Hi, I'm getting the error on running python main.py --cfg cfg/eval_coco.yml --gpu 1

Traceback (most recent call last):
  File "main.py", line 12, in <module>
    enable(os.environ["TELEMETRY"])
  File "/usr/lib/python2.7/UserDict.py", line 40, in __getitem__
    raise KeyError(key)
KeyError: 'TELEMETRY'

Hi,in my experiment,the inception score of the pretrained model on coco dataset is 16.16.Do you know why the IS is not stable?How can make the IS to 25.89?

Hey. First of all need to say that job and result is absolutely amazing. Thank you for your work and sharing the code with the community.

Going through the steps to run your code with your data is quite easy and I get the same result as yours. But when i try to figure out how to test your architecture on external data i faced the issue of preprocessed metadata for each dataset you work with.

So can you please list some steps on how to feed some external data (aka bunch of images with captions) to your model (including pretraining DAMSM and embedding vectors). Suppose this is will be very useful information to expand your research on broad spheres.

Can someone explain what the function of cond_wrong_errD is in the discriminator loss function? It seems not a part of the discriminator loss mentioned in the paper (Eq. 5). Also, it does not make sense to me. Why ignore the last entry in the batch?

cond_wrong_logits = netD.COND_DNET(real_features[:(batch_size - 1)], conditions[1:batch_size])
cond_wrong_errD = nn.BCELoss()(cond_wrong_logits, fake_labels[1:batch_size])

Hi, thanks for sharing your brilliant work. I am a little confused when I read your code in trainer.py line 227 where a set of fixed noise is sampled before the training loop of epoch, which means the whole following training process used the fixed noise. As far as I know about GAN, the noise should not be fixed. If a set of fixed noise vectors is used for training, the model will lose generalization ability, which means when the input noise vector is randomly sampled, the model cannot generate a good image. When I use your code and trained model to generate images with randomly sampled noise, the results are very bad. Please let me know where I have missed understanding or incorrect setting. Thanks in advance!

Thanks for this excellent work! I have some problem with AttnGAN training. I am not sure whether i have a correct thought of the following code.

/code/trainer.py
line 246 ~ 271
imgs, captions, cap_lens, class_ids, keys = prepare_data(data)
errD = discriminator_loss(netsD[i], imgs[i], fake_imgs[i],
                                              sent_emb, real_labels, fake_labels)

I think netsD[i] is the ith discriminator used in AttnGAN. In paper, there are three discriminators. I also think that imgs are corresponding to captions and have a size equal to batch size. So, how can parameter i represents both the netsD index and img index.

Traceback (most recent call last): File "/AttnGAN-master3/code/main.py", line 140, in algo.train() File "\AttnGAN-master3\code\trainer.py", line 285, in train sent_emb, real_labels, fake_labels) File "\AttnGAN-master3\code\miscc\losses.py", line 143, in discriminator_loss cond_real_logits = netD.COND_DNET(real_features, conditions) File "\Anaconda3\envs\torch-env\lib\site-packages\torch\nn\modules\module.py", line 489, in call result = self.forward(*input, **kwargs) File "\AttnGAN-master3\code\model.py", line 567, in forward h_c_code = torch.cat((h_code, c_code), 1) RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 4 and 1 in dimension 2 at c:\a\w\1\s\tmp_conda_3.6_091443\conda\conda-bld\pytorch_1544087948354\work\aten\src\thc\generic/THCTensorMath.cu:83

Process finished with exit code 1

Training using config file "bird_attn2.yml" worked normally, but when I tried to use DCGAN version, it failed with errors above.

def upBlock(in_planes, out_planes):
    block = nn.Sequential(
        nn.Upsample(scale_factor=2, mode='nearest'),
        conv3x3(in_planes, out_planes * 2),
        nn.BatchNorm2d(out_planes * 2),
        GLU())
    return block

i think the out_planes shouldn't times 2, and then the channel can decrease step by step ,because the argument feed into the function in_planes have been two times than out_planes

Where should I place my text into, if I want to generate new images. I have successfully reproduced your images, but when i put my own images into the example_captions.txt they dont come up. So where should I put new text ???

In many places, the fake images are generated via:

fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs, mask)

The netG is of class G_NET defined in https://github.com/taoxugit/AttnGAN/blob/master/code/model.py#L397.

When I keep noise, sent_emb, words_embs and mask constant and rerun the generation, I get different fake images. Shouldn't the model be outputting a constant output for a constant input? Is there any stochastic behaviour of the G_NET?

The global git ignore file in the /data directory currently excludes example files such as /data/birds/example_filenames.txt and /data/birds/example_captions.txt.

Preserving these files by using git ignore more selectively should help new users test this network more easily.

I was exploring the codebase, and wanted to test some ideas on the same. But my machine has an AMD GPU, thus no support for CUDA. I shall have to use OpenCL instead. Not sure, if there is a way to make some tweaks and make it run here on my machine, or rewriting is the only way to achieve the same?

Can someone help me resume the training? I dont have GPU of my own. Currently i am training it in Colab and it only gives 12 hours which is not enough. So can someone help to to resume training from saved "netg" model. PLEASE!

 if class_ids is not None:
        scores0.data.masked_fill_(masks, -float('inf'))

This code is used in some of the generator losses or in the DAMSM loss.. what is the need for using class when we want to generate images and not do classification?

Please someone help me. The validation set for CUB is 2933, How many images have you generated to calculate the R precision? 30000? in code, How much have you considered the size of R? Is the condition of 30,000 considered? `R_count = 0 R = np.zeros(2928) . . .

if R_count >= 30000: sum = np.zeros(8) np.random.shuffle(R) for i in range(8): sum[i] = np.average(R[i * 3000:(i + 1) * 3000- 1]) R_mean = np.average(sum)*100 R_std = np.std(sum)*100 print("R mean:{:.2f} std:{:.2f}".format(R_mean, R_std)) cont = False `