def msg_gan(image_inputs, noise_inputs, latent_size, res_building,
minibatch_size):
# Multi-scaled input images
real_inputs = []
for factor in [2**res for res in range(3, -1, -1)]:
real_input = layers.downscale2d(image_inputs, factor=factor)
real_input = layers.upscale2d(real_input, factor=factor)
real_inputs += [real_input]
# Define networks
generator = network.Network('generator',
msg_generator,
noise_inputs,
res_building=res_building,
latent_size=latent_size)
discriminator = network.Network('discriminator',
msg_discriminator,
real_inputs,
res_building=res_building,
latent_size=latent_size)
# Retrieve network outputs
fake_images = generator(noise_inputs)
fake_outputs = discriminator(fake_images)
real_outputs = discriminator(real_inputs)
# Losses
gen_loss, disc_loss = losses.RelativisticAverageBCE(
real_outputs, fake_outputs)
disc_loss += losses.GradientPenaltyMSG(discriminator, real_inputs,
fake_images, minibatch_size)
# disc_loss += losses.EpsilonPenalty(real_outputs)
return gen_loss, disc_loss, fake_images
def pro_gan(image_inputs, noise_inputs, latent_size, res_building,
res_training, minibatch_size):
# Define networks
generator = network.Network('generator',
pro_generator,
noise_inputs,
res_building=res_building,
res_training=res_training,
latent_size=latent_size)
discriminator = network.Network('discriminator',
pro_discriminator,
image_inputs,
res_building=res_building,
res_training=res_training,
latent_size=latent_size)
# Retrieve network outputs
fake_images = generator(noise_inputs)
fake_outputs = discriminator(fake_images)
real_outputs = discriminator(image_inputs)
# Losses
gen_loss, disc_loss = losses.RelativisticAverageBCE(
real_outputs, fake_outputs)
disc_loss += losses.GradientPenalty(discriminator, image_inputs,
fake_images, minibatch_size)
disc_loss += losses.EpsilonPenalty(real_outputs)
return gen_loss, disc_loss, fake_images
def infer_pro_disc():
"""
Train the generative adverserial network.
"""
# Graph definition: inputs, model, loss, optimizer, initializer
print('Graph building...')
# Inputs
image_inputs = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[None, 256, 256, 1],
name='image_inputs')
res_training = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[],
name='res_training')
# Network
discriminator = network.Network('discriminator',
pro_discriminator,
image_inputs,
res_building=8,
res_training=res_training,
latent_size=latent_size)
disc_vars = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
outputs = discriminator(image_inputs)
outputs = tf.nn.sigmoid(outputs)
# Initializer
init = tf.compat.v1.global_variables_initializer()
print('Done: graph built.')
inputs = getattr(dataset,
config.data_initilize)(**config.data_initilize_kwargs)
select_minibatch = partial(getattr(dataset, config.data_selector), inputs)
# Saver
saver = tf.compat.v1.train.Saver(var_list=disc_vars)
# Training --> use the configuration file
print('Training...')
with tf.compat.v1.Session() as sess:
# Initialize
init.run()
# Restore parameters
saver.restore(sess, restore)
# Run
n = 20
name = 'logs/' + restore[restore.rfind('/') + 1:] + '.npy'
out_tot = np.load(name)
win = []
for s in range(5):
minibatch = select_minibatch(crt_img=n * s,
res=8,
minibatch_size=n)
# minibatch = out_tot[n*s:n*(s+1),:,:]
feed_dict = {image_inputs: minibatch, res_training: 8}
out = sess.run(outputs, feed_dict)
win += [out]
print(np.mean((np.array(win).flatten() > 0.5).astype(np.float32)))
def classification(
inputs, # Input images
labels, # Input labels
nbof_labels, # Number of different labels (number of identities)
training, # Use dropout or not? (Training or not training?)
regularizer_rate=0): # Use weight regularization?
net = network.Network('classification', wideresnet_se, inputs, training, nbof_labels, regularizer_rate=regularizer_rate)
logit = net(inputs)
# with tf.compat.v1.variable_scope('classification'):
# emb = deep_cnn_v1(inputs, training, 128, regularizer_rate=regularizer_rate)
# emb = tf.nn.l2_normalize(emb)
# logit = losses.cosineface_losses(emb, labels, nbof_labels, regularizer_rate=regularizer_rate)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logit, labels=labels))
reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
reg_losses = tf.compat.v1.add_n(reg_losses, name='reg_loss')
return logit, loss, reg_losses
def infer_pro_gan():
"""
Train the generative adverserial network.
"""
# Graph definition: inputs, model, loss, optimizer, initializer
print('Graph building...')
# Inputs
noise_inputs = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[None, latent_size],
name='noise_inputs')
res_training = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[],
name='res_training')
# Network
generator = network.Network('generator',
pro_generator,
noise_inputs,
res_building=8,
res_training=res_training,
latent_size=latent_size)
gen_vars = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope='generator')
outputs = generator(noise_inputs)
# outputs = tf.math.reduce_mean(outputs, axis=-1)
# outputs = tf.clip_by_value(outputs, -1., 1.)
# outputs = outputs/2 + 0.5
# outputs *= 128
# discriminator = network.Network('discriminator', pro_discriminator, image_inputs, res_building=res_building, res_training=res_training, latent_size=latent_size)
# Initializer
init = tf.compat.v1.global_variables_initializer()
print('Done: graph built.')
