def train():
with tf.Graph().as_default():
# data
dataset = traindataset.DataSet(DATA_DIR, SAMPLE_SIZE)
# tfrecords inputs
images, labels_t = dataset.csv_inputs(CSVFILE)
z = tf.placeholder(tf.float32, [None, Z_DIM], name='z')
dcgan = DCGAN("test", "./checkpoint")
images_inf, logits, logits_, G_sum, z_sum, d_sum, d__sum = dcgan.inference(images, z)
d_loss_fake, d_loss_real, d_loss_real_sum, d_loss_fake_sum, d_loss_sum, g_loss_sum, d_loss, g_loss = dcgan.loss(logits, logits_)
# trainable variables
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'd_' in var.name]
g_vars = [var for var in t_vars if 'g_' in var.name]
# train operations
d_optim = D_train_op(d_loss, d_vars)
g_optim = G_train_op(g_loss, g_vars)
# generate images
generate_images = dcgan.generate_images(z, 4, 4)
# summary
g_sum = tf.merge_summary([z_sum, d__sum, G_sum, d_loss_fake_sum, g_loss_sum])
d_sum = tf.merge_summary([z_sum, d_sum, d_loss_real_sum, d_loss_sum])
#summary_op = tf.merge_all_summaries()
# init operation
init_op = tf.initialize_all_variables()
# Session
sess = tf.Session(config=tf.ConfigProto(log_device_placement=LOG_DEVICE_PLACEMENT))
writer = tf.train.SummaryWriter("./logs", sess.graph_def)
# saver
saver = tf.train.Saver(tf.all_variables())
# run
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# sampling
sample_z = np.random.uniform(-1, 1, size=(SAMPLE_SIZE, Z_DIM))
# sample images
#sample_images = dataset.get_sample()
counter = 1
start_time = time.time()
for epoch in xrange(EPOCHS):
for idx in xrange(0, dataset.batch_idxs):
#batch_images = dataset.create_batch()
batch_z = np.random.uniform(-1, 1, [BATCH_SIZE, Z_DIM]).astype(np.float32)
# discriminative
images_inf_eval, _, summary_str = sess.run([images_inf, d_optim, d_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
#for i, image_inf in enumerate(images_inf_eval):
# print np.uint8(image_inf)
# print image_inf.shape
# #image_inf_reshape = image_inf.reshape([64, 64, 3])
# img = Image.fromarray(np.asarray(image_inf), 'RGB')
# print img
# img.save('verbose/%d_%d.png' % (counter, i))
# generative
_, summary_str = sess.run([g_optim, g_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
# twice optimization
_, summary_str = sess.run([g_optim, g_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
errD_fake = sess.run(d_loss_fake, {z: batch_z})
errD_real = sess.run(d_loss_real, {z: batch_z})
errG = sess.run(g_loss, {z: batch_z})
print("epochs: %02d %04d/%04d time: %4.4f, d_loss: %.8f, g_loss: %.8f" % (epoch, idx, dataset.batch_idxs,time.time() - start_time, errD_fake + errD_real, errG))
if np.mod(counter, 10) == 1:
print("generate samples.")
generated_image_eval = sess.run(generate_images, {z: batch_z})
filename = os.path.join(FLAGS.sample_dir, 'out_%05d.png' % counter)
with open(filename, 'wb') as f:
f.write(generated_image_eval)
counter += 1
coord.request_stop()
coord.join(threads)
sess.close()
def main():
# load data
bedroom = LSUNdataset(dirn='../data', category='bedroom')
# fake_x = np.ones([128, 28, 28, 1], dtype=np.float32) * 0.1
# condition
k = 1 # # of discrim updates for each gen update
l2 = 2.5e-5 # l2 weight decay
b1 = 0.5 # momentum term of adam
nc = 3 # # of channels in image
ny = 10 # # of classes
batch_size = 128 # # of examples in batch
npx = 32 # # of pixels width/height of images
nz = 100 # # of dim for Z
ngfc = 1024 # # of gen units for fully connected layers
ndfc = 1024 # # of discrim units for fully connected layers
ngf = 64 # # of gen filters in first conv layer
ndf = 64 # # of discrim filters in first conv layer
nx = npx*npx*nc # # of dimensions in X
niter = 100 # # of iter at starting learning rate
niter_decay = 100 # # of iter to linearly decay learning rate to zero
lr = 0.0002 # initial learning rate for adam
# tensorflow placeholder
x = tf.placeholder(tf.float32, [None, npx, npx, nc])
# y = tf.placeholder(tf.float32, [None, ny]) # for training w/ label
y = None # without label data
# y_target = tf.placeholder(tf.float32, [None, ny]) # for image generation
# graphs
dcgan = DCGAN(batch_size=batch_size, s_size=4, z_dim=nz, y_dim=None)
logits_list = dcgan.inference(x, y) # (x, y)
g_loss, d_loss = dcgan.loss(logits_list)
train_op = dcgan.train(g_loss, d_loss, learning_rate=lr)
# images
images = dcgan.sample_images(label=None)
# images = dcgan.sample_images()
init = tf.global_variables_initializer()
# Training
n_epochs = 10
with tf.Session() as sess:
sess.run(init)
# loop control
n_sample = bedroom.num_examples
n_loop = n_sample // batch_size # for LSUN bedroom dataset, n_loop = 23696
if n_sample % batch_size != 0:
n_loop += 1
for e in range(1, n_epochs+1):
for i in range(n_loop):
batch_x = bedroom.next_batch(batch_size, img_size=32)
batch_img = batch_x.reshape([-1, 32, 32, 3])
fd_train = {x: batch_img}
# fd_train = {x: batch_img}
sess.run(train_op, feed_dict=fd_train)
g_loss_np, d_loss_np = sess.run([g_loss, d_loss],
feed_dict=fd_train)
# print status
if i % 10 == 0:
print((' ecpoch {:>5d}: ({:>8d} /{:>8d}) :'
'g_loss={:>10.4f}, d_loss={:>10.4f}').format(
e, i, n_loop, g_loss_np, d_loss_np))
if i == 100:
break
# Generate sample images after training
if e in [1, 2, 5, 10]:
fn_sample = '../work/samples/bedroom_' + str(e) + '.jpg'
generated = sess.run(images)
# generated = sess.run(images)
with open(fn_sample, 'wb') as fp:
fp.write(generated)
def train():
with tf.Graph().as_default():
# data
dataset = traindataset.DataSet(DATA_DIR, SAMPLE_SIZE)
# tfrecords inputs
images, labels_t = dataset.csv_inputs(CSVFILE)
z = tf.placeholder(tf.float32, [None, Z_DIM], name='z')
dcgan = DCGAN("test", "./checkpoint")
images_inf, logits, logits_, G_sum, z_sum, d_sum, d__sum = dcgan.inference(
images, z)
d_loss_fake, d_loss_real, d_loss_real_sum, d_loss_fake_sum, d_loss_sum, g_loss_sum, d_loss, g_loss = dcgan.loss(
logits, logits_)
# trainable variables
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'd_' in var.name]
g_vars = [var for var in t_vars if 'g_' in var.name]
# train operations
d_optim = D_train_op(d_loss, d_vars)
g_optim = G_train_op(g_loss, g_vars)
# generate images
generate_images = dcgan.generate_images(z, 4, 4)
# summary
g_sum = tf.merge_summary(
[z_sum, d__sum, G_sum, d_loss_fake_sum, g_loss_sum])
d_sum = tf.merge_summary([z_sum, d_sum, d_loss_real_sum, d_loss_sum])
#summary_op = tf.merge_all_summaries()
# init operation
init_op = tf.initialize_all_variables()
# Session
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=LOG_DEVICE_PLACEMENT))
writer = tf.train.SummaryWriter("./logs", sess.graph_def)
# saver
saver = tf.train.Saver(tf.all_variables())
# run
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# sampling
sample_z = np.random.uniform(-1, 1, size=(SAMPLE_SIZE, Z_DIM))
# sample images
#sample_images = dataset.get_sample()
counter = 1
start_time = time.time()
for epoch in xrange(EPOCHS):
for idx in xrange(0, dataset.batch_idxs):
#batch_images = dataset.create_batch()
batch_z = np.random.uniform(-1, 1, [BATCH_SIZE, Z_DIM]).astype(
np.float32)
# discriminative
images_inf_eval, _, summary_str = sess.run(
[images_inf, d_optim, d_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
#for i, image_inf in enumerate(images_inf_eval):
# print np.uint8(image_inf)
# print image_inf.shape
# #image_inf_reshape = image_inf.reshape([64, 64, 3])
# img = Image.fromarray(np.asarray(image_inf), 'RGB')
# print img
# img.save('verbose/%d_%d.png' % (counter, i))
# generative
_, summary_str = sess.run([g_optim, g_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
# twice optimization
_, summary_str = sess.run([g_optim, g_sum], {z: batch_z})
writer.add_summary(summary_str, counter)
errD_fake = sess.run(d_loss_fake, {z: batch_z})
errD_real = sess.run(d_loss_real, {z: batch_z})
errG = sess.run(g_loss, {z: batch_z})
print(
"epochs: %02d %04d/%04d time: %4.4f, d_loss: %.8f, g_loss: %.8f"
% (epoch, idx, dataset.batch_idxs,
time.time() - start_time, errD_fake + errD_real, errG))
if np.mod(counter, 10) == 1:
print("generate samples.")
