PyTorch implementation of fractored convolution in paper, A Closer Look at Spatiotemporal Convolutions for Action Recognition. [link]
from conv2plus1d import FactorizedConv3d
FactorizedConv3d
functions like torch.nn.Conv3d
and has the same APIs.
See example code in d2plus1d.py
, DenseNet with Conv(2+1)D in block 3, 4
. Also refer to a module test in conv2plus1d.py
.
- The torchsummary is used to test this module. I've modified it with PyTorch 1.0 flavor.