Output of the discriminator is not 1-D

[JannisBush]

Hey,

In the README it is written that the last layer of the discriminator maps the output to a scalar.
So the output shape of the discriminator should be (?,1,1,1) or (?,1), but in fact it is (?,4,4,1).

Is this intended?

To see this, add self.dis_real = dis_real in line 199 and print(self.sess.run(self.dis_real, feed_dict=feed_dic).shape) in line 55 of models.py.
An easy fix for this problem, would be to use the padding="VALID" for the last layer of the discriminator.

Best regards,
Jannis

[knazeri]

@JannisBush This was intended! We have to update the readme!
This class of discriminators is known as PatchGAN. The output of the discriminator is a 4x4 matrix with each index corresponding an image patch in the input. The receptive field of the network is 70x70 which means our discriminator evaluates 70x70 overlapping patches of the input image!

Philip Isola has a good explanation for this: junyanz/pytorch-CycleGAN-and-pix2pix#39 (comment)

This is all mathematically equivalent to if we had manually chopped up the image into 70x70 overlapping patches, run a regular discriminator over each patch, and averaged the results.

[JannisBush]

@knazeri Thanks for your clarification.