def main():
# initialize parameters
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
fraction = parser.fraction
data_path = parser.data_path
style_1 = parser.style_1
style_2 = parser.style_2
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
loss_type = parser.loss_type
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
style_dim = parser.style_dim
reconstruct_coef_1 = parser.reconstruct_coef_1
reconstruct_coef_2 = parser.reconstruct_coef_2
reconstruct_coef_3 = parser.reconstruct_coef_3
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
# load data
partition = np.arange(categorical_cardinality, dtype=np.int32)
np.random.shuffle(partition)
partition = partition[:int(categorical_cardinality * (1 - fraction))]
print('partition:\n', partition)
imageNameTrain1, imageDictTrain1, imageNameTest1, imageDictTest1 = locate(
data_path,
styles=['std/' + style_1 + '/cut'],
max_label=categorical_cardinality,
partition=partition)
imageNameTrain2, imageDictTrain2, imageNameTest2, imageDictTest2 = locate(
data_path,
styles=['std/' + style_2 + '/cut'],
max_label=categorical_cardinality,
partition=partition)
imageNameTrain3, imageDictTrain3, imageNameTest3, imageDictTest3 = locate(
data_path,
styles=['std/0/cut'],
max_label=categorical_cardinality,
partition=partition)
imageNum = len(imageNameTrain1)
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
image2 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image2")
image3 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image3")
image4 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image4")
image5 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image5")
image6 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image6")
is_training = tf.placeholder(tf.bool, [], name="is_training")
forward_loss, reconstruct_loss_1, reconstruct_loss_2, reconstruct_loss_3, generator_loss, discriminator_loss, \
image1_forward_reconstruct, image2_forward_reconstruct, image3_forward_reconstruct, image4_forward_reconstruct, \
class_vector_1, style_vector_1, image1_style_reconstruct, image3_style_reconstruct = ae_with_gan(image1,image2,image3,image4,image5,image6,kernel,stride,class_dim,style_dim,is_training, loss_type,
reconstruct_coef_1,reconstruct_coef_2,reconstruct_coef_3,generator_coef,discriminator_coef,
'ae-with-gan')
encoder_variables = scope_variables("ae-with-gan/encoder")
decoder_variables = scope_variables('ae-with-gan/decoder')
discriminator_variables_1 = scope_variables('ae-with-gan/discriminator_1')
discriminator_variables_2 = scope_variables('ae-with-gan/discriminator_2')
discriminator_variables_3 = scope_variables('ae-with-gan/discriminator_3')
#all_variables = scope_variables('ae-with-gan')
#print([n.name for n in tf.get_default_graph().as_graph_def().node])
#print([n.name for n in all_variables])
forward_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
generator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
discriminator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
forward_train = forward_solver.minimize(forward_loss,
var_list=encoder_variables +
decoder_variables)
generator_train = generator_solver.minimize(generator_loss,
var_list=decoder_variables)
discriminator_train = discriminator_solver.minimize(
discriminator_loss,
var_list=discriminator_variables_1 + discriminator_variables_2 +
discriminator_variables_3)
idxes_1 = np.arange(imageNum, dtype=np.int32)
idxes_2 = np.arange(imageNum, dtype=np.int32)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
np.random.shuffle(idxes_1)
np.random.shuffle(idxes_2)
forward_losses = []
reconstruct_losses_1 = []
reconstruct_losses_2 = []
reconstruct_losses_3 = []
generator_losses = []
discriminator_losses = []
for idx in range(0, imageNum, batch_size):
image1_batch = loader(imageNameTrain1[idxes_1[idx:idx +
batch_size]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_batch = loader(find_truth(
imageNameTrain1[idxes_1[idx:idx + batch_size]],
imageDictTrain3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image3_batch = loader(imageNameTrain2[idxes_2[idx:idx +
batch_size]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image4_batch = loader(find_truth(
imageNameTrain2[idxes_2[idx:idx + batch_size]],
imageDictTrain3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image5_batch = loader(find_truth(
imageNameTrain1[idxes_1[idx:idx + batch_size]],
imageDictTrain2),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image6_batch = loader(find_truth(
imageNameTrain2[idxes_2[idx:idx + batch_size]],
imageDictTrain1),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_training = {
image1: image1_batch,
image2: image2_batch,
image3: image3_batch,
image4: image4_batch,
image5: image5_batch,
image6: image6_batch,
is_training: True
}
# forward
_, _forward_loss, _reconstruct_loss_1, _reconstruct_loss_2, _reconstruct_loss_3 = sess.run(
[
forward_train, forward_loss, reconstruct_loss_1,
reconstruct_loss_2, reconstruct_loss_3
],
feed_dict=feed_dict_training)
forward_losses.append(_forward_loss)
reconstruct_losses_1.append(_reconstruct_loss_1)
reconstruct_losses_2.append(_reconstruct_loss_2)
reconstruct_losses_3.append(_reconstruct_loss_3)
# generator
_, _generator_loss = sess.run(
[generator_train, generator_loss],
feed_dict=feed_dict_training)
generator_losses.append(_generator_loss)
# discriminator
_, _discriminator_loss = sess.run(
[discriminator_train, discriminator_loss],
feed_dict=feed_dict_training)
discriminator_losses.append(_discriminator_loss)
print('epoch: %d\nforward_loss: %f\nself_reconstruct_loss: %f\ntruth_reconstruct_loss: %f\ntransfer_reconstruct_loss: %f\ngenerator_loss: %f\ndiscriminator_loss: %f\n' % \
(epoch, get_mean(forward_losses), get_mean(reconstruct_losses_1), get_mean(reconstruct_losses_2), get_mean(reconstruct_losses_3), get_mean(generator_losses), get_mean(discriminator_losses)))
# test
image1_plot = loader(imageNameTrain1[idxes_1[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_plot = loader(find_truth(imageNameTrain1[idxes_1[0:10]],
imageDictTrain3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image3_plot = loader(imageNameTrain2[idxes_2[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image4_plot = loader(find_truth(imageNameTrain2[idxes_2[0:10]],
imageDictTrain3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image5_plot = loader(find_truth(imageNameTrain1[idxes_1[0:10]],
imageDictTrain2),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image6_plot = loader(find_truth(imageNameTrain2[idxes_2[0:10]],
imageDictTrain1),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image1_plot,
image2: image2_plot,
image3: image3_plot,
image4: image4_plot,
image5: image5_plot,
image6: image6_plot,
is_training: False
}
_image1_forward_reconstruct, _image2_forward_reconstruct, _image3_forward_reconstruct, _image4_forward_reconstruct, _image1_style_reconstruct, _image3_style_reconstruct = sess.run(
[
image1_forward_reconstruct, image2_forward_reconstruct,
image3_forward_reconstruct, image4_forward_reconstruct,
image1_style_reconstruct, image3_style_reconstruct
],
feed_dict=feed_dict_not_training)
images = [
image1_plot, image2_plot, image3_plot, image4_plot,
image5_plot, image6_plot, _image1_forward_reconstruct,
_image2_forward_reconstruct, _image3_forward_reconstruct,
_image4_forward_reconstruct, _image1_style_reconstruct,
_image3_style_reconstruct
]
coefs = [
loss_type, lr, reconstruct_coef_1, reconstruct_coef_2,
reconstruct_coef_3, generator_coef, discriminator_coef
]
plot_batch(images, 'train', epoch, coefs)
image1_plot = loader(imageNameTest1,
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_plot = loader(find_truth(imageNameTest1, imageDictTest3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image3_plot = loader(imageNameTest2,
