search

sign-language-processing / transcription

Text to pose model for sign language pose generation from a text sequence
35 stars 16 forks source link

Complete test metrics and plots, tune learning rate, etc. #14

Closed J22Melody closed 1 year ago

J22Melody commented 1 year ago

Confusion matrix, pr curve, and probs plots:

https://wandb.ai/zifanjiang/pose-to-segments/runs/tvxti7jt?workspace=user-j22melody

Tune learning rate - seems a bigger learning rate helps, but ReduceLROnPlateau does not:

Tune learning rate on overfitting dev data:

https://wandb.ai/zifanjiang/pose-to-segments/runs/5uy5gwq0?workspace=user-j22melody https://wandb.ai/zifanjiang/pose-to-segments/runs/1jut6t98?workspace=user-j22melody

J22Melody commented 1 year ago

@AmitMY ahh I forgot to save the PR's descriptions, now you see it, yes and the figures on wandb.

These newly added figures (confusion matrix, pr curve, probs plots) are only generated in the end once for the test data. I am not plotting them during training because that would be many. Do you wanna see how they change during training? One possible way would be to plot them every 10 epochs or so.

Now it also reminds me that I should probably not plot on test data and tune the model accordingly, I should use the dev data instead!