def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = model.config
# settings
max_epoch = 10000
num_trains_per_epoch = 5000
num_validation_data = 10000
batchsize = 128
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
training_images, training_labels, validation_images, validation_labels = dataset.split_data(images, labels, num_validation_data, seed=args.seed)
training_labels = np.random.randint(0, config.num_classes, training_labels.size).astype(np.int32)
validation_labels = np.random.randint(0, config.num_classes, validation_labels.size).astype(np.int32)
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
image_batch, label_batch = dataset.sample_data(training_images, training_labels, batchsize, binarize=False)
image_batch = np.reshape(image_batch, (-1, 1, 28, 28))
distribution = model.discriminate(image_batch, apply_softmax=False)
loss = F.softmax_cross_entropy(distribution, model.to_variable(label_batch))
sum_loss += float(loss.data)
model.backprop(loss)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
model.save(args.model_dir)
train_accuracy = compute_accuracy(training_images, training_labels)
validation_accuracy = compute_accuracy(validation_images, validation_labels)
progress.show(num_trains_per_epoch, num_trains_per_epoch, {
"loss": sum_loss / num_trains_per_epoch,
"accuracy (validation)": validation_accuracy,
"accuracy (train)": train_accuracy,
})
# write accuracy to csv
csv_results.append([epoch, train_accuracy, validation_accuracy, progress.get_total_time()])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "train_accuracy", "validation_accuracy", "min"]
data.to_csv("{}/result.csv".format(args.model_dir))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
discriminator_config = gan.config_discriminator
generator_config = gan.config_generator
# settings
max_epoch = 1000
num_updates_per_epoch = 500
plot_interval = 5
batchsize_true = 100
batchsize_fake = batchsize_true
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# training
progress = Progress()
for epoch in xrange(1, max_epoch + 1):
progress.start_epoch(epoch, max_epoch)
sum_loss_critic = 0
sum_loss_generator = 0
for t in xrange(num_updates_per_epoch):
for k in xrange(discriminator_config.num_critic):
# clamp parameters to a cube
gan.clip_discriminator_weights()
# gan.decay_discriminator_weights()
# sample true data from data distribution
images_true = dataset.sample_data(images, batchsize_true, binarize=False)
# sample fake data from generator
images_fake = gan.generate_x(batchsize_fake)
images_fake.unchain_backward()
fw_true, activations_true = gan.discriminate(images_true)
fw_fake, _ = gan.discriminate(images_fake)
loss_critic = -F.sum(fw_true - fw_fake) / batchsize_true
sum_loss_critic += float(loss_critic.data) / discriminator_config.num_critic
# update discriminator
gan.backprop_discriminator(loss_critic)
# generator loss
images_fake = gan.generate_x(batchsize_fake)
fw_fake, activations_fake = gan.discriminate(images_fake)
loss_generator = -F.sum(fw_fake) / batchsize_fake
# feature matching
if discriminator_config.use_feature_matching:
features_true = activations_true[-1]
features_true.unchain_backward()
if batchsize_true != batchsize_fake:
images_fake = gan.generate_x(batchsize_true)
_, activations_fake = gan.discriminate(images_fake, apply_softmax=False)
features_fake = activations_fake[-1]
loss_generator += F.mean_squared_error(features_true, features_fake)
# update generator
gan.backprop_generator(loss_generator)
sum_loss_generator += float(loss_generator.data)
if t % 10 == 0:
progress.show(t, num_updates_per_epoch, {})
gan.save(args.model_dir)
progress.show(num_updates_per_epoch, num_updates_per_epoch, {
"wasserstein": -sum_loss_critic / num_updates_per_epoch,
"loss_g": sum_loss_generator / num_updates_per_epoch,
})
if epoch % plot_interval == 0 or epoch == 1:
plot(filename="epoch_{}_time_{}min".format(epoch, progress.get_total_time()))
def main():
# load MNIST images
train_images, train_labels = dataset.load_train_images()
# config
config = aae.config
# settings
# _l -> labeled
# _u -> unlabeled
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
train_images, train_labels, batchsize)
# reconstruction phase
z_l = aae.encode_x_z(images_l)
reconstruction_l = aae.decode_yz_x(label_onehot_l, z_l)
loss_reconstruction = F.mean_squared_error(
aae.to_variable(images_l), reconstruction_l)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
images_l = dataset.sample_labeled_data(train_images, train_labels,
batchsize)[0]
z_fake_l = aae.encode_x_z(images_l)
z_true_l = sampler.gaussian(batchsize,
config.ndim_z,
mean=0,
var=1)
dz_true = aae.discriminate_z(z_true_l, apply_softmax=False)
dz_fake = aae.discriminate_z(z_fake_l, apply_softmax=False)
loss_discriminator = F.softmax_cross_entropy(
dz_true, class_true) + F.softmax_cross_entropy(
dz_fake, class_fake)
