def main(_, run_eval_loop=True):
with tf.name_scope('inputs'):
noise, one_hot_labels = _get_generator_inputs(
FLAGS.num_images_per_class, NUM_CLASSES, FLAGS.noise_dims)
# Generate images.
with tf.variable_scope('Generator'): # Same scope as in train job.
images = networks.conditional_generator((noise, one_hot_labels))
# Visualize images.
reshaped_img = tfgan.eval.image_reshaper(
images, num_cols=FLAGS.num_images_per_class)
tf.summary.image('generated_images', reshaped_img, max_outputs=1)
# Calculate evaluation metrics.
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(images, FLAGS.classifier_filename))
tf.summary.scalar('MNIST_Cross_entropy',
util.mnist_cross_entropy(
images, one_hot_labels, FLAGS.classifier_filename))
# Write images to disk.
image_write_ops = tf.write_file(
'%s/%s'% (FLAGS.eval_dir, 'conditional_gan.png'),
tf.image.encode_png(data_provider.float_image_to_uint8(reshaped_img[0])))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
with tf.name_scope('inputs'):
noise, one_hot_labels = _get_generator_inputs(
FLAGS.num_images_per_class, NUM_CLASSES, FLAGS.noise_dims)
# Generate images.
with tf.variable_scope('Generator'): # Same scope as in train job.
images = networks.conditional_generator((noise, one_hot_labels))
# Visualize images.
reshaped_img = tfgan.eval.image_reshaper(
images, num_cols=FLAGS.num_images_per_class)
tf.summary.image('generated_images', reshaped_img, max_outputs=1)
# Calculate evaluation metrics.
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(images, FLAGS.classifier_filename))
tf.summary.scalar('MNIST_Cross_entropy',
util.mnist_cross_entropy(
images, one_hot_labels, FLAGS.classifier_filename))
# Write images to disk.
image_write_ops = tf.write_file(
'%s/%s'% (FLAGS.eval_dir, 'conditional_gan.png'),
tf.image.encode_png(data_provider.float_image_to_uint8(reshaped_img[0])))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
# Fetch real images.
with tf.name_scope('inputs'):
real_images, _, _ = data_provider.provide_data(
'train', FLAGS.num_images_generated, FLAGS.dataset_dir)
image_write_ops = None
if FLAGS.eval_real_images:
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(real_images, FLAGS.classifier_filename))
else:
# In order for variables to load, use the same variable scope as in the
# train job.
with tf.variable_scope('Generator'):
images = networks.unconditional_generator(
tf.random_normal([FLAGS.num_images_generated, FLAGS.noise_dims]))
sess = tf.Session()
saver = tf.train.Saver()
if not restore_from_checkpoint(sess, saver):
raise NotImplementedError
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
sess.run(images, options=options, run_metadata=run_metadata)
cg = CompGraph('gan_mnist', run_metadata, tf.get_default_graph())
cg_tensor_dict = cg.get_tensors()
cg_sorted_keys = sorted(cg_tensor_dict.keys())
cg_sorted_items = []
for cg_key in cg_sorted_keys:
cg_sorted_items.append(tf.shape(cg_tensor_dict[cg_key]))
cg_sorted_shape = sess.run(cg_sorted_items)
cg.op_analysis(dict(zip(cg_sorted_keys, cg_sorted_shape)),
'gan_mnist.pickle')
exit(0)
#tf.summary.scalar('MNIST_Frechet_distance',
# util.mnist_frechet_distance(
# real_images, images, FLAGS.classifier_filename))
#tf.summary.scalar('MNIST_Classifier_score',
# util.mnist_score(images, FLAGS.classifier_filename))
if FLAGS.num_images_generated >= 100:
reshaped_images = tfgan.eval.image_reshaper(
images[:100, ...], num_cols=10)
uint8_images = data_provider.float_image_to_uint8(reshaped_images)
image_write_ops = tf.write_file(
'%s/%s'% (FLAGS.eval_dir, 'unconditional_gan.png'),
tf.image.encode_png(uint8_images[0]))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
with tf.name_scope('inputs'):
noise_args = (FLAGS.noise_samples, CAT_SAMPLE_POINTS, CONT_SAMPLE_POINTS,
FLAGS.unstructured_noise_dims, FLAGS.continuous_noise_dims)
# Use fixed noise vectors to illustrate the effect of each dimension.
display_noise1 = util.get_eval_noise_categorical(*noise_args)
display_noise2 = util.get_eval_noise_continuous_dim1(*noise_args)
display_noise3 = util.get_eval_noise_continuous_dim2(*noise_args)
_validate_noises([display_noise1, display_noise2, display_noise3])
# Visualize the effect of each structured noise dimension on the generated
# image.
def generator_fn(inputs):
return networks.infogan_generator(
inputs, len(CAT_SAMPLE_POINTS), is_training=False)
with tf.variable_scope('Generator') as genscope: # Same scope as in training.
