ningyu1991 / ganfingerprints Goto Github PK

Both CramerGAN and MMDGAN export images with wrong colors:

The error is located in the export images function (equivalent for both GANs):

def inverse_transform(images):
    return (images+1.)/2.


def save_images(images, size, image_path):
    merged = merge(inverse_transform(images), size)
    return scipy.misc.imsave(image_path, merged)

Values passed to save_images are in the range [0,1]. The call to inverse_transform rescales them to [0.5, 1], which in turn causes the pixel values to be only in the range [127.5, 255], instead of [0, 255]. This results in colors only from the brighter part of the spectrum, as shown above.

Hello,

I am testing your code on generated images from the pre-trained models you have provided. There are quite a few issues encountered while attempted to run the classifier on this dataset we have creating (testing portion of your code).

• The environment had to be build as CUDA 10.0, Python3.7, and tensorflow-gpu 1.13.1, as previous versions were incompatible
• once run with the pk pre-trained model provided, in file GANFingerprints/classifier/util_scripts.py, line 23 C_im = misc.load_network_pkl(model_path) cannot be used as is anymore when executing line 57 logits = C_im.run(im, minibatch_size=1, num_gpus=1, out_dtype=np.float32) as this ends up being a tuple. I ended up changing line 23 to G, D, Gs = misc.load_network_pkl(model_path) and using logits = D.run(im, minibatch_size=1, num_gpus=1, out_dtype=np.float32) to run the classifier.
• upon running the classifier, assignment of labels gets skipped over. This would mean it does not match the labels at the top. What could we use as labels if logits.shape[1] = 1?
• If I use labels as list(labels_1) or list(labels_2) , it classifies everything as from CelebA, which is incorrect as the dataset I am passing in is all fake images generated from sngan model.

Please let me know how I can fix this.
Thanks

for example
labels_1 = ['CelebA_real_data', 'ProGAN_generated_data', 'SNGAN_generated_data', 'CramerGAN_generated_data', 'MMDGAN_generated_data'] in util_scripts.py

I've found that if I use a lexicographical order with my own data set, I get an error response

When I tested it with my own data set, I found it wasn't in lexicographic order。

for exmaple，in util_scripts.py

labels_1 = ['CelebA_real_data', 'ProGAN_generated_data', 'SNGAN_generated_data', 'CramerGAN_generated_data', 'MMDGAN_generated_data'].

But in what order should I use my own data set?What is the rule for this order? Thank you