Hi, Thank you for making your code public.

1. I tried running eval.py by using your uploaded models (all_100000.pth) and (all_50000.pth) from the google drive link mentioned under pretrained models (Section 3). But I get the following errors.Looks like the uploaded models are not correct. Could you please provide the link to the correct models? I am not using the docker image.

python eval.py --n_sample 100 --dist pretrained_models/

Traceback (most recent call last): File "eval.py", line 68, in net_ig.load_state_dict(checkpoint['g'])

RuntimeError: Error(s) in loading state_dict for Generator: Missing key(s) in state_dict: "feat_1024.1.weight_orig", "feat_1024.1.weight", "feat_1024.1.weight_u", "feat_1024.1.weight_orig", "feat_1024.1.weight_u", "feat_1024.1.weight_v", "feat_1024.2.weight", "feat_1024.2.bias", "feat_1024.2.running_mean", "feat_1024.2.running_var".

size mismatch for to_big.weight_orig: copying a param with shape torch.Size([3, 16, 3, 3]) from checkpoint, the shape in current model is torch.Size([3, 8, 3, 3]). size mismatch for to_big.weight_v: copying a param with shape torch.Size([144]) from checkpoint, the shape in current model is torch.Size([72]).

2. I also tried to train the model, but again encountered errors. Which pytorch version are you using?

File "FastGAN-pytorch/operation.py", line 16, in InfiniteSampler yield order[i] IndexError: index -1 is out of bounds for axis 0 with size 0

Thank you for your help.

Why is the loss D for fake images min(0,-1-d(x_fake)) not min(0,-d(x_fake))? I feel like min(0,-1-d(x_fake)) can try to converge to -1 when the D is fed fake images. I thought the ideal output of D from real images to be 1 and from fake images to be 0.

Hello! Thanks for your beautiful work. I noticed that in your network, there are BatchNorms in the whole network. So I confused why you do not use the InstanceNorm since InstanceNorm is better than BN in the image synthesis task?

Thank you for sharing your code; I really appreciate the clarity and simplicity in your implementation.

The "Nature Photograph" dataset which you use in the paper is not in provided few-shot-dataset.zip. Are there any copyright issues with the data, or could you also share it? If not, could you provide instructions on how to get the data? This would be great for reproducibility/comparison.

Thanks!

Dear odegeasslbc, how to continue the training? For example, yesterday I had stopped the training at iter 25000. And today I wan to continue the training by following commands: python train.py --name myImages --path "/myImages" --batch_size 16 --start_iter 25001 and it looks that "--start_iter 25001“ doesn't work. it still start the training from iter 0.

thanks.

Hello, First of all, congrats for your nice work. Just to know what is the meaning of the saved models. I have trained with 50000 iterations and for each 10000 iterations a model is saved. Hence, we have 50000.pth and all_50000.pth. Whats is the difference between them when generating new fake samples? Best Regards.

I have a question w.r.t SEblcok.

Do you have any specific reason why bias option is set to "False" in conv layer of SEblock?

I know that the FastGAN and MobileNet also have the same option.

Thank you.

According to the formula of your loss function, the loss of G is the opposite of the score given by D. Since D will only give a positive number, why does the loss of G need to be added with a negative sign? What makes me even more strange is that when the G loss is printed in the code, a negative sign is added, but when I run it, the output is a negative number, indicating that the loss of G should be a positive number, but according to the formula, it cannot be a positive number. if iteration % 100 == 0: print("GAN: loss d: %.5f loss g: %.5f " %(err_dr, -err_g.item()))

In the eval.py code, why you annotated the net.ig.eval() ? So you didn't use net_ig.eval(). But when we do the test, we would load the model and then do the model.eval().

Hi. Thanks for your excellent research work. I have a question why the code use skip-layer channel-wise excitation to extracts feat_64, but the paper does not?

feat_64 = self.se_64( feat_4, self.feat_64(feat_32) )

According to the download address of the datasets you provided, there was an error when decompressing the dataset after downloading it. Is there any other way to share the dataset? Thank you very much!

Hi,

This is a great and neat implementation. I was wondering if there is a simple way to implement conditioning (on image label) like a conditional GAN. I did try the standard approach used in cGAN but it was not successful. Could you please point me in the right direction?

Cheers,

Hi @odegeasslbc , Thank you so much for your awesome work. Can you give some insights into why using the percept.sum() instead of percept.mean() ?

Thanks!

Only a few of the datasets presented in the paper have pre-trained models in google drive! Any plan to release the other ones or at least release the hyperparameters?

I think the err_dr is just the loss of discriminator receiving real samples.

I think the err_dr is just the loss of discriminator receiving real samples. We should also add the loss of discriminator receiving fake samples. err_ discriminator_total=err_ real_samples+err_ fake_samples. Thank you!