def generate(self, source, destination = None, samples = 20, grid_width=480, grid_height=240):
# Load the input and output of the graph
input = self.graph.get_tensor_by_name("en_input:0" )
output = self.graph.get_tensor_by_name("decoder/main_out:0")
# For generating a single image for a source image
if os.path.isfile(source):
image = loader.load_image(source, self.image_size, self.image_size)
batch = np.asarray([image for _ in range(self.batch_size)])
preds = self.session.run(output, feed_dict = {input: batch})
batch[0] = utils.add_border(batch[0],color = [1.0,0.0,0.0])
grid = np.concatenate((batch[0], preds[0]), axis=1)
grid = (grid * 255.0).astype(np.uint8)
# For generating a grid of images from a directory of images
elif os.path.isdir(source):
data = loader.DataSet(images_dir = source,
hard_load = False,
width = self.image_size,
height = self.image_size)
batch,_= data.next_batch(self.batch_size)
preds = self.session.run(output, feed_dict={input: batch})
for i in range(samples):
batch[i] = utils.add_border(batch[i],color = [1.0,0.0,0.0])
grid = self.construct_image_grid(batch,preds,samples,grid_width,grid_height)
else:
print(source,"must be an image pathname or a directory of images")
sys.exit()
if destination:
cv2.imwrite(destination, grid)
else:
cv2.imshow("images", grid)
cv2.waitKey()
def generate(self):
self.data = loader.DataSet(train_path=self.train_path,
width=self.image_size,
height=self.image_size,
validation_size=self.validation_size,
hard_load=False)
f, a = plt.subplots(4, 10, figsize=(10, 4))
x_batch, y_batch = self.data.valid.next_batch(self.batch_size)
output = self.graph.get_tensor_by_name("decoder/main_out:0")
input = self.graph.get_tensor_by_name("en_input:0")
label = self.graph.get_tensor_by_name("labels:0")
preds = self.session.run(output,
feed_dict={
input: x_batch,
label: y_batch
})
for i in range(10):
a[0][i].imshow(x_batch[i])
a[1][i].imshow(preds[i])
a[2][i].imshow(x_batch[10 + i])
a[3][i].imshow(preds[10 + i])
f.show()
plt.draw()
plt.waitforbuttonpress()
def train_initialization(self,
hard_load=True,
en_iterations=1,
de_iterations=1):
# initialize training dataset
self.data = loader.DataSet(images_dir=self.train_path,
width=self.image_size,
height=self.image_size,
split_ratio=self.validation_size,
hard_load=hard_load,
verbose=self.verbose)
self.en_iterations = max(1, en_iterations)
self.de_iterations = max(1, de_iterations)
# get number of batches for a single epoch
self.num_batches = self.data.train.size // self.batch_size
if self.verbose:
print("Train Data size=", self.data.train.size)
print("Test Data size=", self.data.valid.size)
print("Batch size =", self.batch_size)
print("Learning rate =", self.lr_rate)
self.noise_shape = [self.batch_size, self.latent_dim]
self.input_shape = [
self.batch_size, self.image_size, self.image_size, self.channels
]
self.label_shape = [self.batch_size, self.cond_dim]
self.en_input = tf.placeholder(tf.float32,
shape=self.input_shape,
name='en_input') # Input
self.labels = tf.placeholder(tf.float32,
shape=self.label_shape,
name='labels') # Condition vector
self.noise_input = tf.placeholder(tf.float32,
shape=self.noise_shape,
name='noise_input')
# create the network's model and optimizer
self.create_network()
self.create_optimizer()
# initialize of all global variables
global_variables = tf.global_variables_initializer()
self.session.run(global_variables)
self.saver = tf.train.Saver()
if self.pretrained == True:
if self.verbose == True:
print("Loading pretrained model...", end='')
meta_graph = self.model_path + self.model_name + '.meta'
checkpoint = tf.train.latest_checkpoint(self.model_path) #
self.saver.restore(self.session, checkpoint) # Load the weights
if self.verbose == True: print("done")