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
images_l = dataset.sample_labeled_data(train_images, train_labels,
batchsize)[0]
z_fake_l = aae.encode_x_z(images_l)
dz_fake = aae.discriminate_z(z_fake_l, apply_softmax=False)
loss_generator = F.softmax_cross_entropy(dz_fake, class_true)
aae.backprop_generator(loss_generator)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
})
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = aae.config
# settings
# _l -> labeled
# _u -> unlabeled
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images,
labels,
num_validation_data,
num_labeled_data,
num_types_of_label,
seed=args.seed)
print training_labels_l
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize_u, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize_u, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_supervised = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l,
config.ndim_x, config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u,
config.ndim_x)
# reconstruction phase
q_y_x_u, z_u = aae.encode_x_yz(images_u, apply_softmax=True)
reconstruction_u = aae.decode_yz_x(q_y_x_u, z_u)
loss_reconstruction = F.mean_squared_error(
aae.to_variable(images_u), reconstruction_u)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
z_true_u = sampler.gaussian(batchsize_u,
config.ndim_z,
mean=0,
var=1)
y_true_u = sampler.onehot_categorical(batchsize_u, config.ndim_y)
discrimination_z_true = aae.discriminate_z(z_true_u,
apply_softmax=False)
discrimination_y_true = aae.discriminate_y(y_true_u,
apply_softmax=False)
discrimination_z_fake = aae.discriminate_z(z_fake_u,
apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u,
apply_softmax=False)
loss_discriminator_z = F.softmax_cross_entropy(
discrimination_z_true, class_true) + F.softmax_cross_entropy(
discrimination_z_fake, class_fake)
loss_discriminator_y = F.softmax_cross_entropy(
discrimination_y_true, class_true) + F.softmax_cross_entropy(
discrimination_y_fake, class_fake)
loss_discriminator = loss_discriminator_z + loss_discriminator_y
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
discrimination_z_fake = aae.discriminate_z(z_fake_u,
apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u,
apply_softmax=False)
loss_generator_z = F.softmax_cross_entropy(discrimination_z_fake,
class_true)
loss_generator_y = F.softmax_cross_entropy(discrimination_y_fake,
class_true)
loss_generator = loss_generator_z + loss_generator_y
aae.backprop_generator(loss_generator)
# supervised phase
unnormalized_q_y_x_l, z_l = aae.encode_x_yz(images_l,
apply_softmax=False)
loss_supervised = F.softmax_cross_entropy(
unnormalized_q_y_x_l, aae.to_variable(label_ids_l))
aae.backprop_generator(loss_supervised)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_supervised += float(loss_supervised.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
# validation phase
# split validation data to reduce gpu memory consumption
images_v, _, label_ids_v = dataset.sample_labeled_data(
validation_images, validation_labels, num_validation_data,
config.ndim_x, config.ndim_y)
images_v_segments = np.split(images_v, num_validation_data // 500)
label_ids_v_segments = np.split(label_ids_v,
num_validation_data // 500)
num_correct = 0
for images_v, labels_v in zip(images_v_segments, label_ids_v_segments):
predicted_labels = aae.argmax_x_label(images_v, test=True)
for i, label in enumerate(predicted_labels):
if label == labels_v[i]:
num_correct += 1
validation_accuracy = num_correct / float(num_validation_data)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_s": sum_loss_supervised / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
"accuracy": validation_accuracy
})
# write accuracy to csv
csv_results.append([epoch, validation_accuracy])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy"]
data.to_csv("{}/result.csv".format(args.model_dir))
import numpy as np
import os, sys, time, math, pylab
import seaborn as sns
sns.set(font_scale=2.5)
import pandas as pd
import matplotlib.pyplot as plt
from chainer import cuda
from chainer import functions as F
sys.path.append(os.path.split(os.getcwd())[0])
import dataset
from progress import Progress
from model import imsat, params
from args import args
# load MNIST
train_images, train_labels = dataset.load_train_images()
train_images_l, train_labels_l, train_images_u = dataset.create_semisupervised(
train_images, train_labels, num_labeled_data=args.num_labeled_data)
print("labeled: ", len(train_images_l))
print("unlabeled: ", len(train_images_u))
test_images, test_labels = dataset.load_test_images()
# config
config = imsat.config
def compute_accuracy(images, labels_true):
predict_counts = np.zeros((47, config.num_clusters), dtype=np.float32)
# to avoid cudaErrorMemoryAllocation: out of memory
num_data_in_segment = len(images) // 20
for n in range(20):
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
discriminator_config = gan.config_discriminator
generator_config = gan.config_generator
# settings
# _l -> labeled
# _u -> unlabeled
# _g -> generated
max_epoch = 1000
num_trains_per_epoch = 500
plot_interval = 5
batchsize_l = 100
batchsize_u = 100
batchsize_g = batchsize_u
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = args.num_labeled
if batchsize_l > num_labeled_data:
batchsize_l = num_labeled_data
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images,
labels,
num_validation_data,
num_labeled_data,
discriminator_config.ndim_output,
seed=args.seed)
print training_labels_l
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_supervised = 0
sum_loss_unsupervised = 0
sum_loss_adversarial = 0
sum_dx_labeled = 0
sum_dx_unlabeled = 0
sum_dx_generated = 0
gan.update_learning_rate(get_learning_rate_for_epoch(epoch))
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l,
training_labels_l,
batchsize_l,
discriminator_config.ndim_input,
discriminator_config.ndim_output,
binarize=False)
images_u = dataset.sample_unlabeled_data(
training_images_u,
batchsize_u,
discriminator_config.ndim_input,
binarize=False)
images_g = gan.generate_x(batchsize_g)
images_g.unchain_backward()
# supervised loss
py_x_l, activations_l = gan.discriminate(images_l,
apply_softmax=False)
loss_supervised = F.softmax_cross_entropy(
py_x_l, gan.to_variable(label_ids_l))
log_zx_l = F.logsumexp(py_x_l, axis=1)
log_dx_l = log_zx_l - F.softplus(log_zx_l)
dx_l = F.sum(F.exp(log_dx_l)) / batchsize_l
# unsupervised loss
# D(x) = Z(x) / {Z(x) + 1}, where Z(x) = \sum_{k=1}^K exp(l_k(x))
# softplus(x) := log(1 + exp(x))
# logD(x) = logZ(x) - log(Z(x) + 1)
# = logZ(x) - log(exp(log(Z(x))) + 1)
# = logZ(x) - softplus(logZ(x))
# 1 - D(x) = 1 / {Z(x) + 1}
# log{1 - D(x)} = log1 - log(Z(x) + 1)
# = -log(exp(log(Z(x))) + 1)
# = -softplus(logZ(x))
py_x_u, _ = gan.discriminate(images_u, apply_softmax=False)
log_zx_u = F.logsumexp(py_x_u, axis=1)
log_dx_u = log_zx_u - F.softplus(log_zx_u)
dx_u = F.sum(F.exp(log_dx_u)) / batchsize_u
loss_unsupervised = -F.sum(
log_dx_u) / batchsize_u # minimize negative logD(x)
py_x_g, _ = gan.discriminate(images_g, apply_softmax=False)
log_zx_g = F.logsumexp(py_x_g, axis=1)
loss_unsupervised += F.sum(F.softplus(
log_zx_g)) / batchsize_u # minimize negative log{1 - D(x)}
# update discriminator
gan.backprop_discriminator(loss_supervised + loss_unsupervised)
# adversarial loss
images_g = gan.generate_x(batchsize_g)
py_x_g, activations_g = gan.discriminate(images_g,
apply_softmax=False)
log_zx_g = F.logsumexp(py_x_g, axis=1)
log_dx_g = log_zx_g - F.softplus(log_zx_g)
dx_g = F.sum(F.exp(log_dx_g)) / batchsize_g
loss_adversarial = -F.sum(
log_dx_g) / batchsize_u # minimize negative logD(x)
# feature matching
if discriminator_config.use_feature_matching:
features_true = activations_l[-1]
features_true.unchain_backward()
if batchsize_l != batchsize_g:
images_g = gan.generate_x(batchsize_l)
_, activations_g = gan.discriminate(images_g,
apply_softmax=False)
features_fake = activations_g[-1]
loss_adversarial += F.mean_squared_error(
features_true, features_fake)
# update generator
gan.backprop_generator(loss_adversarial)
sum_loss_supervised += float(loss_supervised.data)
sum_loss_unsupervised += float(loss_unsupervised.data)
sum_loss_adversarial += float(loss_adversarial.data)
sum_dx_labeled += float(dx_l.data)
sum_dx_unlabeled += float(dx_u.data)
sum_dx_generated += float(dx_g.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
gan.save(args.model_dir)
# validation
images_l, _, label_ids_l = dataset.sample_labeled_data(
validation_images,
validation_labels,
num_validation_data,
discriminator_config.ndim_input,
discriminator_config.ndim_output,
binarize=False)
images_l_segments = np.split(images_l, num_validation_data // 500)
label_ids_l_segments = np.split(label_ids_l,
num_validation_data // 500)
sum_accuracy = 0
for images_l, label_ids_l in zip(images_l_segments,
label_ids_l_segments):
y_distribution, _ = gan.discriminate(images_l,
apply_softmax=True,
test=True)
accuracy = F.accuracy(y_distribution, gan.to_variable(label_ids_l))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_l_segments)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_l": sum_loss_supervised / num_trains_per_epoch,
"loss_u": sum_loss_unsupervised / num_trains_per_epoch,
"loss_g": sum_loss_adversarial / num_trains_per_epoch,
"dx_l": sum_dx_labeled / num_trains_per_epoch,
"dx_u": sum_dx_unlabeled / num_trains_per_epoch,
"dx_g": sum_dx_generated / num_trains_per_epoch,
"accuracy": validation_accuracy,
})
# write accuracy to csv
csv_results.append(
[epoch, validation_accuracy,
progress.get_total_time()])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy", "min"]
data.to_csv("{}/result.csv".format(args.model_dir))
if epoch % plot_interval == 0 or epoch == 1:
plot(filename="epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
def main(run_load_from_file=False):
# load MNIST images
images, labels = dataset.load_train_images()
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
# setting
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 = opt.ladder_learning_rate(epoch +
len(training_epoch_loss))
# Recording loss per iteration
training_loss_set = []
sum_loss_reconstruction = 0
sum_loss_supervised = 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)
# sample from data distribution
images_u = dataset.sample_unlabeled_data(images, batch_size_u)
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
z_true_u = sampler.gaussian(batch_size_u,
config.ndim_z,
mean=0,
var=1)
y_true_u = sampler.onehot_categorical(batch_size_u,
config.ndim_y)
# adversarial phase for discriminator
_, loss_discriminator_y = sess.run([h.opt_dy, h.loss_dy],
feed_dict={
h.x: images_u,
h.y: y_true_u,
h.lr: learning_rate
})
_, 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 = loss_discriminator_y + loss_discriminator_z
# adversarial phase for generator
_, loss_generator_y, loss_generator_z = sess.run(
[h.opt_e, h.loss_gy, h.loss_gz],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
loss_generator = loss_generator_y + loss_generator_z
training_loss_set.append([
loss_reconstruction,
loss_discriminator,
loss_discriminator_y,
loss_discriminator_z,
loss_generator,
loss_generator_z,
loss_generator_y,
])
sum_loss_reconstruction += loss_reconstruction
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),
})
# In end of epoch, summary the loss
average_training_loss_per_epoch = np.mean(
np.array(training_loss_set), axis=0)
# append validation accuracy to the training loss
training_epoch_loss.append(average_training_loss_per_epoch)
loss_name_per_epoch = [
'reconstruction',
'discriminator',
'discriminator_y',
'discriminator_z',
'generator',
'generator_z',
'generator_y',
]
if epoch % 1 == 0:
process.show_bar(
epoch, max_epoch, {
'loss_r': average_training_loss_per_epoch[0],
'loss_d': average_training_loss_per_epoch[1],
'loss_g': average_training_loss_per_epoch[4],
})
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)))
if epoch % 10 == 0:
saver.save(sess,
os.path.join(config.ckpt_dir, 'model_ckptpoint'),
global_step=epoch)
pickle.dump((training_epoch_loss, loss_name_per_epoch),
open(config.ckpt_dir + '/pickle.pkl', 'wb'))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = adgm.config
# settings
max_epoch = 1000
num_trains_per_epoch = 500
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images,
labels,
num_validation_data,
num_labeled_data,
num_types_of_label,
seed=args.seed)
print training_labels_l
# init weightnorm layers
if config.use_weightnorm:
print "initializing weight normalization layers ..."
images_l, label_onehot_l, label_id_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l, config.ndim_x,
config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u, config.ndim_x)
adgm.compute_lower_bound(images_l, label_onehot_l, images_u)
# training
temperature = 1
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_lower_bound_l = 0
sum_lower_bound_u = 0
sum_loss_classifier = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l,
config.ndim_x, config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u,
config.ndim_x)
# lower bound loss using gumbel-softmax
lower_bound, lb_labeled, lb_unlabeled = adgm.compute_lower_bound_gumbel(
images_l, label_onehot_l, images_u, temperature)
loss_lower_bound = -lower_bound
# classification loss
a_l = adgm.encode_x_a(images_l, False)
unnormalized_y_distribution = adgm.encode_ax_y_distribution(
a_l, images_l, softmax=False)
loss_classifier = alpha * F.softmax_cross_entropy(
unnormalized_y_distribution, adgm.to_variable(label_ids_l))
# backprop
adgm.backprop(loss_classifier + loss_lower_bound)
sum_lower_bound_l += float(lb_labeled.data)
sum_lower_bound_u += float(lb_unlabeled.data)
sum_loss_classifier += float(loss_classifier.data)
progress.show(t, num_trains_per_epoch, {})
adgm.save(args.model_dir)
# validation
images_l, _, label_ids_l = dataset.sample_labeled_data(
validation_images, validation_labels, num_validation_data,
config.ndim_x, config.ndim_y)
images_l_segments = np.split(images_l, num_validation_data // 500)
label_ids_l_segments = np.split(label_ids_l,
num_validation_data // 500)
sum_accuracy = 0
for images_l, label_ids_l in zip(images_l_segments,
label_ids_l_segments):
y_distribution = adgm.encode_x_y_distribution(images_l,
softmax=True,
test=True)
accuracy = F.accuracy(y_distribution,
adgm.to_variable(label_ids_l))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_l_segments)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"lb_u": sum_lower_bound_u / num_trains_per_epoch,
"lb_l": sum_lower_bound_l / num_trains_per_epoch,
"loss_spv": sum_loss_classifier / num_trains_per_epoch,
"accuracy": validation_accuracy,
"tmp": temperature,
})
# anneal the temperature
temperature = max(0.5, temperature * 0.999)
# write accuracy to csv
csv_results.append([epoch, validation_accuracy])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy"]
data.to_csv("{}/result.csv".format(args.model_dir))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
discriminator_config = gan.config_discriminator
generator_config = gan.config_generator
# labels
a = discriminator_config.a
b = discriminator_config.b
c = discriminator_config.c
# settings
max_epoch = 1000
num_updates_per_epoch = 500
plot_interval = 5
batchsize_true = 100
batchsize_fake = batchsize_true
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# training
progress = Progress()
for epoch in xrange(1, max_epoch + 1):
progress.start_epoch(epoch, max_epoch)
sum_loss_d = 0
sum_loss_g = 0
for t in xrange(num_updates_per_epoch):
# sample true data from data distribution
images_true = dataset.sample_data(images,
batchsize_true,
binarize=False)
# sample fake data from generator
images_fake = gan.generate_x(batchsize_fake)
images_fake.unchain_backward()
d_true = gan.discriminate(images_true, return_activations=False)
d_fake = gan.discriminate(images_fake, return_activations=False)
loss_d = 0.5 * (F.sum((d_true - b)**2) + F.sum(
(d_fake - a)**2)) / batchsize_true
sum_loss_d += float(loss_d.data)
# update discriminator
gan.backprop_discriminator(loss_d)
# generator loss
images_fake = gan.generate_x(batchsize_fake)
d_fake = gan.discriminate(images_fake, return_activations=False)
loss_g = 0.5 * (F.sum((d_fake - c)**2)) / batchsize_fake
sum_loss_g += float(loss_g.data)
# update generator
gan.backprop_generator(loss_g)
if t % 10 == 0:
progress.show(t, num_updates_per_epoch, {})
gan.save(args.model_dir)
progress.show(
num_updates_per_epoch, num_updates_per_epoch, {
"loss_d": sum_loss_d / num_updates_per_epoch,
"loss_g": sum_loss_g / num_updates_per_epoch,
})
if epoch % plot_interval == 0 or epoch == 1:
plot(filename="epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = aae.config
# settings
# _l -> labeled
# _u -> unlabeled
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# create semi-supervised split
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)
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize_u, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize_u, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_supervised = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l,
training_labels_l,
batchsize_l,
ndim_y=num_types_of_label)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u)
# reconstruction phase
z_u = aae.encode_x_z(images_u)
reconstruction_u = aae.decode_z_x(z_u)
loss_reconstruction = F.mean_squared_error(
aae.to_variable(images_u), reconstruction_u)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
z_fake_u = aae.encode_x_z(images_u)
z_fake_l = aae.encode_x_z(images_l)
onehot = np.zeros((1, num_types_of_label), dtype=np.float32)
onehot[0, -1] = 1 # turn on the extra class
label_onehot_u = np.repeat(onehot, batchsize_u, axis=0)
z_true_l = sampler.supervised_gaussian_mixture(
batchsize_l, config.ndim_z, label_ids_l,
num_types_of_label - 1)
z_true_u = sampler.gaussian_mixture(batchsize_u, config.ndim_z,
num_types_of_label - 1)
dz_true_l = aae.discriminate_z(label_onehot_l,
z_true_l,
apply_softmax=False)
dz_true_u = aae.discriminate_z(label_onehot_u,
z_true_u,
apply_softmax=False)
dz_fake_l = aae.discriminate_z(label_onehot_l,
z_fake_l,
apply_softmax=False)
dz_fake_u = aae.discriminate_z(label_onehot_u,
z_fake_u,
apply_softmax=False)
loss_discriminator = F.softmax_cross_entropy(
dz_true_l, class_true) + F.softmax_cross_entropy(
dz_true_u, class_true) + F.softmax_cross_entropy(
dz_fake_l, class_fake) + F.softmax_cross_entropy(
dz_fake_u, class_fake)
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
z_fake_u = aae.encode_x_z(images_u)
z_fake_l = aae.encode_x_z(images_l)
dz_fake_l = aae.discriminate_z(label_onehot_l,
z_fake_l,
apply_softmax=False)
dz_fake_u = aae.discriminate_z(label_onehot_u,
z_fake_u,
apply_softmax=False)
loss_generator = F.softmax_cross_entropy(
dz_fake_l, class_true) + F.softmax_cross_entropy(
dz_fake_u, class_true)
aae.backprop_generator(loss_generator)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
})
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = vat.config
# settings
max_epoch = 1000
num_trains_per_epoch = 500
batchsize_l = 100
batchsize_u = 200
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images,
labels,
num_validation_data,
num_labeled_data,
config.ndim_y,
seed=args.seed)
print training_labels_l
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_supervised = 0
sum_loss_lds = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l,
training_labels_l,
batchsize_l,
config.ndim_x,
config.ndim_y,
binarize=False)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u,
config.ndim_x,
binarize=False)
# supervised loss
unnormalized_y_distribution = vat.encode_x_y(images_l,
apply_softmax=False)
loss_supervised = F.softmax_cross_entropy(
unnormalized_y_distribution, vat.to_variable(label_ids_l))
# virtual adversarial training
lds_l = -F.sum(vat.compute_lds(images_l)) / batchsize_l
lds_u = -F.sum(vat.compute_lds(images_u)) / batchsize_u
loss_lsd = lds_l + lds_u
# backprop
vat.backprop(loss_supervised + config.lambda_ * loss_lsd)
sum_loss_supervised += float(loss_supervised.data)
sum_loss_lds += float(loss_lsd.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
vat.save(args.model_dir)
# validation
images_l, _, label_ids_l = dataset.sample_labeled_data(
validation_images,
validation_labels,
num_validation_data,
config.ndim_x,
config.ndim_y,
binarize=False)
images_l_segments = np.split(images_l, num_validation_data // 500)
label_ids_l_segments = np.split(label_ids_l,
num_validation_data // 500)
sum_accuracy = 0
for images_l, label_ids_l in zip(images_l_segments,
label_ids_l_segments):
y_distribution = vat.encode_x_y(images_l,
apply_softmax=True,
test=True)
accuracy = F.accuracy(y_distribution, vat.to_variable(label_ids_l))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_l_segments)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_spv": sum_loss_supervised / num_trains_per_epoch,
"loss_lds": sum_loss_lds / num_trains_per_epoch,
"accuracy": validation_accuracy,
})
# write accuracy to csv
csv_results.append(
[epoch, validation_accuracy,
progress.get_total_time()])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy", "min"]
data.to_csv("{}/result.csv".format(args.model_dir))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = aae.config
# settings
# _l -> labeled
# _u -> unlabeled
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images, labels, num_validation_data, num_labeled_data,
num_types_of_label)
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize_u, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize_u, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_supervised = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
sum_loss_cluster_head = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u)
# reconstruction phase
qy_x_u, z_u = aae.encode_x_yz(images_u, apply_softmax=True)
representation = aae.encode_yz_representation(qy_x_u, z_u)
reconstruction_u = aae.decode_representation_x(representation)
loss_reconstruction = F.mean_squared_error(
aae.to_variable(images_u), reconstruction_u)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
z_true_u = sampler.gaussian(batchsize_u,
config.ndim_z,
mean=0,
var=1)
y_true_u = sampler.onehot_categorical(batchsize_u, config.ndim_y)
dz_true = aae.discriminate_z(z_true_u, apply_softmax=False)
dy_true = aae.discriminate_y(y_true_u, apply_softmax=False)
dz_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
dy_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_discriminator_z = F.softmax_cross_entropy(
dz_true, class_true) + F.softmax_cross_entropy(
dz_fake, class_fake)
loss_discriminator_y = F.softmax_cross_entropy(
dy_true, class_true) + F.softmax_cross_entropy(
dy_fake, class_fake)
loss_discriminator = loss_discriminator_z + loss_discriminator_y
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
dz_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
dy_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_generator_z = F.softmax_cross_entropy(dz_fake, class_true)
loss_generator_y = F.softmax_cross_entropy(dy_fake, class_true)
loss_generator = loss_generator_z + loss_generator_y
aae.backprop_generator(loss_generator)
# supervised phase
log_qy_x_l, z_l = aae.encode_x_yz(images_l, apply_softmax=False)
loss_supervised = F.softmax_cross_entropy(
log_qy_x_l, aae.to_variable(label_ids_l))
aae.backprop_generator(loss_supervised)
# additional cost function that penalizes the euclidean distance between of every two of cluster
distance = aae.compute_distance_of_cluster_heads()
loss_cluster_head = -F.sum(distance)
aae.backprop_cluster_head(loss_cluster_head)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_supervised += float(loss_supervised.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
sum_loss_cluster_head += float(
aae.nCr(config.ndim_y, 2) *
config.cluster_head_distance_threshold +
loss_cluster_head.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
# validation phase
images_v_segments = np.split(validation_images,
num_validation_data // 1000)
labels_v_segments = np.split(validation_labels,
num_validation_data // 1000)
sum_accuracy = 0
for images_v, labels_v in zip(images_v_segments, labels_v_segments):
qy = aae.encode_x_yz(images_v, apply_softmax=True, test=True)[0]
