def train_input_fn():
with tf.device('/cpu:0'):
images, _, _ = data_provider.provide_data('train',
batch_size,
dataset_dir,
num_threads=num_threads)
noise = tf.random_normal([batch_size, noise_dims])
return noise, images
def image2image():
# Set up the input pipeline.
input_image, target_image = data_provider.provide_data(FLAGS.batch_size)
# Build the generator and discriminator.
gan_model = tfgan.gan_model(
generator_fn=nets.generator, # you define
discriminator_fn=nets.discriminator, # you define
real_data=target_image,
generator_inputs=input_image)
# Build the GAN loss and standard pixel loss.
gan_loss = tfgan.gan_loss(
gan_model,
generator_loss_fn=tfgan.losses.least_squares_generator_loss,
discriminator_loss_fn=tfgan.losses.least_squares_discriminator_loss)
l1_pixel_loss = tf.norm(gan_model.real_data - gan_model.generated_data,
ord=1)
# Modify the loss tuple to include the pixel loss.
gan_loss = tfgan.losses.combine_adversarial_loss(
gan_loss, gan_model, l1_pixel_loss, weight_factor=FLAGS.weight_factor)
def prepare_data():
MNIST_DATA_DIR = '/tmp/mnist-data'
if not tf.gfile.Exists(MNIST_DATA_DIR):
tf.gfile.MakeDirs(MNIST_DATA_DIR)
download_and_convert_mnist.run(MNIST_DATA_DIR)
# FIXME why reset graph here?
tf.reset_default_graph()
# Define our input pipeline. Pin it to the CPU so that the GPU can be reserved
# for forward and backwards propogation.
batch_size = 32
with tf.device('/cpu:0'):
# data provider FIXME why returning three fields?
real_images, _, _ = data_provider.provide_data('train', batch_size,
MNIST_DATA_DIR)
# Sanity check that we're getting images.
check_real_digits = tfgan.eval.image_reshaper(real_images[:20, ...],
num_cols=10)
visualize_digits(check_real_digits)
# FIXME is this a tensor or an iterator?
return real_images
def train_input_fn():
with tf.device('/cpu:0'):
real_images, _, _ = data_provider.provide_data(
'train', batch_size, MNIST_DATA_DIR)
noise = tf.random_normal([batch_size, noise_dims])
return noise, real_images
# In[5]:
MNIST_DATA_DIR = '/tmp/mnist-data'
if not tf.gfile.Exists(MNIST_DATA_DIR):
tf.gfile.MakeDirs(MNIST_DATA_DIR)
download_and_convert_mnist.run(MNIST_DATA_DIR)
tf.reset_default_graph()
# Define our input pipeline. Pin it to the CPU so that the GPU can be reserved
# for forward and backwards propogation.
batch_size = 32
with tf.device('/cpu:0'):
real_images, _, _ = data_provider.provide_data('train', batch_size,
MNIST_DATA_DIR)
# Sanity check that we're getting images.
check_real_digits = tfgan.eval.image_reshaper(real_images[:20, ...],
num_cols=10)
visualize_digits(check_real_digits)
def generator_fn(noise, weight_decay=2.5e-5, is_training=True):
"""Simple generator to produce MNIST images.
Args:
noise: A single Tensor representing noise.
weight_decay: The value of the l2 weight decay.
is_training: If `True`, batch norm uses batch statistics. If `False`, batch
norm uses the exponential moving average collected from population
def train_input_fn():
with tf.device('/cpu:0'):
images, _, _ = data_provider.provide_data(
'train', batch_size, dataset_dir, num_threads=num_threads)
noise = tf.random_normal([batch_size, noise_dims])
return noise, images
MNIST_DATA_DIR = '/tmp/mnist-data'
if not tf.gfile.Exists(MNIST_DATA_DIR):
tf.gfile.MakeDirs(MNIST_DATA_DIR)
download_and_convert_mnist.run(MNIST_DATA_DIR)
tf.reset_default_graph()
# Define our input pipeline. Pin it to the CPU so that the GPU can be reserved
# for forward and backwards propogation.
batch_size = 32
with tf.device('/cpu:0'):
real_images, _, _ = data_provider.provide_data(
'train', batch_size, MNIST_DATA_DIR)
# Sanity check that we're getting images.
check_real_digits = tfgan.eval.image_reshaper(real_images[:20,...], num_cols=10)
visualize_digits(check_real_digits)
def infogan_generator(inputs, categorical_dim, weight_decay=2.5e-5,
is_training=True):
"""InfoGAN discriminator network on MNIST digits.
Based on a paper https://arxiv.org/abs/1606.03657 and their code
https://github.com/openai/InfoGAN.
