def main():
images, labels = dataset.load_test_images()
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
z, _x = sess.run([t.z_r, t.x_r], feed_dict={t.x: _images})
plot.scatter_labeled_z(z, label_id, dir=config.ckpt_dir)
plot.tile_images(_x[:100], dir=config.ckpt_dir)
plot.plot_loss_tendency(config.ckpt_dir)
hist_value, hist_head = plot.load_pickle_to_data(config.ckpt_dir)
for loss_name in ['reconstruction']:
plot.plot_loss_trace(hist_value[loss_name], loss_name, config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator'],
hist_value['generator'], 'z', config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_z'],
hist_value['generator_z'], 'z',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_img'],
hist_value['generator_img'], 'z',
config.ckpt_dir)
def main():
# load MNIST images
images, labels = dataset.load_test_images()
# Settings
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
representation, x_reconstruction = sess.run([t.yz, t.x_r],
feed_dict={t.x: _images})
plot.scatter_labeled_z(representation, label_id, dir=config.ckpt_dir)
plot.tile_images(x_reconstruction[:100], dir=config.ckpt_dir)
# z distributed plot
num_segments = 20
limit = (-5, 5)
x_values = np.linspace(limit[0], limit[1], num_segments)
y_values = np.linspace(limit[0], limit[1], num_segments)
vacant = np.zeros((28 * num_segments, 28 * num_segments))
for i, x_element in enumerate(x_values):
for j, y_element in enumerate(y_values):
x_reconstruction = sess.run(
t.x_r,
feed_dict={
t.yz: np.reshape([x_element, y_element], [1, 2])
})
vacant[(num_segments - 1 - i) * 28:(num_segments - i) * 28,
j * 28:(j + 1) * 28] = x_reconstruction.reshape(28, 28)
vacant = (vacant + 1) / 2
pylab.figure(figsize=(10, 10), dpi=400, facecolor='white')
pylab.imshow(vacant, cmap='gray', origin='upper')
pylab.tight_layout()
pylab.axis('off')
pylab.savefig("{}/clusters.png".format(config.ckpt_dir))
# loss part
hist_value, hist_head = plot.load_pickle_to_data(config.ckpt_dir)
for loss_name in ['reconstruction', 'supervised']:
plot.plot_loss_trace(hist_value[loss_name], loss_name,
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_y'],
hist_value['generator_y'], 'y',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_z'],
hist_value['generator_z'], 'z',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['validation_accuracy'],
hist_value['transform'],
'validation_accuracy', config.ckpt_dir)
def main(): images, labels = dataset.load_test_images() num_scatter = len(images) y_distribution, z = aae.encode_x_yz(images, apply_softmax=False, test=True) y = aae.argmax_onehot_from_unnormalized_distribution(y_distribution) representation = aae.to_numpy(aae.encode_yz_representation(y, z, test=True)) plot.scatter_labeled_z(representation, labels, dir=args.plot_dir)
def main():
images, labels = dataset.load_test_images()
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
x, z_respresentation, x_construction = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
z, _x = sess.run([z_respresentation, x_construction],
feed_dict={x: _images})
scatter_labeled_z(z, label_id, dir=config.ckpt_dir)
tile_images(_x[:100], dir=config.ckpt_dir)
plot_loss_tendency(config.ckpt_dir)
def main(run_load_from_file=False):
# load MNIST images
images, labels = dataset.load_test_images()
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
max_epoch = 510
num_trains_per_epoch = 500
batch_size_u = 100
# training
with tf.device(config.device):
h = build_graph()
sess_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
saver = tf.train.Saver(max_to_keep=2)
with tf.Session(config=sess_config) as sess:
'''
Load from checkpoint or start a new session
'''
if run_load_from_file:
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
training_epoch_loss, _ = pickle.load(
open(config.ckpt_dir + '/pickle.pkl', 'rb'))
else:
sess.run(tf.global_variables_initializer())
training_epoch_loss = []
# Recording loss per epoch
process = Process()
for epoch in range(max_epoch):
process.start_epoch(epoch, max_epoch)
'''
Learning rate generator
'''
learning_rate = 0.0001
# Recording loss per iteration
sum_loss_reconstruction = 0
sum_loss_discrminator_z = 0
sum_loss_discrminator_img = 0
sum_loss_generator_z = 0
sum_loss_generator_img = 0
process_iteration = Process()
for i in range(num_trains_per_epoch):
