search

Tobias-Fischer / rt_gene

RT-GENE: Real-Time Eye Gaze and Blink Estimation in Natural Environments
http://www.imperial.ac.uk/personal-robotics
Other
374 stars 71 forks source link

Pytorch RT-BENE #91

Closed ahmed-alhindawi closed 3 years ago

ahmed-alhindawi commented 3 years ago

Hello,

I've pytorched RT-BENE for easier experimentation. Let me know what you think of this pull request.

ahmed-alhindawi commented 3 years ago

I added some comments inline. Bigger question though - have you tried/tested this? Are the results similar to the TF version? As far as I can see there is no code to do the 3-fold evaluation, which is not a big deal, but it would be good to know if the models that are trained are sensible.

Hey, yes it works and trains nicely. You're correct, there isn't any evaluation code yet; that's something to do this week but the fundamentals to perform 3-fold validation is there; the Trainer trains on specific train/validation subjects who are the same splits as the TF version.

Tobias-Fischer commented 3 years ago

Two tiny comments left, well done!

Tobias-Fischer commented 3 years ago

Great - ready to merge if you think it's ready.

ahmed-alhindawi commented 3 years ago

Great - ready to merge if you think it's ready.

Two things: 1 - I realised there's no way to perform inference with the models, so let me make that code today 2 - I'm going to change the evaluate_mode.py to a jupyter notebook to facilitate more data exploration.

ahmed-alhindawi commented 3 years ago

Ready if you're happy?

Tobias-Fischer commented 3 years ago

Many thanks again @ahmed-alhindawi! Nice job :)