search

facebookresearch / VMZ

VMZ: Model Zoo for Video Modeling
Apache License 2.0
1.04k stars 157 forks source link

Parameter epoch_size greater than dataset size in train phase #9

Closed murilovarges closed 6 years ago

murilovarges commented 6 years ago

Hi,

In the paper, you said "Although Kinetics has only about 240k training videos, we set epoch size to be 1M for temporal jittering".

I would like to know what will happen, the video reader will start from the begin of dataset? In this case will process the dataset about 4 times per epoch, in each time get a different clip for each video using temporal jittering?

I tried to find this information reading source codes in the folder pytorch/caffe2/video/. but I remain with doubt.

Thank's Murilo.

dutran commented 6 years ago

Yes, it uses temporal jitter, which is implemented here https://github.com/pytorch/pytorch/blob/master/caffe2/video/video_input_op.h#L235

murilovarges commented 6 years ago

Thank you @dutran,

In https://github.com/pytorch/pytorch/blob/master/caffe2/core/db.h#L232 If the cursor reaches its end, the reader will go back to the head of the db.

dutran commented 6 years ago

Yes, the reader will make about 4 times over the list, but due to temporal jittering, it is likely to see a different clip even with the same video.

cr-hxfan commented 5 years ago

Hi, @dutran May I know how much the accuracy improved by temporal jittering? Did you test it before? Thanks in advance.