process_iteration.start_epoch(i, num_trains_per_epoch)
# Inputs
'''
_l -> labeled
_u -> unlabeled
'''
images_u = dataset.sample_unlabeled_data(images, batch_size_u)
if config.distribution_sampler == 'swiss_roll':
z_true_u = sampler.swiss_roll(batch_size_u, config.ndim_z,
config.num_types_of_label)
elif config.distribution_sampler == 'gaussian_mixture':
z_true_u = sampler.gaussian_mixture(
batch_size_u, config.ndim_z, config.num_types_of_label)
elif config.distribution_sampler == 'uniform_desk':
z_true_u = sampler.uniform_desk(batch_size_u,
config.ndim_z,
radius=2)
elif config.distribution_sampler == 'gaussian':
z_true_u = sampler.gaussian(batch_size_u,
config.ndim_z,
var=1)
elif config.distribution_sampler == 'uniform':
z_true_u = sampler.uniform(batch_size_u,
config.ndim_z,
minv=-1,
maxv=1)
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
# adversarial phase for discriminator_z
images_u_s = dataset.sample_unlabeled_data(
images, batch_size_u)
_, loss_discriminator_z = sess.run([h.opt_dz, h.loss_dz],
feed_dict={
h.x: images_u,
h.z: z_true_u,
h.lr: learning_rate
})
_, loss_discriminator_img = sess.run([h.opt_dimg, h.loss_dimg],
feed_dict={
h.x: images_u,
h.x_s: images_u_s,
h.lr: learning_rate
})
# adversarial phase for generator
_, loss_generator_z = sess.run([h.opt_e, h.loss_e],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
_, loss_generator_img = sess.run([h.opt_d, h.loss_d],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
sum_loss_reconstruction += loss_reconstruction
sum_loss_discrminator_z += loss_discriminator_z
sum_loss_discrminator_img += loss_discriminator_img
sum_loss_generator_z += loss_generator_z
sum_loss_generator_img += loss_generator_img
if i % 1000 == 0:
process_iteration.show_table_2d(
i, num_trains_per_epoch, {
'reconstruction':
sum_loss_reconstruction / (i + 1),
'discriminator_z':
sum_loss_discrminator_z / (i + 1),
'discriminator_img':
sum_loss_discrminator_img / (i + 1),
'generator_z':
sum_loss_generator_z / (i + 1),
'generator_img':
sum_loss_generator_img / (i + 1),
})
average_loss_per_epoch = [
sum_loss_reconstruction / num_trains_per_epoch,
sum_loss_discrminator_z / num_trains_per_epoch,
sum_loss_discrminator_img / num_trains_per_epoch,
sum_loss_generator_z / num_trains_per_epoch,
sum_loss_generator_img / num_trains_per_epoch,
(sum_loss_discrminator_z + sum_loss_discrminator_img) /
num_trains_per_epoch,
(sum_loss_generator_z + sum_loss_generator_img) /
num_trains_per_epoch
]
training_epoch_loss.append(average_loss_per_epoch)
training_loss_name = [
'reconstruction', 'discriminator_z', 'discriminator_img',
'generator_z', 'generator_img', 'discriminator', 'generator'
]
if epoch % 1 == 0:
process.show_bar(
epoch, max_epoch, {
'loss_r': average_loss_per_epoch[0],
'loss_d': average_loss_per_epoch[5],
'loss_g': average_loss_per_epoch[6]
})
plt.scatter_labeled_z(
sess.run(h.z_r, feed_dict={h.x: images[:1000]}),
[int(var) for var in labels[:1000]],
dir=config.log_dir,
filename='z_representation-{}'.format(epoch))
if epoch % 10 == 0:
saver.save(sess,
os.path.join(config.ckpt_dir, 'model_ckptpoint'),
global_step=epoch)
pickle.dump((training_epoch_loss, training_loss_name),
open(config.ckpt_dir + '/pickle.pkl', 'wb'))
description))
if __name__ == "__main__":
# main
main()
file_path = '{}/{}-metric_plus.pkl'.format(config.logs_path,
config.description)
[q_errors, r_adjs, z_adjs, z_true], name = pickle_load(file_path)
# Load data
# latent space
label = np.tile(np.array(dp.test.e), [test_times, 1])
scatter_labeled_z(z_adjs,
label,
dir=config.latent_path,
filename='latent-Q2')
vis.kdeplot(z_adjs[:, 0],
z_adjs[:, 1],
config.latent_path,
name='Kde_Latent_Space')
vis.jointplot(z_adjs[:, 0],
z_adjs[:, 1],
config.latent_path,
name='Scatter_latent')
vis.jointplot(z_true[:, 0],
z_true[:, 1],
config.latent_path,
name='Scatter_latent')
#
def main(): images, labels = dataset.load_test_images() num_scatter = len(images) x, _, label_ids = dataset.sample_labeled_data(images, labels, num_scatter) z = aae.to_numpy(aae.encode_x_z(x, test=True)) plot.scatter_labeled_z(z, label_ids, dir=args.plot_dir)
def main(run_load_from_file=False):
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
max_epoch = 510
num_trains_per_epoch = 500
batch_size_l = 100
batch_size_u = 100
