plotted points are not matching what is known from DLC `analyze_videos`

[MMathisLab]

plotted points are not matching what is known from DLC analyze_videos: @AlexEMG this is the newly exported model, fyi.

same video:

Originally posted by @MMathisLab in #1 (comment)

Moreover, this is true for all model_zoo models.

This is using the default model_export function, so this is crucial that it must work.

[MMathisLab]

you can test this with any model zoo, simple download by:

import deeplabcut
import urllib.request
import tarfile
from tqdm import tqdm
from deeplabcut.utils import auxiliaryfunctions
modelname = 'full_cat'
deeplabcut.utils.auxfun_models.DownloadModel(modelname, '/content/')

[MMathisLab]

@gkane26 can you elaborate how you've tried to fix this? WE said try the non-pb files, right, but how do I point to only those, as in dlclive.anv you point to a folder

[gkane26]

I haven't yet tried loading the snapshot file -- can't do that in DLC-live because you need a lot of DLC functions to recreate the network, then load the weights. I can try doing so this afternoon, but what that amounts to is essentially just running DLC's analyze_videos. I think the issue has to be that we're either choosing the wrong input/output nodes in DLC export.py or in dlclive.graph.py. I've tried some different combinations of but I'm pretty sure it's using the correct ones now

[MMathisLab]

I think @AlexEMG is looking into the TF graph business now

[sneakers-the-rat]

I think this needs a fix! I was experiencing this all day today and had resigned myself to thinking I was doing something wrong -- the numbers from dlc proper (specifically those in the plot.png generated from plot trajectories) were pretty consistently wrong from the numbers dlclive was giving across multiple models and retraining even when my camera and imaging plane were firmly clamped in place and unchanging. I think it might be an indexing error

[MMathisLab]

It was fixed in the PR #1 from GK, see discussion