accuracy = F.accuracy(qy, aae.to_variable(labels_v))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_v_segments)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_s": sum_loss_supervised / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
"loss_c": sum_loss_cluster_head / num_trains_per_epoch,
"accuracy": validation_accuracy
})
# write accuracy to csv
csv_results.append([epoch, validation_accuracy])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy"]
data.to_csv("{}/result.csv".format(args.model_dir))
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = aae.config
# settings
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
sum_loss_cluster_head = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_u = dataset.sample_unlabeled_data(images, batchsize)
# reconstruction phase
qy_x_u, z_u = aae.encode_x_yz(images_u, apply_softmax=True)
representation = aae.encode_yz_representation(qy_x_u, z_u)
reconstruction_u = aae.decode_representation_x(representation)
loss_reconstruction = F.mean_squared_error(aae.to_variable(images_u), reconstruction_u)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
z_true_u = sampler.gaussian(batchsize, config.ndim_z, mean=0, var=1)
y_true_u = sampler.onehot_categorical(batchsize, config.ndim_y)
discrimination_z_true = aae.discriminate_z(z_true_u, apply_softmax=False)
discrimination_y_true = aae.discriminate_y(y_true_u, apply_softmax=False)
discrimination_z_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_discriminator_z = F.softmax_cross_entropy(discrimination_z_true, class_true) + F.softmax_cross_entropy(discrimination_z_fake, class_fake)
loss_discriminator_y = F.softmax_cross_entropy(discrimination_y_true, class_true) + F.softmax_cross_entropy(discrimination_y_fake, class_fake)
loss_discriminator = loss_discriminator_z + loss_discriminator_y
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
discrimination_z_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_generator_z = F.softmax_cross_entropy(discrimination_z_fake, class_true)
loss_generator_y = F.softmax_cross_entropy(discrimination_y_fake, class_true)
loss_generator = loss_generator_z + loss_generator_y
aae.backprop_generator(loss_generator)
# additional cost function that penalizes the euclidean distance between of every two of cluster
distance = aae.compute_distance_of_cluster_heads()
loss_cluster_head = -F.sum(distance)
aae.backprop_cluster_head(loss_cluster_head)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
sum_loss_cluster_head += float(aae.nCr(config.ndim_y, 2) * config.cluster_head_distance_threshold + loss_cluster_head.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
progress.show(num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
"loss_c": sum_loss_cluster_head / num_trains_per_epoch,
})
def main(run_load_from_file=False):
# load MNIST images
images, labels = dataset.load_train_images()
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
# setting
max_epoch = 510
num_trains_per_epoch = 500
batch_size_l = 100
batch_size_u = 100
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images, labels, num_validation_data, 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
training_loss_set = []
sum_loss_reconstruction = 0
sum_loss_supervised = 0
sum_loss_discrminator = 0
sum_loss_generator = 0
sum_loss_cluster_head = 0
process_iteration = Process()
for i in range(num_trains_per_epoch):
process_iteration.start_epoch(i, num_trains_per_epoch)
# sample from data distribution
images_l, label_onehot_l, label_id_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batch_size_l)
images_u = dataset.sample_unlabeled_data(
training_images_u, batch_size_u)
# additional cost function that penalizes the euclidean between of every two of cluster
if epoch == 0:
for j in range(5):
starting_labels, ending_labels = dataset.cluster_create_dataset(
config.ndim_y)
_, loss_transform = sess.run(
[h.opt_t, h.loss_t],
feed_dict={
h.g_s: starting_labels,
h.g_e: ending_labels,
h.lr: learning_rate
})
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
z_true_u = sampler.gaussian(batch_size_u,
config.ndim_z,
mean=0,
var=1)
y_true_u = sampler.onehot_categorical(batch_size_u,
config.ndim_y)
# adversarial phase for discriminator
_, loss_discriminator_y = sess.run([h.opt_dy, h.loss_dy],
feed_dict={
h.x: images_u,
h.y: y_true_u,
h.lr: learning_rate
})
_, 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 = loss_discriminator_y + loss_discriminator_z
# adversarial phase for generator
_, loss_generator_y, loss_generator_z = sess.run(
[h.opt_e, h.loss_gy, h.loss_gz],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
loss_generator = loss_generator_y + loss_generator_z
# supervised phase
_, loss_generator_supervised = sess.run([h.opt_ey, h.loss_ey],
feed_dict={
h.x: images_l,
h.y_s: label_id_l,
h.lr: learning_rate
})
training_loss_set.append([
