williamFalcon / pix2pix-keras

Image-to-Image Translation with Conditional Adversarial Networks (Pix2pix) implementation in keras
MIT License
140 stars 50 forks source link

Code compilation error, DCGAN, UNETGenerator incompatible shapes #4

Open hoyleb opened 6 years ago

hoyleb commented 6 years ago

Hi!

I'm trying to get main.py to run, fresh out the box, after I've set up the conda environment as suggested.

In generator.py UNETGenerator I get errors that the combined net works are not compatible size, so I have to resize each decoder to have the same size as the encoder layer, e.g. from de_6 = Convolution2D(nb_filter=512, nb_row=4, nb_col=4, border_mode='same')(de_6) to de_6 = Convolution2D(nb_filter=128, nb_row=4, nb_col=4, border_mode='same')(de_6)

otherwise this fails: de_6 = merge([de_6, en_2], mode=merge_mode, concat_axis=1)

Then once adjusted I get an error with DCGAN

telling me: "ValueError: Dimensions must be equal, but are 1 and 64 for 'Conv2D_168' (op: 'Conv2D') with input shapes: [?,510,64,1], [4,4,64,64]."

Could you suggest what I'm doing wrong?

Thanks,

Ben