Hi, what a wonderfull work!!!
When i was eval the network picture, IS value is up to 4.* without fine-tuned inception net in respective datasets.
Could u release the fine-tuned inceptionV3 network pth file on respective dataset(birds, cars, dogs)?
thx!!!

How to generate more images?

hello,
i'm interested in your work but i'm confused about how you transfer the bounding box to the bird mask.
after reading your code, i still cannot figure out the process. what does "warped_bbox" mean in your code?
looking for your reply
the related code is shown below

`x1 = self.warped_bbox[0][i]
x2 = self.warped_bbox[2][i]
y1 = self.warped_bbox[1][i]
y2 = self.warped_bbox[3][i]

                a1 = max(torch.tensor(0).float().cuda(), torch.ceil((x1 - self.recp_field)/self.patch_stride))
                a2 = min(torch.tensor(self.n_out - 1).float().cuda(), torch.floor((self.n_out - 1) - ((126 - self.recp_field) - x2)/self.patch_stride)) + 1
                b1 = max(torch.tensor(0).float().cuda(), torch.ceil((y1 - self.recp_field)/self.patch_stride))
                b2 = min(torch.tensor(self.n_out - 1).float().cuda(), torch.floor((self.n_out - 1) - ((126 - self.recp_field) - y2)/self.patch_stride)) + 1

                if (x1 != x2 and y1 != y2):
                        weights_real[i, :, a1.type(torch.int) : a2.type(torch.int) , b1.type(torch.int) : b2.type(torch.int)] = 0.0`

Hi! When i test your fine-tuned model file, it happens some errors:
Error(s) in loading state_dict for Inception3: Missing key(s) in state_dict: "AuxLogits.fc.weight", "AuxLogits.fc.bias", "fc.weight", "fc.bias". Unexpected key(s) in state_dict: "fc_new.weight", "fc_new.bias", "AuxLogits.fc_new.weight", "AuxLogits.fc_new.bias".
Could u mind send a fine-tuned inception model file(inceptionv2.py) for me, thx!!!

Other Question~, how to fine-tuned the inception model on bird datasets?

Hi, I found that you do this after every time you call train_Gnet function.

 for p, avg_p in zip(self.netG.parameters(), avg_param_G):
     avg_p.mul_(0.999).add_(0.001, p.data)

avg_param_G is the final parameters of your saved model. It looks like you update your model again after opt.step(). Why don't you save the final model directly but use this avg_param_G?

Could you explain this to me? Thank you!

In the paper it's mentioned that from the automatically learned representations , clustering is done on real images. Can you please show or direct,e to the example/code used for doing it?

Thanks for your work！
I can‘t download your formatted CUB data. So I download the original CUB data and resize images to 128×128 for train. I don't know whether it's right. Could you share your method of preparing dataset.

And another question, how do you get the real background images during background stage？ Because the discriminator D_b needs both real background images and fake background images.

Hi, thanks for releasing code. It's a great job. But I have some tiny questions about the classification score of background. In line 337 of model.py, the background score is set to 0
classi_score = self.uncond_logits1(x_code) # Background vs Foreground classification score (0 - background and 1 - foreground)
It's consistent with your statement in your paper. But in lines 227 and 330 of trainer.py, the score of background seems to be set to 1.
errD_real_uncond_classi = criterion(real_logits[0], weights_real) # Background/foreground classification loss
errG_classi = criterion_one(outputs[0], real_labels) # Background/Foreground classification loss for the fake background image (on patch level)
So, I am confused, could you point anything wrong about my understanding, that will be very helpful, thank you.

Thank you for sharing.
Did you try high res ver (like 256x or 512x)?
If not, what difficulties can be considered?

Could u pls upload the formatted CUB data and pretrained models to the Baidu Cloud(https://pan.baidu.com)? It's very hard for us to download large files in China. Sorry for that, it's just unstable to download something from Google driver. I don't know if it's convenient for u to enter this website (https://pan.baidu.com). If possible, pls upload it for us. Thank u very much.

Can you upload the pre trained cars model?

Hi, nice work!

I am evaluating your released model on Birds dataset. Particularly, I want to get the number claimed in your paper to make sure that everything I have done is correct. However, in your paper, it is claimed IS=52.53±0.45 and FID=11.25. But I got IS=43.20±0.54 and FID=22.08. I think I might have made some mistakes in details.

What I did

I generate 30K 128x128 child images with your released model on Birds, wherein 150 images for each child category.
(1) I compute the IS with your released finetuned inception model. The generated images are resized to 299x299 normalized to be within [-1, 1] before fed into the network. Mean and std are computed over 10 splits.
(2) I compute the FID with default inception model using

calculate_fid_given_paths([/path/to/generated/images, /path/to/real/images], batch_size=1, cuda=True, dims=2048)

Note that /path/to/real/images are original CUB images without any cropping or resizing.

Evaluation codes

IS: https://github.com/sbarratt/inception-score-pytorch
FID: https://github.com/mseitzer/pytorch-fid

Questions

1. Do you use the finetuned inception model for computing FID?
2. Should I use the images that are cropped with 1.5x bounding boxes from original images to compute FID?
3. Should I first resize the real images to 128x128 and then feed them into the inception network (which automatically resizes input to 299x299) to compute FID?
4. Is the number I got lies in the normal variation? I suppose the quality of generated images may be different for different times of generation.

B.T.W. I am also curious about the results of LR-GAN you got in Table1. Do you train LR-GAN on the original CUB images or on the cropped images?

B.T.W. I am using pytorch==1.3.0. Not sure if there is any version issue.

Thank you for sharing.
I have a question about how to calculate the fid score between 'birds' dataset. Did u use the fine-tuned inceoption network to calculate the fid score or use other official fid calculate function ?

Hi, thanks for publishing your code. But after I reading the code, it makes me confused that the parent discriminator seems not to be updated anymore after initialized. It seems to be different from the paper. Have I ever missed anything important?
Looking forward to your answer, thanks!

Im confused about your paper word 'We evaluate image generation using Inception Score (IS) [42] and Frechet Inception Distance (FID) [20], which are computed on 30K randomly generated images (equal number of images for each child code c)'
How can i generate 30k test image on CUB200 dataset? : )
Because in CUB200, parent class is 20, child class is 200, background class is 200.So that the final image is about 20x200x200= 800K. How can i get the 30K test image?

Also, how calc the FID value of CUB200 dataset? (e.g how to get the CUB200 mu and sigma?)

Hope for your reply~~~

Sorry, I just can't wait it. What a wonderful job!!!

Please aviod "inconsistent use of tabs and spaces in indentation"!