# Saver
saver = tf.compat.v1.train.Saver(var_list=gen_vars)
# Training --> use the configuration file
print('Training...')
with tf.compat.v1.Session() as sess:
# Initialize
init.run()
# Restore parameters
saver.restore(sess, restore)
# Run
n = 5
out_tot = np.empty((0, 256, 256, 1))
for s in range(4):
feed_dict = {
noise_inputs: 2 * np.random.random((n * n, latent_size)) - 1,
res_training: 8
}
out = sess.run(outputs, feed_dict)
out_tot = np.append(out_tot, out, axis=0)
# for i in range(len(out)):
# im = out[i]*255
# im = im.astype(np.uint8)
# sk.io.imsave('logs/images/{}.jpg'.format(s*len(out)+i), im)
# print(out.shape)
# # fig = plt.figure()
# # for i in range(n):
# # for j in range(n):
# # plt.subplot(n,n,i*n+j+1)
# # plt.imshow(out[i*n+j])
# # plt.show()
# # print(out[0,:5,:32])
# for i in range(len(out)):
# # keyed_pianoroll2midi(out[i], file_path='../../../data/music/test_{}.mid'.format(i))
# # im = out[i]*255/np.max(out[i])
# # im = im.astype(np.uint8)
# keyed_pianoroll2midi(out[i]*100, file_path='logs/musics/{}.mid'.format(s*len(out)+i))
name = 'logs/' + restore[restore.rfind('/') + 1:] + '.npy'
print(name)
print(out_tot.shape)
np.save(name, np.array(out_tot))
def infer_pro_win(res=2):
"""
Train the generative adverserial network.
"""
# Graph definition: inputs, model, loss, optimizer, initializer
print('Graph building...')
# Inputs
noise_inputs = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[None, latent_size],
name='noise_inputs')
res_training = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[],
name='res_training')
image_inputs = tf.compat.v1.placeholder(tf.compat.v1.float32,
shape=[None, 256, 256, 1],
name='image_inputs')
# Network
generator = network.Network('generator',
pro_generator,
noise_inputs,
res_building=8,
res_training=res_training,
latent_size=latent_size)
discriminator = network.Network('discriminator',
pro_discriminator,
image_inputs,
res_building=8,
res_training=res_training,
latent_size=latent_size)
var_list = tf.compat.v1.global_variables()
gen_outputs = generator(noise_inputs)
fake_outputs = discriminator(gen_outputs)
real_outputs = discriminator(image_inputs)
rf = real_outputs - tf.compat.v1.math.reduce_mean(fake_outputs)
fr = fake_outputs - tf.compat.v1.math.reduce_mean(real_outputs)
# outputs = tf.math.reduce_mean(outputs, axis=-1)
# outputs = tf.clip_by_value(outputs, -1., 1.)
# outputs = outputs/2 + 0.5
# outputs *= 128
# Initializer
init = tf.compat.v1.global_variables_initializer()
print('Done: graph built.')
inputs = getattr(dataset,
config.data_initilize)(**config.data_initilize_kwargs)
select_minibatch = partial(getattr(dataset, config.data_selector), inputs)
# Saver
saver = tf.compat.v1.train.Saver(var_list=var_list)
# Training --> use the configuration file
print('Training...')
with tf.compat.v1.Session() as sess:
# Initialize
init.run()
# Restore parameters
saver.restore(sess, restore)
# Run
n = 4
win = []
for s in range(4):
minibatch = select_minibatch(crt_img=n * n * s,
res=8,
minibatch_size=n * n)
feed_dict = {
noise_inputs: 2 * np.random.random((n * n, latent_size)) - 1,
res_training: res,
image_inputs: minibatch,
}
rf_, fr_ = sess.run([rf, fr], feed_dict)
print(rf_)
print(fr_)
win = np.append(win, (rf_ < .0).astype(np.float32))
win = np.append(win, (fr_ > .0).astype(np.float32))
# out_tot += [out]
print(np.mean(win))
# plt.show()
# Load graph
# Network parameters
image_inputs = tf.placeholder(
tf.float32,
shape=[None, config.image_size, config.image_size, 3],
name='image_inputs')
image_normed = image_inputs / 127.5
image_normed = image_normed - 1
label_inputs = tf.placeholder(tf.int64, shape=[None], name='label_inputs')
nbof_labels = config.nbof_labels
emb_size = config.latent_size
training = tf.placeholder(tf.bool, shape=[], name='training')
# Network
net = network.Network('classification', network_classification.v3,
image_normed, labels, emb_size, training, nbof_labels)
logit = net(image_normed)
var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES,
scope='classification')
pred = tf.nn.softmax(logit)
pred = tf.argmax(pred, axis=1)
# Load network
# saver_path = 'logs/Cifar10-classification_v1-run_20191026172524/models/'
# saver_path = 'logs/Cifar10-classification_v1-run_20191028120938/models/' # without batch norm (1568)
saver_path = 'logs/Cifar10-classification_v1-run_20191028144237/models/' # with custom norm custom
# saver_path = 'logs/Cifar10-classification_v1-run_20191028151112/models/' # with keras batch norm
saver = tf.compat.v1.train.Saver(var_list=var_list)
# Run session on samples
with tf.Session() as sess:
saver.restore(sess, saver_path + 'model-336.ckpt')
