python scripts/translate_images.py PATH_TO_TRAINED_MODEL -n 100 raise RuntimeError

[sonwe1e]

It raise RuntimeError when I run python ./uvcgan-main/scripts/translate_images.py ./uvcgan-main/outdir/selfie2anime/model_d\(cyclegan\)_m\(cyclegan\)_d\(basic\)_g\(vit-unet\)_cyclegan_vit-unet-12-self-lsgan-paper-cycle_high-256/ -n 100

And I found the shape of X and shape of embed is not the same

The position of error is

    def forward(self, x):
        # x     : (N, L, input_features)
        # embed : (1, height * width, embed_features)
        #       = (1, L, embed_features)
        embed = self.embed(self.y_const, self.x_const)
        print(x.shape)
        print(embed.shape)

        # embed : (1, L, embed_features)
        #      -> (N, L, embed_features)
        embed = embed.expand((x.shape[0], *embed.shape[1:]))
        print(x.shape)
        print(embed.shape)

        # result : (N, L, embed_features + input_features)
        result = torch.cat([embed, x], dim=2)

        # (N, L, features)
        return self.output(result)

Looking forward to your reply.

[usert5432]

Hi @sonwe1e, I think this error is caused by working on images of larger size. From the error message, it looks like the model was trained on images of size (256, 256) pixels, but it is evaluated on images of size (512, 512) pixels.

Would my assumption on images sizes be correct?

[sonwe1e]

Hi! Thanks for your reply. I found there is no resize in valid dataset, so I add resize augment to valid config to match the image size.