Args:
inputs: A 3-tuple of Tensors (unstructured_noise, categorical structured
DATA_DIR = "assets/data/"
RESULT_DIR = "assets/results/"
# short cuts
tfgan = tf.contrib.gan
slim = tf.contrib.slim
if not tf.gfile.Exists(DATA_DIR): # check if data directory already exists
tf.gfile.MakeDirs(DATA_DIR)
download_and_convert_mnist.run(DATA_DIR) # download data if missing
# input pipeline
batch_size = 32
with tf.device('/cpu:0'): # pin it to the cpu to free up gpu for propagation
images, one_hot_labels, _ = data_provider.provide_data(
'train', batch_size, DATA_DIR)
# Sanity check that we're getting images.
imgs_to_visualize = tfgan.eval.image_reshaper(images[:20, ...], num_cols=10)
visualizer.pre_train_image(imgs_to_visualize, save=False)
noise_dims = 64
conditional_gan_model = tfgan.gan_model(generator_fn=tf_nets.generator,
discriminator_fn=tf_nets.discriminator,
real_data=images,
generator_inputs=(tf.random_normal(
[batch_size,
noise_dims]), one_hot_labels))
# Sanity check that currently generated images are garbage.
cond_generated_data_to_visualize = tfgan.eval.image_reshaper(
def train(is_train):
if not tf.gfile.Exists(MNIST_DATA_DIR):
tf.gfile.MakeDirs(MNIST_DATA_DIR)
#download_and_convert_mnist.run(MNIST_DATA_DIR)
tf.reset_default_graph()
# Define our input pipeline. Pin it to the CPU so that the GPU can be reserved
# for forward and backwards propogation.
batch_size = 32
with tf.device('/cpu:0'):
real_images, _, _ = data_provider.provide_data(
'train', batch_size, MNIST_DATA_DIR)
# Sanity check that we're getting images.
#check_real_digits = tfgan.eval.image_reshaper(
# real_images[:20,...], num_cols=10)
#print('visualize_digits')
#visualize_digits(check_real_digits)
#plt.show()
gan_model = tfgan.gan_model(
generator_fn,
discriminator_fn,
real_data=real_images,
generator_inputs=tf.random_normal([batch_size, noise_dims]))
improved_wgan_loss = tfgan.gan_loss(
gan_model,
# We make the loss explicit for demonstration, even though the default is
# Wasserstein loss.
generator_loss_fn=tfgan.losses.wasserstein_generator_loss,
discriminator_loss_fn=tfgan.losses.wasserstein_discriminator_loss,
gradient_penalty_weight=1.0)
# Sanity check that we can evaluate our losses.
print("Sanity check that we can evaluate our losses")
for gan_loss, name in [(improved_wgan_loss, 'improved wgan loss')]:
evaluate_tfgan_loss(gan_loss, name)
#generator_optimizer = tf.train.AdamOptimizer(0.001, beta1=0.5)
#discriminator_optimizer = tf.train.AdamOptimizer(0.0001, beta1=0.5)
generator_optimizer = tf.train.RMSPropOptimizer(0.001)
discriminator_optimizer = tf.train.RMSPropOptimizer(0.0001)
gan_train_ops = tfgan.gan_train_ops(
gan_model,
improved_wgan_loss,
generator_optimizer,
discriminator_optimizer)
# ### Evaluation
num_images_to_eval = 500
MNIST_CLASSIFIER_FROZEN_GRAPH = os.path.join(
RESEARCH_FOLDER,
'gan/mnist/data/classify_mnist_graph_def.pb')
# For variables to load, use the same variable scope as in the train job.
with tf.variable_scope('Generator', reuse=True):
eval_images = gan_model.generator_fn(
tf.random_normal([num_images_to_eval, noise_dims]),
is_training=False)
# Calculate Inception score.
eval_score = util.mnist_score(eval_images, MNIST_CLASSIFIER_FROZEN_GRAPH)
# Calculate Frechet Inception distance.
with tf.device('/cpu:0'):
real_images, _, _ = data_provider.provide_data(
'train', num_images_to_eval, MNIST_DATA_DIR)
frechet_distance = util.mnist_frechet_distance(
real_images, eval_images, MNIST_CLASSIFIER_FROZEN_GRAPH)
# Reshape eval images for viewing.
generated_data_to_visualize = tfgan.eval.image_reshaper(
eval_images[:20,...], num_cols=10)
# This code block should take about **1 minute** to run on a GPU kernel, and about **8 minutes** on CPU.
train_step_fn = tfgan.get_sequential_train_steps()
global_step = tf.train.get_or_create_global_step()
loss_values, mnist_scores, frechet_distances = [], [], []
tf.summary.scalar('dis_loss', gan_loss.discriminator_loss)
tf.summary.scalar('gen_loss', gan_loss.generator_loss)
merged = tf.summary.merge_all()
saver = tf.train.Saver()
saver_hook = tf.train.CheckpointSaverHook(
checkpoint_dir= "./models",
save_steps=1000,
saver=saver)
print("Graph trainable nodes:")
for v in tf.trainable_variables():
print (v.name)
with tf.train.SingularMonitoredSession(hooks=[saver_hook],
checkpoint_dir="./models") as sess:
start_time = time.time()
train_writer = tf.summary.FileWriter("./summary", sess.graph)
if is_train:
for i in xrange(2000):
cur_loss, _ = train_step_fn(
sess, gan_train_ops, global_step, train_step_kwargs={})
loss_values.append((i, cur_loss))
if i % 10 == 0:
merged_val = sess.run(merged)
train_writer.add_summary(merged_val, i)
print("Step:{}".format(i))
mnist_score, f_distance, digits_np = sess.run(
[eval_score, frechet_distance, generated_data_to_visualize])
mnist_scores.append((i, mnist_score))
frechet_distances.append((i, f_distance))
print('Current loss: %f' % cur_loss)
print('Current MNIST score: %f' % mnist_scores[-1][1])
print('Current Frechet distance: %f' % frechet_distances[-1][1])
visualize_training_generator(i, start_time, digits_np)
else: #generate from trained model
generated = sess.run(eval_images)
print("generated[0] shape:{}".format(generated[0].shape))
plt.imshow(np.squeeze(generated[0]), cmap='gray')
plt.show()
