Some questions about MinibatchStatConcatLayer

[lyhangustc]

I have some questions about MinibatchStatConcatLayer. They are mostly about the shape of the vals in the code.
Let's say the shape of the input x is [b, c, h, w]

After this line:

PyTorch-progressive_growing_of_gans/models/base_model.py

Line 80 in 8337fc9

vals = self.adjusted_std(x, dim=0, keepdim=True)# per activation, over minibatch dim

the shape of vals should be [1, c, h, w].

1. For the case of "all ", we should get vals of shape [1, 1, 1, 1], since it is to "average everything --> 1 value per minibatch". However, this line
   

   PyTorch-progressive_growing_of_gans/models/base_model.py

   Line 83 in 8337fc9

   vals = torch.mean(vals, dim=1, keepdim=True)#vals = torch.mean(vals, keepdim=True)

   
   outputs vals of shape [1, 1, h, w].
   I think we should use vals = torch.mean(vals, keepdim=True) instead, which you have commented for an unknown reason.

2. What is the purpose of this line?
   

   PyTorch-progressive_growing_of_gans/models/base_model.py

   Line 89 in 8337fc9

   target_shape = [target_shape[0]] + [s for s in target_shape[1:]]

   
   It seems like 'target_shape = target_shape' and we still get [b, c, h, w].