clovaai / stargan-v2-tensorflow Goto Github PK

Not an issue. But just a question.

This might be a beginner level question. I will give a scenario. Like say we have cat, dog and tiger image. And say we are transitioning from cat to dog.

Will the quality of cat to dog transition will be better if we only train with cat and dog images (like cyclegan)?

Or we train with all cat,dog and tiger images?

Like cat as source image. Dog and Tiger as Target Images. Transitioning from cat to dog. And cat to tiger.

Will the quality and accuracy be same when we do multiple transitioning and single transitioning?

Hi, can you share the pre-trained models of the networks trained using this code? Also, what were the FID and LPIPS score comparison between TensorFlow and PyTorch code?

Shouldn't we update style encoder when we call g_train_step() with x_ref and not with z_trg?

 if z_trgs is not None:
      f_train_variable = self.mapping_network.trainable_variables
      e_train_variable = self.style_encoder.trainable_variables

      f_gradient = g_tape.gradient(g_loss, f_train_variable)
      e_gradient = g_tape.gradient(g_loss, e_train_variable)

      self.f_optimizer.apply_gradients(zip(f_gradient, f_train_variable))
      self.e_optimizer.apply_gradients(zip(e_gradient, e_train_variable))

Should be:

 if z_trgs is not None:
      f_train_variable = self.mapping_network.trainable_variables
      f_gradient = g_tape.gradient(g_loss, f_train_variable)
      self.f_optimizer.apply_gradients(zip(f_gradient, f_train_variable))
 else:
      e_train_variable = self.style_encoder.trainable_variables
      e_gradient = g_tape.gradient(g_loss, e_train_variable)
      self.e_optimizer.apply_gradients(zip(e_gradient, e_train_variable))

Thank you for your excellent works! Could you please to release your pretrained models, I'll be very appreciated about that.

hi,
Can you describe the directory structure of your data set? and how to split source and reference image?
I trained with my own data set and the results are strange

I am getting this error in Colab

`

Traceback (most recent call last):
File "main.py", line 117, in
main()
File "main.py", line 103, in main
gan.build_model()
File "/content/stargan-v2-tensorflow/StarGAN_v2.py", line 197, in build_model
self.mapping_network_ema = MappingNetwork(self.style_dim, self.hidden_dim, self.num_domains, sn=False, name='MappingNetwork')
File "/content/stargan-v2-tensorflow/networks.py", line 85, in init
self.shared_layers, self.unshared_layers = self.architecture_init()
File "/content/stargan-v2-tensorflow/networks.py", line 96, in architecture_init
shared_layers = Sequential(layers)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/training/tracking/base.py", line 457, in _method_wrapper
result = method(self, *args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/sequential.py", line 142, in init
self.add(layer)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/training/tracking/base.py", line 457, in _method_wrapper
result = method(self, *args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/sequential.py", line 189, in add
'to pass a unique name.' % (layer.name,))
ValueError: All layers added to a Sequential model should have unique names. Name "relu" is already the name of a layer in this model. Update the name argument to pass a unique name `

Can someone help if there are changes in the hyper parameters if I want to change image size (output) to 512x512 or even 1024x1024, please?

Thanks,
Steve

This is not an issue but I am curious why there is such a big difference. Is it about implementation or difference btw Tensorflow vs Pytorch? thanks

Please help. I used tensorflow 2.10.1
Traceback (most recent call last):
File "D:\stargan-v2-tensorflow\main.py", line 117, in
main()
File "D:\stargan-v2-tensorflow\main.py", line 103, in main
gan.build_model()
File "D:\stargan-v2-tensorflow\StarGAN_v2.py", line 146, in build_model
self.style_encoder = StyleEncoder(self.img_size, self.style_dim, self.num_domains, max_conv_dim=self.hidden_dim, sn=False, name='StyleEncoder')
File "D:\stargan-v2-tensorflow\networks.py", line 141, in init
self.shared_layers, self.unshared_layers = self.architecture_init()
File "D:\stargan-v2-tensorflow\networks.py", line 161, in architecture_init
shared_layers = Sequential(blocks)
File "C:\Users\user\anaconda3\envs\tf-gpu\lib\site-packages\tensorflow\python\trackable\base.py", line 205, in _method_wrapper
result = method(self, *args, **kwargs)
File "C:\Users\user\anaconda3\envs\tf-gpu\lib\site-packages\keras\utils\traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
File "C:\Users\user\anaconda3\envs\tf-gpu\lib\site-packages\keras\engine\sequential.py", line 190, in add
raise ValueError(
ValueError: All layers added to a Sequential model should have unique names. Name "leaky_relu" is already the name of a layer in this model. Update the name argument to
pass a unique name.

Hello together!

When trying to train (only using the Files that were provided, no own data (yet)).
I encounter the following error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: buffer_size must be greater than zero. [Op:ShuffleDatasetV2]

Full Error Stack:
Traceback (most recent call last):
File "main.py", line 117, in
main()
File "main.py", line 103, in main
gan.build_model()
File "C:\Users\TOGERLA\Desktop\stargan-v2-tensorflow-master\StarGAN_v2.py", line 138, in build_model
img_and_domain = img_and_domain.shuffle(buffer_size=dataset_num, reshuffle_each_iteration=True).repeat()
File "C:\Users\TOGERLA\Desktop\stargan-v2-tensorflow-master\venv\lib\site-packages\tensorflow_core\python\data\ops\dataset_ops.py", line 1193, in shuffle
return ShuffleDataset(self, buffer_size, seed, reshuffle_each_iteration)
File "C:\Users\TOGERLA\Desktop\stargan-v2-tensorflow-master\venv\lib\site-packages\tensorflow_core\python\data\ops\dataset_ops.py", line 3535, in init
**self._flat_structure)
File "C:\Users\TOGERLA\Desktop\stargan-v2-tensorflow-master\venv\lib\site-packages\tensorflow_core\python\ops\gen_dataset_ops.py", line 5221, in shuffle_dataset_v2
_ops.raise_from_not_ok_status(e, name)
File "C:\Users\TOGERLA\Desktop\stargan-v2-tensorflow-master\venv\lib\site-packages\tensorflow_core\python\framework\ops.py", line 6606, in raise_from_not_ok_status
six.raise_from(core._status_to_exception(e.code, message), None)
File "", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: buffer_size must be greater than zero. [Op:ShuffleDatasetV2]

I might be able to fix the issue myself, but I'll still try to document it here for other people to reproduce :)

Have a nice Christmas!
Tobias

    def build_model(self):
        if self.phase == 'train':
            """ Input Image"""
            img_class = Image_data(self.img_size, self.img_ch, self.dataset_path, self.domain_list, self.augment_flag)
            img_class.preprocess()

            dataset_num = len(img_class.images)
            print("Dataset number : ", dataset_num)

            img_and_domain = tf.data.Dataset.from_tensor_slices((img_class.images, img_class.shuffle_images, img_class.domains))

            gpu_device = '/gpu:1'

            img_and_domain = img_and_domain.shuffle(buffer_size=dataset_num, reshuffle_each_iteration=True).repeat()
            img_and_domain = img_and_domain.map(map_func=img_class.image_processing, num_parallel_calls=AUTOTUNE).batch(self.batch_size, drop_remainder=True)
            img_and_domain = img_and_domain.apply(prefetch_to_device(gpu_device, buffer_size=AUTOTUNE))

How to set up multiple gpu?

from where I can get main.py file ?

Seems missing self.img_ch as the second argument

When I try to install tensorflow with pip install Tensorflow==2.1.0 I'm getting an error No matching distribution found for Tensorflow==2.1.0. Only offered versions are 2.2.0 and higher.
Should the code be working with those?

I am trying to train starGan on Alderley Day/Night Dataset to turn day images to night, but after 613 epoch the loss turns NAN, GAN type is gan-gp and every thing is set to default, can you help to elaborate what could be the cause of this divergence

Can not run batch_size = 2 with 256x256 on 10G NVIDIA GPU card ??!?

This model is a bit too much for NVIDIA. Any hint gents ?

Steve

Hi,

According to the official TF documentation, ds_weight decay implemented on this way will not affect the loss weight inside g_train function. Please, take a look here: https://www.tensorflow.org/guide/function#depending_on_python_global_and_free_variables
Have you observed that behavior in your experiments as well?

Maybe loss weights can be passed as an argument to this function or directly converted to tf.Variable ?