Hello, I can train your model on CPU with the following configuration:CUDA8.0 Tensorflow(1.6.0) Keras (2.1.5) .
But when I install Tensorflow-gpu(1.6.0) instead of Tensorflow(1.6.0), the program does not work ,even if I tried Tensorflow-gpu1.5.0 to 1.0.0. My GPU is NVIDIA 1080.
Could you please find the error and tell me about your configuration of CUDA ,Tensorflow and so on,which is not showing in yout 'requirements.txt'.
Thanks !

Hi, neka-nat,

I use Tensorflow-gpu (2.3.0) and keras (2.4.0).
When I run the test.py, I met the error as follows.

ValueError: in user code:

ValueError: Layer sequential_1 expects 1 inputs, but it received 2 input tensors. Inputs received: [<tf.Tensor 'IteratorGetNext:0' shape=(None, 256, 256, 3) dtype=float32>, <tf.Tensor 'IteratorGetNext:1' shape=(None, 256, 256, 1) dtype=float32>]

Please give me some advice, thanks!

thanks for your work. I am running your code and have some questions.
where did you use the joint loss in your code? And I notice the weight for discrimitor loss is 0.0004 in the original paper. where did you set the parameter?
thanks, any help appreciated!

The result seems not good as the result in the paper.It there just need more training or other tricks?

Hi, as the title, sorry for that I have not run successfully the code for now so i did not know it it can fulfill random inpainting. I noticed the script in https://github.com/nekanat/image_completion_keras/blob/master/train.py.
may be the code can fulfill random inpainting? is it?

Hi, neka-nat, Have you wrote code for testing ? If so , can you upload it ?

I have finished the training for master branch, I try to use "preds = com_model.predict(test_img)" to test my pictures after each training epoch, test_img.shape=(1,256,256,3).
But the test results are too bad which is totally different from the result of training. Give me some help please. How to predict images with arbitrary mask shape.

Hello, I find you have three branches in this project, and you last submit your code in the anim branch.
Can you tell me the difference between these three branches if you are in rest?
Thanks a lot

Hi,

When I'm trying to run the test script it gives the following error:

ValueError: Layer sequential expects 1 inputs, but it received 2 input tensors. Inputs received: [<tf.Tensor 'IteratorGetNext:0' shape=(None, 128, 128, 3) dtype=float32>, <tf.Tensor 'IteratorGetNext:1' shape=(None, 128, 128, 1) dtype=float32>]

I don't understand where it comes from, since the cmp_model.predict([inputs, masks]) works fine in the train file. I have made no alteration to the code. I hope you can help me!

@neka-nat hello, in the original paper, Computational times with different resolution of the input image are tested(such as 512x512,768x768 and 1024x1024), how can I train the model with (256,256,3) then use the model to test images with arbitrary shape ？

I have tried such as change input_shape to (None, None, 3) ,But it can't work. Only can change to (512,512,3), so the model only can deal with images with shape of (512,512,3)
Can you give me some advice ?

Hello, I am a new deep learning person and there is a mistake in train.py but I don't know what is wrong with it. "Local variable 'inputs' referenced before assignment" Check the inputs: they are clearly defined.

Hello,I trained the cmp_model Tc = 18 times and d_model Td=2 times, then after the all_module trained 40 times ,the program was killed with the following error, and I find my computer available memory is reduced from 14G to 100M until the program is killed:
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Abandoned (core dumped)

It is need some tricks in the training?

Hi neka-nat.
I put the required pictures where they are needed. I trained my model and ı have generator.h5, discriminator.h5 .
But my test.py only shows model summary. Can u help me please?
My code output is:

WARNING:tensorflow:No training configuration found in save file: the model was not compiled. Compile it manually.
Model: "sequential"

