Skip to content

Mismatch between test set metrics during training and evaluation #3083

Description

@dimokaramanlis

Is there an existing issue for this?

  • I have searched the existing issues

Operating System

Windows 10

DeepLabCut version

3.0.0rc10

What engine are you using?

pytorch

DeepLabCut mode

multi animal

Device type

NVIDIA RTX A4000

Bug description 🐛

To DLC developers,

I have been using DLC to track two mice (12 keypoints each) very successfully in the past using a resnet50 core with tensorflow.

Since moving to pytorch, I trained new models using the same datasets so that we can have longer-term support. In the latest version (rc10), I keep coming across a (large) mismatch between the metrics reported during model training and model evaluation.

For example (see file below), my model (dlcrnet) is being trained normally and both train and test set errors are decreasing. The model of epoch 150 is selected as "best" and has a test rmse of ~4.4. See below:

learning_stats.csv

When I'm running the evaluation later, the same model only gets a test rmse of 10, which is much worse than what is reported during training. Visual inspection of the predictions is also off. See below.

DLC_DlcrnetStride16Ms5_TwoMicePredictionAug29shuffle1_snapshot_best-150-results.csv

Do you have an idea of what may be the issue? I have never had a similar issue with tensorflow models, or pytorch models trained with the rc9 version. As as side note, model iinference with rc9 was much slower.

Thank you very much for the amazing software.

Best,
Dimos

Steps To Reproduce

No response

Relevant log output

Anything else?

No response

Code of Conduct

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions