Pretrained network E

[nvidia]

Pretrained network E

[jthsieh]

Thank you so much! I just realized that since the latest_net_G you provided was trained without features, I can't use it with features. Do you also have a checkpoint for G trained with features? Thanks a lot!

[tcwang0509]

It can be found here: https://drive.google.com/file/d/1AV_2-ltcyER08yjhRaut1tbaMzfeQWfu/view?usp=sharing. We should also release the UI soon, once it gets approved.

[adynathos]

I would also be grateful for the pre-trained weights for netE and netG-with-features, that is the checkpoint for label2city_1024p_feat or label2city_512p_feat.

---

I tried the latest_net_E.pth provided in this thread, but it fails to load - I think it is for an older version of the architecture.
The file contains keys for model.28.weight and model.28.bias but that is now a ReflectionPad, and there are no weights for model.29.

...
(26): InstanceNorm2d(16, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
(27): ReLU(inplace)
(28): ReflectionPad2d((3, 3, 3, 3))
(29): Conv2d(16, 3, kernel_size=(7, 7), stride=(1, 1))
(30): Tanh()

Keys in latest_net_E.pth:

...
model.26.running_mean
model.26.running_var
model.28.weight
model.28.bias

---

Also the latest_net_G.pth for label2city_1024p is of wrong shape for using features:

While copying the parameter named "model.1.weight", whose dimensions in the model are torch.Size([64, 39, 7, 7]) and whose dimensions in the checkpoint are torch.Size([64, 36, 7, 7]).

---

Thank you.

Best wishes,
Krzysztof Lis