Closed philgzl closed 1 year ago
Hi! You are absolutely right, we let the attention block in the bottleneck, that is a mistake that comes from the confusion between multiples experiments where we did indeed remove the bottleneck attention. We will correct if we are getting accepted. Thanks for reporting !
In the paper, when discussing the network architecture and comparing NCSN++M with NCSN++, it is said that "the attention layer in the bottleneck is removed". However, it seems to me that by setting
attn_resolutions = (0,)
here, no attention layers are used at all except in the bottleneck layer here. This is also the case in the previous sp-uhh/sgmse repo (here and here). So it seems to me the correct statement describing the new model should be "the attention layers are removed except in the bottleneck layer".