In case that helps, there is a dataset of ~800 pokemon images on Kaggle.
(their GAN is available on github as well: https://github.com/kvpratama/gan/tree/master/pokemon , which includes their original images)

Readme says that before running pokeGAN.py one should run resize.py and RGBA2RGB.py.
There are few problems here and I'll work backwards from pokeGAN.py to explain them. In pokeGAN.py the height and width of the images are hardcoded to be 128 and 128. However in resize.py resized images height and width are 256. This is the first wtf.
Second wtf is that pokeGAN.py reads the data from 'data' directory (line 29), but resize.py puts data in a directory named 'resizedData'. Third thing is that RGBA2RGB.py takes this a step further and puts files from resizedData to resized_black (which is bad naming, but let it be).

To conclude, either the pokeGAN.py reads images from wrong directories or helper scripts put processed images into wrong directories. I understand that this is not your code, but let's fix this if people of YouTube will try to use it 😉

When using tf.layers.conv2d_transpose what defines the output tensor shape?

In the generator code, when the input was 4x4x512, for the output to be 8x8x256 the number of filters can be given, but how are the height and width of the output defined? Or else is it always two times the input height and width?

Thanks.

Results are getting black color. Is it need to uncomment the line before saving to get proper result

imgtest = imgtest * 255.0

imgtest.astype(np.uint8)

        save_images(imgtest, [8,8] ,newPoke_path + '/epoch' + str(i) + '.jpg')

Possibly the dependencies list is incomplete. When I first ran the code it threw error "module 'scipy.misc' has no attribute 'imsave'"

Googling led to this being about pillow. It was suggested to uninstall scipy, install pillow, then install scipy again. I did this and the error was gone.

Perhaps pillow should be on the dependencies list?

What type of training times are you getting with 5k epochs on a GPU? Does 20 hours seem in the right ballpark?

Hi everyone,

I got some error like above when I'm trying to run pokeGAN.py. Is there anaybody know the solution?

Where is the Input_fn? it is practically impossible to read and write data to the disk, do you imagine how inefficient it makes the code? You have to make the crop and resize on the fly using the TF DATASET API.

When I'm trying to run pokeGAN.py, line 225 throws the following error: tensorflow.python.framework.errors_impl.OutOfRangeError: RandomShuffleQueue '_0_shuffle_batch/random_shuffle_queue' is closed and has insufficient elements (requested 64, current size 0)
[[Node: shuffle_batch = QueueDequeueManyV2[component_types=[DT_FLOAT], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](shuffle_batch/random_shuffle_queue, shuffle_batch/n)]]

While this error is throwing during the train_image = sess.run(image_batch) instruction, it seems to be related to line 58, where is defined a shuffle batch instruction:

    images_batch = tf.train.shuffle_batch(
                                    [image], batch_size = BATCH_SIZE,
                                    num_threads = 4, capacity = 200 + 3* BATCH_SIZE,
                                    min_after_dequeue = 200)

I couldn't find anyone else on the internet with the same mistake, but it happens when I try to run this code without any changes.

Hi
Does anyone have any idea why i am getting this error. Here is the error log.

2018-06-04 10:46:48 PSTTraceback (most recent call last):
2018-06-04 10:46:48 PSTFile "gan.py", line 296, in <module>
2018-06-04 10:46:48 PSTtrain_network()
2018-06-04 10:46:48 PSTFile "gan.py", line 267, in train_network
2018-06-04 10:46:48 PSTsave_images(imgtest, [8,8] ,new_image_path + '/epoch' + str(i) + '.jpg')
2018-06-04 10:46:48 PSTFile "/code/utils.py", line 34, in save_images
2018-06-04 10:46:48 PSTreturn imsave(inverse_transform(images), size, image_path)
2018-06-04 10:46:48 PSTFile "/code/utils.py", line 67, in imsave
2018-06-04 10:46:48 PSTimage = np.squeeze(merge(images, size))
2018-06-04 10:46:48 PSTFile "/code/utils.py", line 53, in merge
2018-06-04 10:46:48 PSTimg[j * h:j * h + h, i * w:i * w + w, :] = image
2018-06-04 10:46:48 PSTValueError: could not broadcast input array from shape (128,128,3) into shape (0,128,3)

As soon as finish the first epoch this occurs.

        if i%50 == 0:
            # save images
            if not os.path.exists(new_image_path):
                os.makedirs(new_image_path)
            sample_noise = np.random.uniform(-1.0, 1.0, size=[batch_size, random_dim]).astype(np.float32)
            imgtest = sess.run(fake_image, feed_dict={random_input: sample_noise, is_train: False})
            # imgtest = imgtest * 255.0
            # imgtest.astype(np.uint8)
            save_images(imgtest, [8,8] ,new_image_path + '/epoch' + str(i) + '.png')  # <<<< HERE 

I've been trying to run this code with a separate collection of images. I keep running into a problem where the images are cloudy/washed-out. I've been looking through this previous issue raised here https://github.com/llSourcell/Pokemon_GAN/issues/9?_pjax=%23js-repo-pjax-container.

It seems to be an issue with the channels, but I can't seem to figure it out. I've already made sure that the images I'm working with are all .jpegs. Thanks for taking the time.

After completing the 5000 epoch run, my output image is nothing like yours... thoughts?!

In my opinion, to use the stored ckpt, the pointed replacement is needed.

I noted that for activation relu is used in the generator and leaky relu is used in the Discriminator. Is there a particular reason behind this?

In the code, it says CONV->BIAS->ACTIVATION but it is CONV->BATCH NORMALISATION->ACTIVATION. Correct me if I am wrong!

Hi, I just wanted to comment that you are using transparent images with 4 channels, but when you read them, you use 3 channels only. So why not change the input and output channels to be 4?

Sorry to bother, but there should be a

os.mkdir(dst)

before the for loop in RGBA2RGB.py to create the folder else it may give an error .
Was facing the problem so thought might point out!

Big fan :)

Hello,

I'm having a minor issue when trying to continue an interrupted training.

It is creating these files in /tmp/ as it's supposed to:

And these under model/NewPokemon in my working directory:

On other code I worked on, I would import the saved weights and biases with something like;

saver = tf.train.import_meta_graph('model.meta')
saver.restore(sess, tf.train.latest_checkpoint('./'))

On this code I just couldn't make it continue from where it is left off. It starts from 0 back again when I run PokeGAN.py even though I'm running inside the directory and the folder structure seems correct.

I haven't modified the code except the fact that, mine saves the model every epoch instead of every 500 - just changed the modulo i%500 to i%1.

Any help would be much appreciated.

Thanks very much!

That's how images looks like when I run resize.py script.
Quite sure it's not good.

These two questions came up, when I went threw the code:

• In line 127 of pokeGan.py the bn1 tensor is created. But actually it is never used.. instead the next line continues with lrelu(conv1,...). Any reason for that?

• Why is tanh used to generate the final images (line 113). Shouldn't the images values be between 0 and 1? So would not the sigmoid activation be the better choice here?

Any answers / ideas?