0["ghost_batches"] == 16 + assert e1["batch_size"] == 300 and e1["ghost_batches"] == 15 + assert np.isfinite(e1["train_loss"]) and 0.0 <= e1["val_auc"] <= 1.0 + + +def test_finetune_layers_must_match() -> None: + """finetune_layers prefixes that match no parameter at all would freeze everything: an error naming them + and the prefixes that exist, raised before the first epoch. One unmatched prefix among matching ones is + allowed (the default `head_reduce.` / `aux_heads.` match nothing in the HH -> 4b network).""" + with tempfile.TemporaryDirectory(prefix="corenet_ft_") as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + with pytest.raises( + ConfigError, + match=r"finetune_layers: \['nosuch\.'\] match no trainable parameter.*'head_main\.'.*'none' freezes nothing", + ): + _train(data, tmp / "bad", "--finetune-layers", "nosuch.") + assert not (tmp / "bad" / "history_fold0.json").exists() + h = _train(data, tmp / "mixed", "--finetune-layers", "head_main.", "nosuch.") + assert h["finetune_trainable_parameters"] == 18 # head_main of the 4b network + model = CoReNet(load_config(HH4B), device="cpu") + train.check_finetune_layers(model, train.HEAD_PREFIXES) + train.check_finetune_layers(model, ["head_main.", "nosuch."]) + with pytest.raises( + ConfigError, match=r"\['nosuch\.', 'typo'\] match no trainable parameter" + ): + train.check_finetune_layers(model, ["nosuch.", "typo"]) + with pytest.raises(ConfigError, match="nosuch"): + train.freeze_heads(model, ["nosuch."]) + assert train.count_trainable(model) > 0 # nothing was frozen by the failed call + + +def test_early_stopping_needs_validation() -> None: + """--early-stop-patience stops with an error when the validation set is empty (--exclude-folds names the + held-out fold) or has no validation AUC (one class in the validation fold); without it both runs proceed.""" + with tempfile.TemporaryDirectory(prefix="corenet_es_") as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + with pytest.raises( + SystemExit, + match=r"--early-stop-patience 1 needs validation events, but --exclude-folds \[0\] leaves none", + ): + _train( + data, + tmp / "empty", + "--exclude-folds", + "0", + "--early-stop-patience", + "1", + epochs=1, + ) + h = _train(data, tmp / "empty_ok", "--exclude-folds", "0", epochs=1) + assert h["n_val"] == 0 and "val_loss" not in h["history"][0] + one = tmp / "oneclass.h5" + shutil.copy(data, one) + with h5py.File(one, "a") as f: + label, fold = f["EVENT/label"][...], f["EVENT/fold"][...] + label[fold == 1] = 0 + del f["EVENT/label"] + f.create_dataset("EVENT/label", data=label) + with pytest.raises( + SystemExit, + match=r"epoch 0: no validation AUC, so --early-stop-patience cannot be applied: the validation folds \[1\] hold \{'rest': \d+, 'signal': 0\}", + ): + _train(one, tmp / "oneclass", "--early-stop-patience", "1", epochs=1) + h = _train(one, tmp / "oneclass_ok", epochs=1) + assert h["n_val"] > 0 and "val_auc" not in h["history"][0] + assert np.isfinite(h["history"][0]["val_loss"]) + + +def test_file_without_folds() -> None: + """A file with neither EVENT/fold nor --n-folds: training stops with a clear message, corenet-evaluate + needs --fold -1 (every event; the output carries fold -1) and corenet-export works.""" + from corenet import export + from corenet.data.h5 import FOLD_RULE_NONE, NO_FOLD + + with tempfile.TemporaryDirectory(prefix="corenet_nofold_") as tmp: + tmp = Path(tmp) + data = write_synthetic(tmp / "hh4b_all.h5", "hh4b") + _train(data, tmp / "m", epochs=1) + ckpt = tmp / "m" / "model_fold0.pkl" + nofold = tmp / "nofold.h5" + shutil.copy(data, nofold) + with h5py.File(nofold, "a") as f: + del f["EVENT/fold"] + with pytest.raises( + SystemExit, + match="no EVENT/fold dataset; training needs the fold of every event: pass --n-folds K", + ): + _train(nofold, tmp / "nofold_train", epochs=1) + assert not (tmp / "nofold_train").exists() + with pytest.raises( + SystemExit, + match="no EVENT/fold dataset, so fold 0 cannot be selected: pass --fold -1", + ): + evaluate.main( + [ + "--checkpoint", + str(ckpt), + "--input", + str(nofold), + "--output", + str(tmp / "e0.h5"), + ] + ) + metrics = evaluate.main( + [ + "--checkpoint", + str(ckpt), + "--input", + str(nofold), + "--output", + str(tmp / "eall.h5"), + "--fold", + "-1", + ] + ) + assert metrics["n_events"] == 2000 and 0.0 <= metrics["auc"]["main"] <= 1.0 + with h5py.File(tmp / "eall.h5", "r") as f: + assert f.attrs["fold"] == -1 and f.attrs["n_folds"] == 0 + assert f.attrs["fold_rule"] == FOLD_RULE_NONE + assert (f["EVENT/fold"][...] == NO_FOLD).all() and len( + f["EVENT/event"] + ) == 2000 + meta = export.main( + [ + "--checkpoint", + str(ckpt), + "--input", + str(nofold), + "--output-dir", + str(tmp / "export"), + "--n-events", + "256", + ] + ) + assert set(meta["files"]) == {"torchscript", "onnx"} + assert (tmp / "export" / "model_fold0.pt").exists() + + def test_finetune_recipe_from_config() -> None: """The three fine-tuning keys in the trainer: section, and their validation.""" cfg = yaml.safe_load(HH4B.read_text()) @@ -591,7 +781,10 @@ def test_n_folds_rule() -> None: and meta["fold_rule"] == FOLD_RULE_COLUMN == "EVENT/fold" ) assert meta["exclude_folds"] == [0] - with pytest.raises(KeyError, match="no EVENT/fold dataset"): + with pytest.raises( + SystemExit, + match="no EVENT/fold dataset, so fold 4 cannot be selected: pass --fold -1", + ): evaluate.main( [ "--checkpoint", diff --git a/tests/test_validate.py b/tests/test_validate.py index cf26007..0530a2d 100644 --- a/tests/test_validate.py +++ b/tests/test_validate.py @@ -294,6 +294,15 @@ def _replace(f, name, data): lambda f: f.__delitem__("EVENT/fold"), "EVENT/fold not found: corenet-train needs a fold index", ), + ( + lambda f: _replace( + f, + "EVENT/fold", + np.where(f["EVENT/fold"][...] == 2, 7, f["EVENT/fold"][...]), + ), + "EVENT/fold: fold ids [0, 1, 3, 4, 7] are not 0..K-1 (the ids of K folds must be exactly 0..K-1; " + "--exclude-folds next = fold (k + 1) % K): renumber the column or pass --n-folds K", + ), ( lambda f: _replace(f, "EVENT/weight", np.full(400, np.nan, dtype=np.float32)), "EVENT/weight: 400 non-finite weights", @@ -404,3 +413,16 @@ def test_fold_report(tmp_path): "folds = squares(event) % 3, seed 0 (--n-folds): EVENT/event not found" in _errors(r) ) + + +def test_fold_ids_not_contiguous_with_n_folds(tmp_path): + """A column whose ids are not 0..K-1 is a FAIL by itself and a WARN with --n-folds (the column is ignored).""" + data = write_synthetic(tmp_path / "d.h5", "bbww", n=400) + with h5py.File(data, "a") as f: + _replace(f, "EVENT/fold", f["EVENT/fold"][...] + 1) + r = validate.validate(BBWW, data) + assert "EVENT/fold: fold ids [1, 2, 3, 4, 5] are not 0..K-1" in _errors(r) + r = validate.validate(BBWW, data, n_folds=5) + assert not r.errors + assert "fold ids [1, 2, 3, 4, 5] are not 0..K-1" in "\n".join(r.warnings) + assert "--n-folds 5 ignores the column" in "\n".join(r.warnings)