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image4_plot = loader(find_truth(imageNameTest2, imageDictTest3),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image5_plot = loader(find_truth(imageNameTest1, imageDictTest2),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image6_plot = loader(find_truth(imageNameTest2, imageDictTest1),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image1_plot,
image2: image2_plot,
image3: image3_plot,
image4: image4_plot,
image5: image5_plot,
image6: image6_plot,
is_training: False
}
_image1_forward_reconstruct, _image2_forward_reconstruct, _image3_forward_reconstruct, _image4_forward_reconstruct, _image1_style_reconstruct, _image3_style_reconstruct = sess.run(
[
image1_forward_reconstruct, image2_forward_reconstruct,
image3_forward_reconstruct, image4_forward_reconstruct,
image1_style_reconstruct, image3_style_reconstruct
],
feed_dict=feed_dict_not_training)
images = [
image1_plot, image2_plot, image3_plot, image4_plot,
image5_plot, image6_plot, _image1_forward_reconstruct,
_image2_forward_reconstruct, _image3_forward_reconstruct,
_image4_forward_reconstruct, _image1_style_reconstruct,
_image3_style_reconstruct
]
coefs = [
loss_type, lr, reconstruct_coef_1, reconstruct_coef_2,
reconstruct_coef_3, generator_coef, discriminator_coef
]
plot_batch(images, 'test', epoch, coefs)
coefs = [
loss_type, lr, reconstruct_coef_1, reconstruct_coef_2,
reconstruct_coef_3, generator_coef, discriminator_coef
]
suffix = ''
for coef in coefs:
suffix += str(coef) + '-'
saver.save(sess,
os.path.join(os.path.join('ckpt', suffix[:-1]), 'model'))
def main():
# initialize parameters
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
reconstruct_coef = parser.reconstruct_coef
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
# load data
imageName, imageDict = locate(data_path,
styles=styles,
max_label=categorical_cardinality)
_, imageTrue = locate(data_path, max_label=categorical_cardinality)
imageNum = len(imageName)
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
image2 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image2")
is_training = tf.placeholder(tf.bool, [], name="is_training")
forward_loss, reconstruct_loss, generator_loss, discriminator_loss, image1_forward_reconstruct, _ = cycle_consistent_vae_with_gan(
image1, image2, kernel, stride, class_dim, is_training,
reconstruct_coef, generator_coef, discriminator_coef,
'cycle-consistent-vae-with-gan')
encoder_variables = scope_variables(
"cycle-consistent-vae-with-gan/encoder")
decoder_variables = scope_variables(
'cycle-consistent-vae-with-gan/decoder')
discriminator_variables = scope_variables(
'cycle-consistent-vae-with-gan/discriminator')
forward_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
generator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
discriminator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
forward_train = forward_solver.minimize(forward_loss,
var_list=encoder_variables +
decoder_variables)
generator_train = generator_solver.minimize(generator_loss,
var_list=decoder_variables)
discriminator_train = discriminator_solver.minimize(
discriminator_loss, var_list=discriminator_variables)
idxes_1 = np.arange(imageNum, dtype=np.int32)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
np.random.shuffle(idxes_1)
forward_losses = []
generator_losses = []
discriminator_losses = []
for idx in range(0, imageNum, batch_size):
image1_batch = loader(imageName[idxes_1[idx:idx + batch_size]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_batch = loader(find_truth(
imageName[idxes_1[idx:idx + batch_size]], imageTrue),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_training = {
image1: image1_batch,
image2: image2_batch,
is_training: True
}
# forward
_, _forward_loss = sess.run([forward_train, forward_loss],
feed_dict=feed_dict_training)
forward_losses.append(_forward_loss)
# generator
_, _generator_loss = sess.run(
[generator_train, generator_loss],
feed_dict=feed_dict_training)
generator_losses.append(_generator_loss)
# discriminator
_, _discriminator_loss = sess.run(
[discriminator_train, discriminator_loss],
