What is performance in comparison with original implementation?

[John1231983]

Great implementation. Could you provide the reproduce result that can use to compare with original implementation in CAFFE2? Thanks

[John1231983]

I think the first conv should be conv2d. Am I right?
The correct version likes

       self.spatial_conv = nn.Conv2d(in_channels, intermed_channels, kernel=3,
                                    stride=1, padding=1, bias=bias)
        self.bn = nn.BatchNorm2d(intermed_channels)
        self.relu = nn.ReLU()
        self.temporal_conv = nn.Conv3d(intermed_channels, out_channels, temporal_kernel_size, 
                                    stride=temporal_stride, padding=temporal_padding, bias=bias)

[yechanp]

I think it is okay.
It should be kept as conv3d. but it actually performs like conv2d because one of kernel size is 1.

[JinXiaozhao]

self.conv3 = SpatioTemporalResLayer(64, 128, 3, layer_sizes[1], block_type=block_type, downsample=True)
why downsample=True?input size = 64 output size =128,I can't understand.can you help me ?
Thanks! @irhum

[Litou1]

My finding is that it's actually slower than C3D with fp16.
With fp32, R2+1D is faster.

pytorch 1.3
cuda 10.2
cudnn 7.6.5

I think the newer cudnn is quite efficient in performing 3D convolution for fp16 inputs.