weights initialization

[zeal-up]

Hi, thanks for your job to reproduce S3D model. I have some questions both of the model. The original paper has no mentioned about the details of initializing the model's weights from a ImageNet pretrained model, so I'm confused about how to transfer weights from a 2D conv to a 2D conv + 1D conv? For I can imagine is that we average a 2D kernel and replicate the value along the temporal dimension to construct 1D kernel(kx1x1 convolution in the S3D model).

Another question is, in your README, you say that you initialize the reproduced model with pretrianed I3D, it seems that you only transfer pretrained weights to 2D kernel? Why don't you also transfered to 1D temporal kernel use a similar way?