categorical_images = generator_fn(display_noise1)
reshaped_categorical_img = tfgan.eval.image_reshaper(
categorical_images, num_cols=len(CAT_SAMPLE_POINTS))
tf.summary.image('categorical', reshaped_categorical_img, max_outputs=1)
with tf.variable_scope(genscope, reuse=True):
continuous1_images = generator_fn(display_noise2)
reshaped_continuous1_img = tfgan.eval.image_reshaper(
continuous1_images, num_cols=len(CONT_SAMPLE_POINTS))
tf.summary.image('continuous1', reshaped_continuous1_img, max_outputs=1)
with tf.variable_scope(genscope, reuse=True):
continuous2_images = generator_fn(display_noise3)
reshaped_continuous2_img = tfgan.eval.image_reshaper(
continuous2_images, num_cols=len(CONT_SAMPLE_POINTS))
tf.summary.image('continuous2', reshaped_continuous2_img, max_outputs=1)
# Evaluate image quality.
all_images = tf.concat(
[categorical_images, continuous1_images, continuous2_images], 0)
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(all_images, FLAGS.classifier_filename))
# Write images to disk.
image_write_ops = []
if FLAGS.write_to_disk:
image_write_ops.append(_get_write_image_ops(
FLAGS.eval_dir, 'categorical_infogan.png', reshaped_categorical_img[0]))
image_write_ops.append(_get_write_image_ops(
FLAGS.eval_dir, 'continuous1_infogan.png', reshaped_continuous1_img[0]))
image_write_ops.append(_get_write_image_ops(
FLAGS.eval_dir, 'continuous2_infogan.png', reshaped_continuous2_img[0]))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
# Fetch real images.
with tf.name_scope('inputs'):
real_images, _, _ = data_provider.provide_data(
'train', FLAGS.num_images_generated, FLAGS.dataset_dir)
image_write_ops = None
if FLAGS.eval_real_images:
tf.summary.scalar(
'MNIST_Classifier_score',
util.mnist_score(real_images, FLAGS.classifier_filename))
else:
# In order for variables to load, use the same variable scope as in the
# train job.
with tf.variable_scope('Generator'):
images = networks.unconditional_generator(tf.random_normal(
[FLAGS.num_images_generated, FLAGS.noise_dims]),
is_training=False)
tf.summary.scalar(
'MNIST_Frechet_distance',
util.mnist_frechet_distance(real_images, images,
FLAGS.classifier_filename))
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(images, FLAGS.classifier_filename))
if FLAGS.num_images_generated >= 100 and FLAGS.write_to_disk:
reshaped_images = tfgan.eval.image_reshaper(images[:100, ...],
num_cols=10)
uint8_images = data_provider.float_image_to_uint8(reshaped_images)
image_write_ops = tf.write_file(
'%s/%s' % (FLAGS.eval_dir, 'unconditional_gan.png'),
tf.image.encode_png(uint8_images[0]))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[
tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)
],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
# Fetch real images.
with tf.name_scope('inputs'):
real_images, _, _ = data_provider.provide_data(
'train', FLAGS.num_images_generated, FLAGS.dataset_dir)
image_write_ops = None
if FLAGS.eval_real_images:
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(real_images, FLAGS.classifier_filename))
else:
# In order for variables to load, use the same variable scope as in the
# train job.
with tf.variable_scope('Generator'):
images = networks.unconditional_generator(
tf.random_normal([FLAGS.num_images_generated, FLAGS.noise_dims]),
is_training=False)
tf.summary.scalar('MNIST_Frechet_distance',
util.mnist_frechet_distance(
real_images, images, FLAGS.classifier_filename))
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(images, FLAGS.classifier_filename))
if FLAGS.num_images_generated >= 100 and FLAGS.write_to_disk:
reshaped_images = tfgan.eval.image_reshaper(
images[:100, ...], num_cols=10)
uint8_images = data_provider.float_image_to_uint8(reshaped_images)
image_write_ops = tf.write_file(
'%s/%s'% (FLAGS.eval_dir, 'unconditional_gan.png'),
tf.image.encode_png(uint8_images[0]))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
def main(_, run_eval_loop=True):
with tf.name_scope('inputs'):
noise_args = (FLAGS.noise_samples, CAT_SAMPLE_POINTS,
CONT_SAMPLE_POINTS, FLAGS.unstructured_noise_dims,
FLAGS.continuous_noise_dims)
# Use fixed noise vectors to illustrate the effect of each dimension.
display_noise1 = util.get_eval_noise_categorical(*noise_args)
display_noise2 = util.get_eval_noise_continuous_dim1(*noise_args)
display_noise3 = util.get_eval_noise_continuous_dim2(*noise_args)
_validate_noises([display_noise1, display_noise2, display_noise3])
# Visualize the effect of each structured noise dimension on the generated
# image.
generator_fn = lambda x: networks.infogan_generator(
x, len(CAT_SAMPLE_POINTS))
with tf.variable_scope(
'Generator') as genscope: # Same scope as in training.
categorical_images = generator_fn(display_noise1)
reshaped_categorical_img = tfgan.eval.image_reshaper(
categorical_images, num_cols=len(CAT_SAMPLE_POINTS))
tf.summary.image('categorical', reshaped_categorical_img, max_outputs=1)
with tf.variable_scope(genscope, reuse=True):
continuous1_images = generator_fn(display_noise2)
reshaped_continuous1_img = tfgan.eval.image_reshaper(
continuous1_images, num_cols=len(CONT_SAMPLE_POINTS))
tf.summary.image('continuous1', reshaped_continuous1_img, max_outputs=1)
with tf.variable_scope(genscope, reuse=True):
continuous2_images = generator_fn(display_noise3)
reshaped_continuous2_img = tfgan.eval.image_reshaper(
continuous2_images, num_cols=len(CONT_SAMPLE_POINTS))
tf.summary.image('continuous2', reshaped_continuous2_img, max_outputs=1)
# Evaluate image quality.
all_images = tf.concat(
[categorical_images, continuous1_images, continuous2_images], 0)
tf.summary.scalar('MNIST_Classifier_score',
util.mnist_score(all_images, FLAGS.classifier_filename))
# Write images to disk.
image_write_ops = []
image_write_ops.append(
_get_write_image_ops(FLAGS.eval_dir, 'categorical_infogan.png',
reshaped_categorical_img[0]))
image_write_ops.append(
_get_write_image_ops(FLAGS.eval_dir, 'continuous1_infogan.png',
reshaped_continuous1_img[0]))
image_write_ops.append(
_get_write_image_ops(FLAGS.eval_dir, 'continuous2_infogan.png',
reshaped_continuous2_img[0]))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop: return
tf.contrib.training.evaluate_repeatedly(
FLAGS.checkpoint_dir,
hooks=[
tf.contrib.training.SummaryAtEndHook(FLAGS.eval_dir),
tf.contrib.training.StopAfterNEvalsHook(1)
],
eval_ops=image_write_ops,
max_number_of_evaluations=FLAGS.max_number_of_evaluations)
discriminator_optimizer = tf.train.AdamOptimizer(0.0001, beta1=0.5)
gan_train_ops = tf.contrib.gan.gan_train_ops(
gan_model,
improved_wgan_loss,
generator_optimizer,
discriminator_optimizer)
num_images_to_eval = 500
MNIST_CLASSIFIER_FROZEN_GRAPH = str(Path('mnist_data/frozen.pb').resolve())
# 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]))
eval_score = util.mnist_score(eval_images, MNIST_CLASSIFIER_FROZEN_GRAPH)
# We have the option to train the discriminator more than one step for every
# step of the generator. In order to do this, we use a `GANTrainSteps` with
# desired values. For this example, we use the default 1 generator train step
# for every discriminator train step.
train_step_fn = tf.contrib.gan.get_sequential_train_steps()
global_step = tf.train.get_or_create_global_step()
loss_values, mnist_score_values = [], []
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
with tf.contrib.slim.queues.QueueRunners(sess):
start_time = time.time()
for i in range(801):