Layer (type) Output Shape Param #

conv2d (Conv2D) (None, 256, 256, 64) 4864

batch_normalization (BatchNo (None, 256, 256, 64) 256

activation (Activation) (None, 256, 256, 64) 0

conv2d_1 (Conv2D) (None, 128, 128, 128) 73856

batch_normalization_1 (Batch (None, 128, 128, 128) 512

activation_1 (Activation) (None, 128, 128, 128) 0

conv2d_2 (Conv2D) (None, 128, 128, 128) 147584

batch_normalization_2 (Batch (None, 128, 128, 128) 512

activation_2 (Activation) (None, 128, 128, 128) 0

conv2d_3 (Conv2D) (None, 64, 64, 256) 295168

batch_normalization_3 (Batch (None, 64, 64, 256) 1024

activation_3 (Activation) (None, 64, 64, 256) 0

conv2d_4 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_4 (Batch (None, 64, 64, 256) 1024

activation_4 (Activation) (None, 64, 64, 256) 0

conv2d_5 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_5 (Batch (None, 64, 64, 256) 1024

activation_5 (Activation) (None, 64, 64, 256) 0

conv2d_6 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_6 (Batch (None, 64, 64, 256) 1024

activation_6 (Activation) (None, 64, 64, 256) 0

conv2d_7 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_7 (Batch (None, 64, 64, 256) 1024

activation_7 (Activation) (None, 64, 64, 256) 0

conv2d_8 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_8 (Batch (None, 64, 64, 256) 1024

activation_8 (Activation) (None, 64, 64, 256) 0

conv2d_9 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_9 (Batch (None, 64, 64, 256) 1024

activation_9 (Activation) (None, 64, 64, 256) 0

conv2d_10 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_10 (Batc (None, 64, 64, 256) 1024

activation_10 (Activation) (None, 64, 64, 256) 0

conv2d_11 (Conv2D) (None, 64, 64, 256) 590080

batch_normalization_11 (Batc (None, 64, 64, 256) 1024

activation_11 (Activation) (None, 64, 64, 256) 0

conv2d_transpose (Conv2DTran (None, 128, 128, 128) 524416

batch_normalization_12 (Batc (None, 128, 128, 128) 512

activation_12 (Activation) (None, 128, 128, 128) 0

conv2d_12 (Conv2D) (None, 128, 128, 128) 147584

batch_normalization_13 (Batc (None, 128, 128, 128) 512

activation_13 (Activation) (None, 128, 128, 128) 0

conv2d_transpose_1 (Conv2DTr (None, 256, 256, 64) 131136

batch_normalization_14 (Batc (None, 256, 256, 64) 256

activation_14 (Activation) (None, 256, 256, 64) 0

conv2d_13 (Conv2D) (None, 256, 256, 32) 18464

batch_normalization_15 (Batc (None, 256, 256, 32) 128

activation_15 (Activation) (None, 256, 256, 32) 0

conv2d_14 (Conv2D) (None, 256, 256, 3) 867

batch_normalization_16 (Batc (None, 256, 256, 3) 12

activation_16 (Activation) (None, 256, 256, 3) 0

Total params: 6,076,495
Trainable params: 6,070,537
Non-trainable params: 5,958

#hello neka-nat.
In the train.py line 98 and 99. your code is
'''
d_container = Container([org_img, in_pts], discriminator([org_img, in_pts]))
d_model = Model([org_img, in_pts], d_container([org_img, in_pts]))
'''
in line 76 you create discriminator.'discriminator = model_discriminator(input_shape, local_shape)'
can you tell me the difference between discriminator and d_model?

I am confused about why you such design discriminator and d_model. Please tell me if you are free.
Thanks

hello,neka-nat!
I am just a bignner. My GPU is low, so i need to load G-discriminator to continue traning. But i can't load G-discriminator like before,It may be caused by the code in line 101,93 of 'mode.py'.I tried some methods that didn't work,So,how do you do it?
Looking forward to your answers..

Do we need to normalize the training images into the range of (-1,1) before feed them into the network?

Thanks a lot for your contribution. When learning the code in the train.py, I find d_container.trainable is set to False. It means that the discriminator does not need to be trained? However, it seems that d_model is trained through "d_loss_real = d_model.train_on_batch([inputs, points], valid)". Can you tell me the reason? Thank you very much! I am looking forward for your response.

hello,and I have a question,how can I get the datasets,

Thank you for the great work. Don't you think d_container.trainable should be set to True after training all_model.
Something like this:

d_container.trainable = True 
all_model.compile(loss=['mse', 'binary_crossentropy'], loss_weights=[1.0, alpha], optimizer=optimizer) 

I get this error on Keras 2.2.2.
Any idea how to replace the merge?

  File "train.py", line 160, in <module>
    main()
  File "train.py", line 156, in main
    example_gan()
  File "train.py", line 84, in example_gan
    in_img = merge([org_img, mask],mode=lambda x: x[0] * (1 - x[1]),output_shape=input_shape)
TypeError: 'module' object is not callable```