feed_dict=feed_dict_training)
discriminator_losses.append(_discriminator_loss)
print(
'epoch: %d\nforward_loss: %f, generator_loss: %f, discriminator_loss: %f\n'
% (epoch, get_mean(forward_losses), get_mean(generator_losses),
get_mean(discriminator_losses)))
image1_plot = loader(imageName[idxes_1[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_plot = loader(find_truth(imageName[idxes_1[0:10]],
imageTrue),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image1_plot,
image2: image2_plot,
is_training: False
}
image1_reconstruct = sess.run(image1_forward_reconstruct,
feed_dict=feed_dict_not_training)
plot(image1_plot, image1_reconstruct, image2_plot, epoch)
saver.save(sess, 'ckpt/model')
def test_with_graph_manually_set_up():
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
reconstruct_coef = parser.reconstruct_coef
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
imageName, imageDict = locate(data_path,
styles=styles,
max_label=categorical_cardinality)
_, imageTrue = locate(data_path, max_label=categorical_cardinality)
imageNum = len(imageName)
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
image2 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image2")
is_training = tf.placeholder(tf.bool, [], name="is_training")
forward_loss, reconstruct_loss, generator_loss, discriminator_loss, image1_forward_reconstruct, vector = cycle_consistent_vae_with_gan(
image1, image2, kernel, stride, class_dim, is_training,
reconstruct_coef, generator_coef, discriminator_coef,
'cycle-consistent-vae-with-gan')
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
saver = tf.train.Saver()
idxes_1 = np.arange(imageNum, dtype=np.int32)
np.random.shuffle(idxes_1)
with tf.Session(config=config) as sess:
saver.restore(sess, tf.train.latest_checkpoint('ckpt/'))
image1_test = loader(imageName[idxes_1[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_test = loader(find_truth(imageName[idxes_1[0:10]], imageTrue),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image1_test,
image2: image2_test,
is_training: False
}
image_reconstruct, latent_vector = sess.run(
[image1_forward_reconstruct, vector],
feed_dict=feed_dict_not_training)
print(latent_vector.shape)
print(latent_vector)
plot(image1_test, image_reconstruct, image2_test, 0)
def test_with_graph_automatically_loaded():
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
reconstruct_coef = parser.reconstruct_coef
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
imageName, imageDict = locate(data_path,
styles=styles,
max_label=categorical_cardinality)
_, imageTrue = locate(data_path, max_label=categorical_cardinality)
imageNum = len(imageName)
#image1 = tf.placeholder(tf.float32,[None, image_size, image_size, channel_size],name="image1")
#image2 = tf.placeholder(tf.float32,[None, image_size, image_size, channel_size],name="image2")
#is_training = tf.placeholder(tf.bool,[],name="is_training")
saver = tf.train.import_meta_graph('ckpt/server-2/model.meta')
graph = tf.get_default_graph()
with graph.as_default():
variable_names = [v.name for v in tf.all_variables()]
print(variable_names)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
idxes_1 = np.arange(imageNum, dtype=np.int32)
np.random.shuffle(idxes_1)
with tf.Session(config=config) as sess:
saver.restore(sess, tf.train.latest_checkpoint('ckpt/server-2/'))
image1_test = loader(imageName[idxes_1[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image2_test = loader(find_truth(imageName[idxes_1[0:10]],
imageTrue),
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
image1 = graph.get_tensor_by_name('image1:0')
image2 = graph.get_tensor_by_name('image2:0')
is_training = graph.get_tensor_by_name('is_training:0')
feed_dict_not_training = {
image1: image1_test,
image2: image2_test,
is_training: False
}
class_vector = graph.get_operation_by_name(
'cycle-consistent-vae-with-gan/encoder/class_vector/fully_connected/Maximum'
)
image_reconstruct = graph.get_operation_by_name(
