Upstream our changes to PyTorch r(2+1)d architecture

moabitcoin / ig65m-pytorch

PyTorch 3D video classification models pre-trained on 65 million Instagram videos

MIT License

265 stars 30 forks source link

Upstream our changes to PyTorch r(2+1)d architecture #5

Closed daniel-j-h closed 5 years ago

daniel-j-h commented 5 years ago

The torchvision r(2+1)d architecture needs two modifications to get it in sync with the official Caffe2 implementation (see https://github.com/facebookresearch/VMZ/issues/89) and our provided code:

Number of midplanes in the downsampling blocks
Batchnorm

We should upstream both modifications to torchvision.

bjuncek commented 5 years ago

see https://github.com/pytorch/vision/issues/1265

need some discussion, but is easily done

daniel-j-h commented 5 years ago

@bjuncek great, this would be sweet to have upstream! :heart: Please consider adapting the Batchnorm blocks, too. Then the architectures are 100% in sync and weights can easily be transfered.

daniel-j-h commented 5 years ago

Closing this ticket since there is nothing actionable on our end. We will follow up with you folks on vmz to upstream some of our changes when we ported the weights.