# create semi-supervised split
# Load minist images
images, labels = dataset.load_train_images()
num_labeled_data = 10000
num_types_of_label = 11 # additional label corresponds to unlabeled data
training_images_l, training_labels_l, training_images_u, _, _ = dataset.create_semisupervised(images, labels, 0, num_labeled_data, num_types_of_label)
# training
with tf.device(config.device):
h = build_graph()
sess_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
saver = tf.train.Saver(max_to_keep=2)
with tf.Session(config=sess_config) as sess:
'''
Load from checkpoint or start a new session
'''
if run_load_from_file:
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
training_epoch_loss, _ = pickle.load(open(config.ckpt_dir + '/pickle.pkl', 'rb'))
else:
sess.run(tf.global_variables_initializer())
training_epoch_loss = []
# Recording loss per epoch
process = Process()
for epoch in range(max_epoch):
process.start_epoch(epoch, max_epoch)
'''
Learning rate generator
'''
learning_rate = opt.ladder_learning_rate(epoch + len(training_epoch_loss))
# Recording loss per iteration
sum_loss_reconstruction = 0
sum_loss_discrminator = 0
sum_loss_generator = 0
process_iteration = Process()
for i in range(num_trains_per_epoch):
process_iteration.start_epoch(i, num_trains_per_epoch)
# Inputs
'''
_l -> labeled
_u -> unlabeled
'''
images_l, label_onehot_l, label_id_l = dataset.sample_labeled_data(training_images_l, training_labels_l, batch_size_l, ndim_y=num_types_of_label)
images_u = dataset.sample_unlabeled_data(training_images_u, batch_size_u)
onehot = np.zeros((1, num_types_of_label), dtype=np.float32)
onehot[-1] = 1
label_onehot_u = np.repeat(onehot, batch_size_u, axis=0)
z_true_l = sampler.supervised_swiss_roll(batch_size_l, config.ndim_z, label_id_l, num_types_of_label - 1)
z_true_u = sampler.swiss_roll(batch_size_u, config.ndim_z, num_types_of_label - 1)
# z_true_l = sampler.supervised_gaussian_mixture(batch_size_l, config.ndim_z, label_id_l, num_types_of_label - 1)
# z_true_u = sampler.gaussian_mixture(batch_size_u, config.ndim_z, num_types_of_label - 1)
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r], feed_dict={
h.x: images_u,
h.lr: learning_rate
})
# adversarial phase for discriminator
_, loss_discriminator_l = sess.run([h.opt_d, h.loss_d], feed_dict={
h.x: images_l,
h.label: label_onehot_l,
h.z: z_true_l,
h.lr: learning_rate
})
_, loss_discriminator_u = sess.run([h.opt_d, h.loss_d], feed_dict={
h.x: images_u,
h.label: label_onehot_u,
h.z: z_true_u,
h.lr: learning_rate
})
loss_discriminator = loss_discriminator_l + loss_discriminator_u
# adversarial phase for generator
_, loss_generator_l= sess.run([h.opt_e, h.loss_e,], feed_dict={
h.x: images_l,
h.label: label_onehot_l,
h.lr: learning_rate
})
_, loss_generator_u = sess.run([h.opt_e, h.loss_e], feed_dict={
h.x: images_u,
h.label: label_onehot_u,
h.lr: learning_rate
})
loss_generator = loss_generator_l + loss_generator_u
sum_loss_reconstruction += loss_reconstruction / batch_size_u
sum_loss_discrminator += loss_discriminator
sum_loss_generator += loss_generator
if i % 1000 == 0:
process_iteration.show_table_2d(i, num_trains_per_epoch, {
'reconstruction': sum_loss_reconstruction / (i + 1),
'discriminator': sum_loss_discrminator / (i + 1),
'generator': sum_loss_generator / (i + 1),
})
average_loss_per_epoch = [
sum_loss_reconstruction / num_trains_per_epoch,
sum_loss_discrminator / num_trains_per_epoch,
sum_loss_generator / num_trains_per_epoch,
]
training_epoch_loss.append(average_loss_per_epoch)
training_loss_name = [
'reconstruction',
'discriminator',
'generator'
]
if epoch % 1 == 0:
process.show_bar(epoch, max_epoch, {
'loss_r': average_loss_per_epoch[0],
'loss_d': average_loss_per_epoch[1],
'loss_g': average_loss_per_epoch[2]
})
plt.tile_images(sess.run(h.x_, feed_dict={h.x: images_u}),
dir=config.log_dir,
filename='x_rec_epoch_{}'.format(str(epoch).zfill(3)))
plt.scatter_labeled_z(sess.run(h.z_r, feed_dict={h.x: images[:1000]}), [int(var) for var in labels[:1000]],
dir=config.log_dir,
filename='z_representation-{}'.format(epoch))
if epoch % 10 == 0:
saver.save(sess, os.path.join(config.ckpt_dir, 'model_ckptpoint'), global_step=epoch)
pickle.dump((training_epoch_loss, training_loss_name), open(config.ckpt_dir + '/pickle.pkl', 'wb'))
plt.plot_double_scale_trend(config.ckpt_dir)