'cycle-consistent-vae-with-gan/decoder/conv4/Conv2d_transpose/Sigmoid'
)
_class_vector, _image_reconstruct = sess.run(
[class_vector, image_reconstruct],
feed_dict=feed_dict_not_training)
print(_class_vector.shape)
plot(image1_test, _image_reconstruct, image2_test, 0)
def main():
# initialize parameters
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
fraction = parser.fraction
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
augment = parser.augment
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
loss_type = parser.loss_type
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
style_dim = parser.style_dim
reconstruct_coef_1 = parser.reconstruct_coef_1
reconstruct_coef_3 = parser.reconstruct_coef_3
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
# load data
partition = make_partition(300, categorical_cardinality, fraction)
imageNameTrain, imageNameTest = locate(data_path,
styles=styles,
max_label=categorical_cardinality,
partition=partition)
styleNum, charNum, imageNum = imageNameTrain.shape[
0], imageNameTrain.shape[
1], imageNameTrain.shape[0] * imageNameTrain.shape[1]
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
image1_binary = binary(image1, image_size, channel_size, 0.7)
'''
image1_averge_org = tf.reduce_mean(image1,[1,2,3])
image1_average = tf.reshape(tf.reduce_mean(image1,[1,2,3]),[-1,1,1,1])
#image1_average = tf.tile(image1_average,[1, image_size, image_size, channel_size])
image1_average = tf.cast(tf.ones_like(image1),tf.float32)*0.7
image1_mask = tf.cast(tf.less(image1,image1_average),tf.float32)
image1_binary = image1*image1_mask+(1-image1_mask)
'''
idxes_1 = np.arange(imageNum, dtype=np.int32)
idxes_2 = np.arange(imageNum, dtype=np.int32)
np.random.shuffle(idxes_1)
np.random.shuffle(idxes_2)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
image1_plot, _, _, _, _, _, _ = loader(imageNameTrain,
idxes_1[0:10],
idxes_2[0:10],
styleNum,
charNum,
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
augment=augment,
force_grayscale=force_grayscale)
feed_dict = {image1: image1_plot}
_image1_binary = sess.run(image1_binary, feed_dict=feed_dict)
images = [image1_plot, _image1_binary]
coefs = [
loss_type, lr, reconstruct_coef_1, reconstruct_coef_3,
generator_coef, discriminator_coef
]
plot_batch(images, 'train', 0, coefs)
exit(0)
def transfer_style():
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
style_dim = parser.style_dim
reconstruct_coef_1 = parser.reconstruct_coef_1
reconstruct_coef_2 = parser.reconstruct_coef_2
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
imageName, imageDict = locate(data_path,
styles=styles,
max_label=categorical_cardinality)
_, imageTrue = locate(data_path, max_label=categorical_cardinality)
imageNum = len(imageName)
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
image2 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image2")
is_training = tf.placeholder(tf.bool, [], name="is_training")
image1_forward_reconstruct, image2_forward_reconstruct = transfer(
image1, image2, kernel, stride, class_dim, style_dim, is_training,
'cycle-consistent-vae-with-gan')
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
saver = tf.train.Saver()
idxes_1 = np.arange(imageNum, dtype=np.int32)
np.random.shuffle(idxes_1)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
saver.restore(
sess, tf.train.latest_checkpoint('ckpt/server-14/1.0-1.0-5e-05-'))
for idx in range(0, imageNum, batch_size):
image_batch = loader(imageName[idxes_1[idx:idx + batch_size]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image_batch[:batch_size // 2, :],
image2: image_batch[batch_size // 2:, :],
is_training: False
}
_image1_forward_reconstruct, _image2_forward_reconstruct = sess.run(
[image1_forward_reconstruct, image2_forward_reconstruct],
feed_dict=feed_dict_not_training)
print(_image1_forward_reconstruct.shape,
_image2_forward_reconstruct.shape)
exit(0)
def main():
# initialize parameters
parser = init()
os.environ["CUDA_VISIBLE_DEVICES"] = parser.gpu