loss_reconstruction, loss_discriminator,
loss_discriminator_y, loss_discriminator_z, loss_generator,
loss_generator_z, loss_generator_y,
loss_generator_supervised, loss_transform
])
sum_loss_reconstruction += loss_reconstruction
sum_loss_discrminator += loss_discriminator
sum_loss_generator += loss_generator
sum_loss_supervised += loss_generator_supervised
sum_loss_cluster_head += loss_transform
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),
'supervise': sum_loss_supervised / (i + 1),
'cluster_head': sum_loss_cluster_head / (i + 1)
})
# In end of epoch, summary the loss
average_training_loss_per_epoch = np.mean(
np.array(training_loss_set), axis=0)
# validation phase
images_v_segments = np.split(validation_images,
num_validation_data // 1000)
labels_v_segments = np.split(validation_labels,
num_validation_data // 1000)
sum_accuracy = 0
for images_v, labels_v in zip(images_v_segments,
labels_v_segments):
y_v = sess.run(h.y_r, feed_dict={h.x: images_v})
accuracy = opt.compute_accuracy(y_v, labels_v)
sum_accuracy += accuracy
validation_accuracy = sum_accuracy / len(images_v_segments)
# append validation accuracy to the training loss
average_loss_per_epoch = np.append(average_training_loss_per_epoch,
validation_accuracy)
training_epoch_loss.append(average_loss_per_epoch)
loss_name_per_epoch = [
'reconstruction', 'discriminator', 'discriminator_y',
'discriminator_z', 'generator', 'generator_z', 'generator_y',
'supervised', 'transform', 'validation_accuracy'
]
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[4],
'loss_v': average_loss_per_epoch[9],
})
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)))
if epoch % 10 == 0:
saver.save(sess,
os.path.join(config.ckpt_dir, 'model_ckptpoint'),
global_step=epoch)
pickle.dump((training_epoch_loss, loss_name_per_epoch),
open(config.ckpt_dir + '/pickle.pkl', 'wb'))
plt.plot_double_scale_trend(config.ckpt_dir)
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = adgm.config
# settings
max_epoch = 1000
num_trains_per_epoch = 500
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(images, labels, num_validation_data, num_labeled_data, num_types_of_label, seed=args.seed)
print training_labels_l
# init weightnorm layers
if config.use_weightnorm:
print "initializing weight normalization layers ..."
images_l, label_onehot_l, label_id_l = dataset.sample_labeled_data(training_images_l, training_labels_l, batchsize_l, config.ndim_x, config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u, batchsize_u, config.ndim_x)
adgm.compute_lower_bound(images_l, label_onehot_l, images_u)
# training
temperature = 1
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_lower_bound_l = 0
sum_lower_bound_u = 0
sum_loss_classifier = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(training_images_l, training_labels_l, batchsize_l, config.ndim_x, config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u, batchsize_u, config.ndim_x)
# lower bound loss using gumbel-softmax
lower_bound, lb_labeled, lb_unlabeled = adgm.compute_lower_bound_gumbel(images_l, label_onehot_l, images_u, temperature)
loss_lower_bound = -lower_bound
# classification loss
a_l = adgm.encode_x_a(images_l, False)
unnormalized_y_distribution = adgm.encode_ax_y_distribution(a_l, images_l, softmax=False)
loss_classifier = alpha * F.softmax_cross_entropy(unnormalized_y_distribution, adgm.to_variable(label_ids_l))
# backprop
adgm.backprop(loss_classifier + loss_lower_bound)
sum_lower_bound_l += float(lb_labeled.data)
sum_lower_bound_u += float(lb_unlabeled.data)
sum_loss_classifier += float(loss_classifier.data)
progress.show(t, num_trains_per_epoch, {})
adgm.save(args.model_dir)
# validation
images_l, _, label_ids_l = dataset.sample_labeled_data(validation_images, validation_labels, num_validation_data, config.ndim_x, config.ndim_y)
images_l_segments = np.split(images_l, num_validation_data // 500)
label_ids_l_segments = np.split(label_ids_l, num_validation_data // 500)
sum_accuracy = 0
for images_l, label_ids_l in zip(images_l_segments, label_ids_l_segments):
y_distribution = adgm.encode_x_y_distribution(images_l, softmax=True, test=True)
accuracy = F.accuracy(y_distribution, adgm.to_variable(label_ids_l))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_l_segments)
progress.show(num_trains_per_epoch, num_trains_per_epoch, {
"lb_u": sum_lower_bound_u / num_trains_per_epoch,
"lb_l": sum_lower_bound_l / num_trains_per_epoch,
"loss_spv": sum_loss_classifier / num_trains_per_epoch,
"accuracy": validation_accuracy,
"tmp": temperature,
})
# anneal the temperature
temperature = max(0.5, temperature * 0.999)
# write accuracy to csv
csv_results.append([epoch, validation_accuracy])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy"]
data.to_csv("{}/result.csv".format(args.model_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)
