search

piergiaj / pytorch-i3d

Apache License 2.0
971 stars 250 forks source link

Finetune on charades #14

Closed XudongLinthu closed 5 years ago

XudongLinthu commented 5 years ago

Hi, Thanks a lot for sharing your reimplementation of I3D. You mentioned you reproduced the results on Charades with their training methods. Could you please share your hyper-parameters setting? Or it is the default setting of train_i3d.py? I couldn't reproduce their results. Thanks a lot! Xudong

piergiaj commented 5 years ago

The learning rate was set to 0.1, then dropped to 0.01 after 45k steps. The batch size was set to 6 videos per-GPU, an aggregated over a total of 64 GPUs (I used 8 GPUs, for 8 steps each). The effective batch size that the gradients are accumulated over is 384. I trained for 65k steps.

The weight regularization was set to 1e-6 and batch norm was the same as in the files. Dropout was set to 0.5.

I believe that the settings in train_i3d are mostly correct, other than possibly the distributed training settings (which will depend on your system).

XudongLinthu commented 5 years ago

Thanks a lot!