categorical_cardinality = parser.categorical_cardinality
data_path = parser.data_path
styles = parser.styles
image_size = parser.image_size
force_grayscale = parser.force_grayscale
channel_size = 1 if force_grayscale else 3
seed = parser.seed
lr = parser.lr
batch_size = parser.batch_size
epochs = parser.epochs
kernel = parser.kernel
stride = parser.stride
class_dim = parser.class_dim
style_dim = parser.style_dim
continuous_dim = parser.continuous_dim
reconstruct_coef = parser.reconstruct_coef
continuous_coef = parser.continuous_coef
generator_coef = parser.generator_coef
discriminator_coef = parser.discriminator_coef
# load data
imageName, imageDict = locate(data_path,
styles=styles,
max_label=categorical_cardinality)
imageNum = len(imageName)
image1 = tf.placeholder(tf.float32,
[None, image_size, image_size, channel_size],
name="image1")
is_training = tf.placeholder(tf.bool, [], name="is_training")
zc_vector = tf.placeholder(tf.float32,
[None, class_dim + style_dim + continuous_dim],
name="zc_vector")
reconstruct_loss, continuous_loss, generator_loss, discriminator_loss, image_reconstruct = infogan(
zc_vector, image1, kernel, stride, class_dim, style_dim,
continuous_dim, is_training, reconstruct_coef, continuous_coef,
generator_coef, discriminator_coef, 'infogan')
decoder_variables = scope_variables('infogan/decoder')
discriminator_variables = scope_variables('infogan/discriminator')
mutual_variables = scope_variables('infogan/mutual_fc')
reconstruct_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
continuous_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
generator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
discriminator_solver = tf.train.AdamOptimizer(learning_rate=lr, beta1=0.5)
reconstruct_train = reconstruct_solver.minimize(reconstruct_loss,
var_list=decoder_variables)
continuous_train = continuous_solver.minimize(continuous_loss,
var_list=decoder_variables +
discriminator_variables +
mutual_variables)
generator_train = generator_solver.minimize(generator_loss,
var_list=decoder_variables)
discriminator_train = discriminator_solver.minimize(
discriminator_loss, var_list=discriminator_variables)
idxes_1 = np.arange(imageNum, dtype=np.int32)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = parser.gpu_fraction
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
np.random.shuffle(idxes_1)
reconstruct_losses = []
continuous_losses = []
generator_losses = []
discriminator_losses = []
for idx in range(0, imageNum, batch_size):
image1_batch = loader(imageName[idxes_1[idx:idx + batch_size]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
vector = get_vector(imageName[idxes_1[idx:idx + batch_size]],
style_dim, continuous_dim)
feed_dict_training = {
image1: image1_batch,
zc_vector: vector,
is_training: True
}
# decoder
_, _, _, _reconstruct_loss, _continuous_loss, _generator_loss = sess.run(
[
reconstruct_train, continuous_train, generator_train,
reconstruct_loss, continuous_loss, generator_loss
],
feed_dict=feed_dict_training)
reconstruct_losses.append(_reconstruct_loss)
continuous_losses.append(_continuous_loss)
generator_losses.append(_generator_loss)
# discriminator
_, _discriminator_loss = sess.run(
[discriminator_train, discriminator_loss],
feed_dict=feed_dict_training)
discriminator_losses.append(_discriminator_loss)
print(
'epoch: %d\nreconstruct_loss: %f, continuous_loss: %f\ngenerator_loss: %f, discriminator_loss: %f\n'
% (epoch, get_mean(reconstruct_losses),
get_mean(continuous_losses), get_mean(generator_losses),
get_mean(discriminator_losses)))
image1_plot = loader(imageName[idxes_1[0:10]],
desired_height=image_size,
desired_width=image_size,
value_range=(0.0, 1.0),
force_grayscale=force_grayscale)
feed_dict_not_training = {
image1: image1_plot,
zc_vector: vector,
is_training: False
}
_image_reconstruct = sess.run(image_reconstruct,
feed_dict=feed_dict_not_training)
plot(image1_plot, _image_reconstruct, epoch)
saver.save(sess, 'ckpt/infogan/model')