Tensor size mismatch error for batch_size>1 - File:gb_encoder.py

[m-abdelkarim]

When the batch_size is set for any value >1, the following error is triggered:

" File "D:\Photorealism\PhotorealismEnhancement\code\epe\network\gb_encoder.py", line 137, in forward features += classmap[:,c,:,:] * self.class_encoders[c](\ RuntimeError: The size of tensor a (2) must match the size of tensor b (128) at non-singleton dimension 1"

The error takes place in the following line. As seen in the second screenshot, the shapes dont match. Even when I try to reshape the classmap[:,c,:,:] so that the multiplication works, an assert is thrown in the discriminators.py file as shown in the 3rd screenshot. Could anyone help in this regards?

[witchoco]

yes... I am facing this same issue... besides removing the constraints you listed above, there is some other slight modifications I've made: epe/network/discriminators.py, around line 164:

precomputed segmentation

                            #_, _, hy, wy = y.shape # original implementation for batch_size=1 constraint
                            #y = self.embedding(y.reshape(-1)) # original implementation for batch_size=1 constraint
                            #y = y.permute(1,0).reshape(1,c,hy,wy) # original implementation for batch_size=1 constraint
                            #changed as:
                            ny, _, hy, wy = y.shape
                            y = self.embedding(y.reshape(ny, -1))
                            y = y.permute(0,2,1).reshape(ny,c,hy,wy)

Now it's working... But I'm not sure if it's correct or any impact on effect... I'm not familiar with embedding... Could any body please share with us some explanation why the batch_size=1 constraint is there? Thank you guys so much~