AttributeError: 'Distribution' object has no attribute 'dist_type'

[Ciccios96]

Hello, when i try to run the torchvision.utils.save_image() function inside the save_and_sample() function i get the error in the title. Can i get some help? This error is present in the Cifar10 train script.

[Gass2109]

The same issue with the version 0.8.2 of torchvision (it works properly with the version 0.4.2).

[universome]

The problem arises because grid variable is not a normal torch.Tensor, but rather a tweaked Distribution variable. To fix the issue, I had to tweak lines:

in train_fns.py: from

torchvision.utils.save_image(fixed_Gz.float().cpu(), image_filename,

to

torchvision.utils.save_image(torch.from_numpy(fixed_Gz.float().cpu().numpy()), image_filename,

in utils.py: add line:

out_ims = torch.from_numpy(out_ims.numpy())

after line:

out_ims = torch.stack(ims, 1).view(-1, ims[0].shape[1], ims[0].shape[2],
                                        ims[0].shape[3]).data.float().cpu()

What we basically do here is converting the tweaked Distribution tensor into normal torch.Tensor via torch.from_numpy(t.numpy()).

---

Also, JFYI, here are the scores I got for cifar-10 when launching the launch_cifar_ema.sh script:

• best IS = 8.375 at itr=95000
• best FID = 6.335 at itr=95000 The paper reported IS=9.22 and FID=14.73 in Appendix C.2