hi
I see some cgan code use y in the input before first layer.
So why do you use in block 3 and block 4 in SNResNetConcatDiscriminator?
and also , i am wonder why do you use a lot of y in gen?

Thanks!!

There are some implementation wrong.
Point out that the place to put the condition is different.

Add condition to output after convolution such like,

Unofficial(this repository)

• https://github.com/crcrpar/pytorch.sngan_projection/blob/master/models/discriminators/snresnet.py#L92
• https://github.com/crcrpar/pytorch.sngan_projection/blob/master/models/discriminators/snresnet64.py#L92

Official

• https://github.com/pfnet-research/sngan_projection/blob/master/dis_models/snresnet_32.py#L29
• https://github.com/pfnet-research/sngan_projection/blob/master/dis_models/snresnet_256.py#L36

thanks.

Hello

May i have one question?

utils.py > line 68~69
if distribution is None:
distribution == 'normal' -> I think it maybe " distribution = 'normal' "

if i'm wrong, please tell me what means ==

Thanks for your nice code :)

• SAGAN

17K iteration

Trained 20K iterations

Some modes are lost...?

class ID 76

class ID 104

class ID 126

Due to the lack of GPUs, I ran this on 1080Ti for 12 hours (around 20K iterations).
The dataset is tiny-ImageNet. The original images are themselves a bit difficult to tell what is in each image.

Config

same as the paper

Results

After 20K iteration.

Randomly sampled class outputs

ID 20

ID 181

ID 195

cGAN with concat discriminator
hmm

Is there any training or testing results?

Hi, train 64x64 dataset with your code . But it failed , and I test something ,like D network and loss
Here is my result:

1. sn_projection Discriminator is good, and it work fine. but wrong use in code when computed dis_fake which I think input of D should be true_label not pesudo_y . you can try it , pesudo_y is not useful at all.

2. Hinge loss have some trouble and I was using LSGAN loss is work fine!

I found some differences in the networks from https://github.com/pfnet-research/sngan_projection . Please check below.

pytorch.sngan_projection/models/discriminators/snresnet.py

activation=activation, downsample=False)

pytorch.sngan_projection/models/discriminators/resblocks.py

h_ch = in_ch

self.c_sc = utils.spectral_norm(nn.Conv2d(in_ch, out_ch, 1, 1, 0))

return F.avg_pool2d(x, 2)

pytorch.sngan_projection/models/generators/resnet.py

self.conv7 = nn.Conv2d(num_features, 3, 3, 1, 1)

Hi, I found a bug in fid score.

The normalization before inception model is written as:

but it should be:
https://github.com/mseitzer/pytorch-fid/blob/4e366b2fc9fb933bec9f6f24c5e87c3bd9452eda/inception.py#L130-L131

or it should be fixed like:

if self.normalize_input:
    mean = torch.tensor((0.485, 0.456, 0.406)).view(1, 3, 1, 1).cuda()
    std = torch.tensor((0.229, 0.224, 0.225)).view(1, 3, 1, 1).cuda()
    x = (x - mean) / std

not relativistic

16K iteration

After 20K iterations

Pointed out in #20

~Should be closed by #19 ~
Create a new one to fix it.

I type python train_64.py --no_tensorboard and run it...
I see the 'Initialized models...' print out and then I got a RuntimeError: CUDA error: initialization error. I am pretty sure my pyTorch version is '1.1.0'. Welcome to any solution. Thank you.

Following is the wrong feedback I got:

Traceback (most recent call last): File "train_64.py", line 402, in <module> main() File "train_64.py", line 341, in main real, y = sample_from_data(args, device, train_loader) File "train_64.py", line 199, in sample_from_data real, y = next(data_loader) File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 582, in __next__ return self._process_next_batch(batch) File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 608, in _process_next_batch raise batch.exc_type(batch.exc_msg) RuntimeError: Traceback (most recent call last): File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 99, in _worker_loop samples = collate_fn([dataset[i] for i in batch_indices]) File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/_utils/collate.py", line 68, in default_collate return [default_collate(samples) for samples in transposed] File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/_utils/collate.py", line 68, in <listcomp> return [default_collate(samples) for samples in transposed] File "/vm/data/rdata/shared/conda_envs/fastaiv1/lib/python3.6/site-packages/torch/utils/data/_utils/collate.py", line 59, in default_collate return torch.tensor(batch)

s/gen/dis/g

gen_* loss ignores the dis_real term

Collapsed.