Max pooling layer is not at usual position

Hello,

The max pooling is after the second resnet stage in both Slowfast and Resnet architectures contrary to the architecture described within the Slowfast paper (and the original resnet paper). This can be seen for exemple [here](https://github.com/facebookresearch/SlowFast/blob/b168a41781cf71c5b33856420fdfd94e8f336e45/slowfast/models/video_model_builder.py#L407).

What is the reasoning behind that choice? Why not put the max pooling after `s1` as is being done in a traditional resnet?

Thanks