search

moabitcoin / ig65m-pytorch

PyTorch 3D video classification models pre-trained on 65 million Instagram videos
MIT License
265 stars 30 forks source link

Port missing r(2+1)d 34-layer IG65-M model #3

Closed daniel-j-h closed 5 years ago

daniel-j-h commented 5 years ago

Blocked by official weights not being released so far:

https://github.com/facebookresearch/VMZ/issues/87

sandhawalia commented 5 years ago

The model download link is back up seems to be back up. The porting scripts are failing

python convert.py --frames 32 --classes 487 /nas/3rd_party/fair/video-model-zoo/VMZ-models-6c925c47b7d6545b64094a083f111258b37cbeca/r2plus1d_34_clip32_ig65m_from_scratch_f102649996.pkl /nas/team-space/experiments/video-ig65m/converted-models/r2plus1d_34_clip32_ig65m_from_scratch
Traceback (most recent call last):
  File "convert.py", line 204, in <module>
    main(parser.parse_args())
  File "convert.py", line 178, in main
    copy_fc(model.fc, blobs)
  File "convert.py", line 70, in copy_fc
    copy_tensor(module.bias.data, blobs, "last_out_L" + str(n) + "_b")
  File "convert.py", line 43, in copy_tensor
    tensor = torch.from_numpy(blobs[name])
KeyError: 'last_out_L487_b'
sandhawalia commented 5 years ago

Update. r2plus1d_34_clip32_ig65m is trained on 359 classes. Needs 359 as an option for --classes in convert.py

sandhawalia commented 5 years ago

https://github.com/moabitcoin/ig65m-pytorch/pull/10