I train 500 epoch, and get the different results

Isn't the code just training generator? The discriminator loss doesn't seem to be going down at all. Also there is a line .trainable = False for discriminator, I'm not super familiar with kera, but does this mean all params in discriminator are freezed? If that's the case, isn't this just training a CNN

Hi, your trained model download link is broken. Hope you'll fix the link soon.

def dcrm_loss(y_true, y_pred): return -tf.reduce_mean(tf.log(tf.maximum(y_true, EPSILON)) + tf.log(tf.maximum(1. - y_pred, EPSILON)))

Here is your dcrm_loss. However, I think the correct loss for discriminator is just binary_crossentropy, which is:

-y_true*tf.reduce_mean(tf.log(tf.maximum(y_pred, EPSILON)) + (1-y_true)*tf.log(tf.maximum(1. - y_pred, EPSILON)))

No matter what I do, every time I run outpaint.ipynb I get an error saying there's no keras_contrib module even though I see it listed using pip freeze. I looked at the issues on the keras-contrib Github page and people who have run into the same issue said they installed Python 3.7 but that hasn't been released yet. Do you have any idea why this might be happening and how to resolve it??

i downloaded your trained model .but i didn't see test.py , how should i test your model

In keras 2.1.5, there is no ReLU function in keras.layers

Fine, author.
May you provide some details on the training for help?
I've got a total black outpainting for the images for the images after 100+ epochs.
I'm using python 3.5
keras 2.2.0
keras-contrib 2.0.8 (the version it takes were 0.0.2 in coding)
and etc
Fine, I though the etc would not affect the result greatly...
Tensorflow 1.8.0
Cuda 9.1
Cudnn 7.0.5

Thank you for reading.

https://drive.google.com/file/d/1548iAtsNf3wLSc1i5zYy-HX8_TW95wi_/view?usp=sharing

I want to train your network with my data. please let me know how can I train your network with my own data.

it's my code :

 def mask_width(img):
    image = img.copy()
    height = image.shape[0]
    width = image.shape[1]
    new_width = int(width * MASK_PERCENTAGE)
    mask = np.ones([height, new_width, 3])
    missing_x = img[:, :new_width]
    missing_y = img[:, width - new_width:]
    missing_part = np.concatenate((missing_x, missing_y), axis=1)
    image = image[:, :width - new_width]
    image = image[:, new_width:]
    return image, missing_part
    
def get_masked_images(images):
    mask_images = []
    missing_images = []
    for image in images:
        mask_image, missing_image = mask_width(image)
        mask_images.append(mask_image)
        missing_images.append(missing_image)
    return np.array(mask_images), np.array(missing_images)

def get_demask_images(original_images, generated_images):
    demask_images = []
    for o_image, g_image in zip(original_images, generated_images):
        width = g_image.shape[1] // 2
        x_image = g_image[:, :width]
        y_image = g_image[:, width:]
        o_image = np.concatenate((x_image,o_image, y_image), axis=1)
        demask_images.append(o_image)
    return np.asarray(demask_images)

i also change def dcrm_loss
d_input_shape = (int(INPUT_SHAPE[0] * (MASK_PERCENTAGE * 2)), int(INPUT_SHAPE[1] * (MASK_PERCENTAGE * 2)), INPUT_SHAPE[2])

and change def gen_loss
g_input_shape = (int(INPUT_SHAPE[0] * (MASK_PERCENTAGE * 2)), int(INPUT_SHAPE[1] * (MASK_PERCENTAGE * 2)), INPUT_SHAPE[2])

and i train:
Error when checking input: expected input_2 to have shape (256, 128, 3) but got array with shape (128, 128, 3)
......
thanks for your time!

In outpaint.ipynb, below line gives the following error.
CONF_GENERATED_IMAGE = DCRM(GENERATED_IMAGE)
ValueError: number of input channels does not match corresponding dimension of filter, 3 != 256

I found that the shape of GENERATED_IMAGE is (None, 3, 128, 4) based on the summary but DCRM wants (None, 256, 128, 3) as an input. Could you guess what's going on please?

I only added 2 lines on top of the outpaint.ipynb, which are
import os os.environ['KERAS_BACKEND']='tensorflow'
otherwise, I get an error on the code
DCRM = build_discriminator()
TypeError: Failed to convert object of type <class 'theano.tensor.var.TensorVariable'> to Tensor. Contents: /dense_2_target. Consider casting elements to a supported type.

