No Predictions with a Bottom Up-ID model

[cece3014]

Hi SLEAP team!
I have been working in SLEAP the last few months and am truly amazed by the results of the bottom-up model I have been able to train (thank you to @gitttt-1234 for so much guidance). For my particular dataset that is divided into age classes, I would love to compare how the bottom up ID model performs in relation to a pose-estimation only model. When i ran a training with IDs attached to the instances, I got 0 predictions. Perhaps there are different model hyperparameters that would be optimal for an ID model that are different than a bottom-up model?

I would appreciate any insight you have, thank you again.

Example image of dataset
(all of these predictions were made using the bottom-up model, not an ID mdoel)

Model Hyperparameters for bottom-up model

Model Hyperparameters for bottom-up ID model

[gitttt-1234]

Hi @cece3014 !

Could you let us know how the training curves looked like? you could share the training_log.csv here so that we could view the loss curves. Based on your config, I’d suggest trying the following changes: Enable rotation augmentations — without them, the model can overfit to the exact poses seen during training. In the model config, check the up-interpolate option and in the heads section, set the second sigma to 15.0 and the second loss weight to 2.0. Give these a try and let us know if the predictions improve!

Thanks,

Divya

[cece3014]

Hi @gitttt-1234

Good morning - Absolutely! You mean the training_log for the bottom_up_IDmodel correct?

Is it only possible to create one ID model? I wanted to attempt a new modl using the parameters suggested but I can only select the previous model that produced 0 results.

Thank you.

[gitttt-1234]

Hi @cece3014 !

You don't have to choose anything in the drop down, but just update the values that should work!

Thanks,

Divya

[cece3014]

Hi @gitttt-1234,

I hope your holiday season was eventful! I adjusted the model configurations as suggested for training but unfortunately had 0 predictions. Are there other options to explore?

[gitttt-1234]

Hi @cece3014 !

Apologies for the delay!

How many epochs did the model train for and how did the training curves look like? Could you share the training_log.csv here? If the model has only trained for very few epochs, then try adjusting the early stopping (set delta as 1e-8 / increase the patience) as training with multiple heads can be little unstable and the default parameters might not work! In your previous model, you set input scaling: 0.25 and output stride: 4. Give it a try with these same parameters and let us know if that works!

Thanks,

Divya

[cece3014]

Hi @gitttt-1234 ,

Not a problem at all, thank you for your suggestions! I uploaded the previous bottom-up ID model's training_log csv to the link provided. In the meantime I ran another bottom-up ID with theses adjustments, but still no predictions. Would you suggest to continue to adjust early stopping?

Thank you so much.