questions about ShiftModule used in BottleneckShift

[ZGCTroy]

Thanks for sharing the code !

I find the ShiftModule which uses a conv1d to shift like TSM. In your BottleneckShift, you add a mSEModule and ShiftModule, making it different from the normal Bottleneck.

Does the mSEMoudle refer to Long-term TDM in your paper? How about the ShiftModule, does the ShiftModule refer to the shift module of TSM.

'''
out = self.mse(out)
out = self.shift(out
'''
In your code, you simultaneously applied the mse and shift. Does this mean the TDM simultaneously use L-TDM and TSM shift ? Or you just set ShiftModule(mode = 'fixed') when running TDM.

[yztongzhan]

Hi @ZGCTroy, Thanks for your comments.

1. Yes! The mSEMoudle refer to Long-term TDM in your paper.
2. The ShiftModule is a 1D-Conv which is followed by TEA. you can refer the original TEA paper for more detail.
3. As shown in the figure 2 of our paper, we add the 1D-Conv after L-TDM. And, our baseline is ResNet with 1D temporal convolutions in latter stages for better temporal modeling.