generated_image_eval = sess.run(generate_images,
{z: batch_z})
filename = os.path.join(FLAGS.sample_dir,
'out_%05d.png' % counter)
with open(filename, 'wb') as f:
f.write(generated_image_eval)
counter += 1
coord.request_stop()
coord.join(threads)
sess.close()
def main():
# load data
cifar = load_data('../data/')
# fake_x = np.ones([128, 28, 28, 1], dtype=np.float32) * 0.1
# condition
k = 1 # # of discrim updates for each gen update
l2 = 2.5e-5 # l2 weight decay
b1 = 0.5 # momentum term of adam
nc = 3 # # of channels in image
ny = 10 # # of classes
batch_size = 128 # # of examples in batch
npx = 32 # # of pixels width/height of images
nz = 100 # # of dim for Z
ngfc = 1024 # # of gen units for fully connected layers
ndfc = 1024 # # of discrim units for fully connected layers
ngf = 64 # # of gen filters in first conv layer
ndf = 64 # # of discrim filters in first conv layer
nx = npx * npx * nc # # of dimensions in X
niter = 100 # # of iter at starting learning rate
niter_decay = 100 # # of iter to linearly decay learning rate to zero
lr = 0.0002 # initial learning rate for adam
# tensorflow placeholder
x = tf.placeholder(tf.float32, [None, npx, npx, nc])
y = tf.placeholder(tf.float32, [None, ny]) # for training w/ label
y_target = tf.placeholder(tf.float32, [None, ny]) # for image generation
# graphs
dcgan = DCGAN(batch_size=batch_size, s_size=4, z_dim=nz, y_dim=ny)
logits_list = dcgan.inference(x, y) # (x, y)
g_loss, d_loss = dcgan.loss(logits_list)
train_op = dcgan.train(g_loss, d_loss, learning_rate=lr)
# images
images = dcgan.sample_images(label=y_target)
# images = dcgan.sample_images()
init = tf.global_variables_initializer()
# Training
n_epochs = 300
with tf.Session() as sess:
sess.run(init)
# loop control
n_sample = cifar.train.num_examples
n_loop = n_sample // batch_size
if n_sample % batch_size != 0:
n_loop += 1
for e in range(1, n_epochs + 1):
for i in range(n_loop):
batch_x, batch_y = cifar.train.next_batch(batch_size)
batch_img = batch_x.reshape([-1, 32, 32, 3])
fd_train = {x: batch_img, y: batch_y}
# fd_train = {x: batch_img}
sess.run(train_op, feed_dict=fd_train)
g_loss_np, d_loss_np = sess.run([g_loss, d_loss],
feed_dict=fd_train)
print('ecpoch {:>5d}: g_loss={:>11.4f}, d_loss={:>11.4f}'.format(
e, g_loss_np, d_loss_np))
# Generate sample images after training
if e in [10, 20, 50, 100, 200, 300]:
_, batch_yv = cifar.validation.next_batch(batch_size)
fn_sample = '../work/samples/cifar_' + str(e) + '.jpg'
generated = sess.run(images, feed_dict={y_target: batch_yv})
# generated = sess.run(images)
with open(fn_sample, 'wb') as fp:
fp.write(generated)