summary of GEN

Layer (type) Output Shape Param #

input_2 (InputLayer) (None, 256, 128, 3) 0

conv2d_6 (Conv2D) (None, 64, 128, 3) 409664

re_lu_1 (ReLU) (None, 64, 128, 3) 0

dropout_6 (Dropout) (None, 64, 128, 3) 0

instance_normalization_1 (In (None, 64, 128, 3) 2

conv2d_7 (Conv2D) (None, 128, 64, 2) 131200

re_lu_2 (ReLU) (None, 128, 64, 2) 0

dropout_7 (Dropout) (None, 128, 64, 2) 0

instance_normalization_2 (In (None, 128, 64, 2) 2

conv2d_8 (Conv2D) (None, 256, 32, 1) 524544

re_lu_3 (ReLU) (None, 256, 32, 1) 0

dropout_8 (Dropout) (None, 256, 32, 1) 0

instance_normalization_3 (In (None, 256, 32, 1) 2

conv2d_9 (Conv2D) (None, 512, 32, 1) 2097664

re_lu_4 (ReLU) (None, 512, 32, 1) 0

dropout_9 (Dropout) (None, 512, 32, 1) 0

instance_normalization_4 (In (None, 512, 32, 1) 2

conv2d_10 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_5 (ReLU) (None, 512, 32, 1) 0

dropout_10 (Dropout) (None, 512, 32, 1) 0

instance_normalization_5 (In (None, 512, 32, 1) 2

conv2d_11 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_6 (ReLU) (None, 512, 32, 1) 0

dropout_11 (Dropout) (None, 512, 32, 1) 0

instance_normalization_6 (In (None, 512, 32, 1) 2

conv2d_12 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_7 (ReLU) (None, 512, 32, 1) 0

dropout_12 (Dropout) (None, 512, 32, 1) 0

instance_normalization_7 (In (None, 512, 32, 1) 2

conv2d_13 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_8 (ReLU) (None, 512, 32, 1) 0

dropout_13 (Dropout) (None, 512, 32, 1) 0

instance_normalization_8 (In (None, 512, 32, 1) 2

conv2d_14 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_9 (ReLU) (None, 512, 32, 1) 0

dropout_14 (Dropout) (None, 512, 32, 1) 0

instance_normalization_9 (In (None, 512, 32, 1) 2

conv2d_15 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_10 (ReLU) (None, 512, 32, 1) 0

dropout_15 (Dropout) (None, 512, 32, 1) 0

instance_normalization_10 (I (None, 512, 32, 1) 2

conv2d_16 (Conv2D) (None, 512, 32, 1) 4194816

re_lu_11 (ReLU) (None, 512, 32, 1) 0

dropout_16 (Dropout) (None, 512, 32, 1) 0

instance_normalization_11 (I (None, 512, 32, 1) 2

conv2d_transpose_1 (Conv2DTr (None, 256, 64, 2) 2097408

re_lu_12 (ReLU) (None, 256, 64, 2) 0

conv2d_transpose_2 (Conv2DTr (None, 128, 128, 4) 524416

re_lu_13 (ReLU) (None, 128, 128, 4) 0

conv2d_17 (Conv2D) (None, 128, 128, 4) 262272

re_lu_14 (ReLU) (None, 128, 128, 4) 0

dropout_17 (Dropout) (None, 128, 128, 4) 0

instance_normalization_12 (I (None, 128, 128, 4) 2

conv2d_18 (Conv2D) (None, 64, 128, 4) 131136

re_lu_15 (ReLU) (None, 64, 128, 4) 0

dropout_18 (Dropout) (None, 64, 128, 4) 0

instance_normalization_13 (I (None, 64, 128, 4) 2

conv2d_19 (Conv2D) (None, 3, 128, 4) 3075

Total params: 35,545,117
Trainable params: 35,545,117
Non-trainable params: 0

summary of DCRM

Layer (type) Output Shape Param #

input_1 (InputLayer) (None, 256, 128, 3) 0

conv2d_1 (Conv2D) (None, 32, 64, 2) 204832

leaky_re_lu_1 (LeakyReLU) (None, 32, 64, 2) 0

dropout_1 (Dropout) (None, 32, 64, 2) 0

conv2d_2 (Conv2D) (None, 64, 32, 1) 51264

leaky_re_lu_2 (LeakyReLU) (None, 64, 32, 1) 0

dropout_2 (Dropout) (None, 64, 32, 1) 0

conv2d_3 (Conv2D) (None, 64, 16, 1) 102464

leaky_re_lu_3 (LeakyReLU) (None, 64, 16, 1) 0

dropout_3 (Dropout) (None, 64, 16, 1) 0

conv2d_4 (Conv2D) (None, 128, 8, 1) 204928

leaky_re_lu_4 (LeakyReLU) (None, 128, 8, 1) 0

dropout_4 (Dropout) (None, 128, 8, 1) 0

conv2d_5 (Conv2D) (None, 128, 4, 1) 409728

leaky_re_lu_5 (LeakyReLU) (None, 128, 4, 1) 0

dropout_5 (Dropout) (None, 128, 4, 1) 0

flatten_1 (Flatten) (None, 512) 0

dense_1 (Dense) (None, 1024) 525312

dense_2 (Dense) (None, 1) 1025

Total params: 1,499,553
Trainable params: 1,499,553
Non-trainable params: 0

if i set epoch=200,will the result become worse? must i set epoch=500 and the train result could be the best?

How can I get the correct color ?
I've tried many ways...