search

selimsef / dfdc_deepfake_challenge

A prize winning solution for DFDC challenge
MIT License
787 stars 209 forks source link

32 frames are selected continuously or randomly ？ #3

Closed miaoct closed 4 years ago

miaoct commented 4 years ago

Could you talk about it in detail ? Thank you !

miaoct commented 4 years ago

And you say "Though only 400 videos can be used for validation as well." How to get tags for 400 videos and how to use them for verification ? Thank you !

selimsef commented 4 years ago

You can have a look at the code for frame extraction here https://github.com/selimsef/dfdc_deepfake_challenge/blob/e6038f17fb16c9ee93f216a9292392158826d53d/kernel_utils.py#L53

selimsef commented 4 years ago

Regarding 400 videos I hardcoded their names. But originally (for the 4 models out of 5) I used all videos from the 0-2 folders, it just takes more time. Left it like that to be reproducible. Can be changed to use public validation set - not sure how it would affect final scores though. https://github.com/selimsef/dfdc_deepfake_challenge/blob/e6038f17fb16c9ee93f216a9292392158826d53d/training/datasets/classifier_dataset.py#L362