def srgan_model(features, labels, mode, params):
del params
global load_flag
if mode == tf.estimator.ModeKeys.PREDICT:
net_g_test = SRGAN_g(features, is_train=False)
predictions = {'generated_images': net_g_test.outputs}
return tf.estimator.EstimatorSpec(mode, predictions=predictions)
net_g = SRGAN_g(features, is_train=True)
net_d, logits_real = SRGAN_d(labels, is_train=True)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True)
t_target_image_224 = tf.image.resize_images(labels,
size=[224, 224],
method=0,
align_corners=False)
t_predict_image_224 = tf.image.resize_images(
net_g.outputs, size=[224, 224], method=0,
align_corners=False) # resize_generate_image_for_vgg
net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2)
_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2)
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs, labels, is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(config.TRAIN.lr_init, trainable=False)
# SRGAN
g_optim = tf.train.AdamOptimizer(lr_v, beta1=config.TRAIN.beta1) \
.minimize(g_loss, var_list=g_vars, global_step=tf.train.get_global_step())
d_optim = tf.train.AdamOptimizer(lr_v, beta1=config.TRAIN.beta1) \
.minimize(d_loss, var_list=d_vars, global_step=tf.train.get_global_step())
joint_op = tf.group([g_optim, d_optim])
load_vgg(net_vgg)
return tf.estimator.EstimatorSpec(mode, loss=g_loss, train_op=joint_op)
def g_init_model(features, labels, mode, params):
del params
if mode == tf.estimator.ModeKeys.PREDICT:
net_g_test = SRGAN_g(features, is_train=False)
predictions = {'generated_images': net_g_test.outputs}
return tf.contrib.tpu.TPUEstimatorSpec(mode, predictions=predictions)
net_g = SRGAN_g(features, is_train=True)
_ = SRGAN_d(labels, is_train=True)
mse_loss = tl.cost.mean_squared_error(net_g.outputs, labels, is_mean=True)
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(config.TRAIN.lr_init, trainable=False)
g_optimizer = tf.train.AdamOptimizer(lr_v, beta1=config.TRAIN.beta1)
g_optimizer = tf.contrib.tpu.CrossShardOptimizer(g_optimizer)
init_ops = g_optimizer.minimize(mse_loss,
var_list=g_vars,
global_step=tf.train.get_global_step())
return tf.contrib.tpu.TPUEstimatorSpec(mode,
loss=mse_loss,
train_op=init_ops)
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.hr_img_path,
regx='.*.png',
printable=False))
train_lr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.lr_img_path,
regx='.*.png',
printable=False))
valid_hr_img_list = sorted(
tl.files.load_file_list(path=config.VALID.hr_img_path,
regx='.*.png',
printable=False))
valid_lr_img_list = sorted(
tl.files.load_file_list(path=config.VALID.lr_img_path,
regx='.*.png',
printable=False))
## If your machine have enough memory, please pre-load the whole train set.
train_hr_imgs = tl.vis.read_images(train_hr_img_list,
path=config.TRAIN.hr_img_path,
n_threads=32)
# for im in train_hr_imgs:
# print(im.shape)
# valid_lr_imgs = tl.vis.read_images(valid_lr_img_list, path=config.VALID.lr_img_path, n_threads=32)
# for im in valid_lr_imgs:
# print(im.shape)
# valid_hr_imgs = tl.vis.read_images(valid_hr_img_list, path=config.VALID.hr_img_path, n_threads=32)
# for im in valid_hr_imgs:
# print(im.shape)
# exit()
###========================== DEFINE MODEL ============================###
## train inference
t_image = tf.placeholder('float32', [batch_size, 96, 96, 3],
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3],
name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(
t_target_image, size=[224, 224], method=0, align_corners=False
) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(
net_g.outputs, size=[224, 224], method=0,
align_corners=False) # resize_generate_image_for_vgg
net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2,
reuse=False)
_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2,
reuse=True)
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs,
t_target_image,
is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
tl.layers.initialize_global_variables(sess)
if tl.files.load_and_assign_npz(
sess=sess,
name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']),
network=net_g) is None:
tl.files.load_and_assign_npz(
sess=sess,
name=checkpoint_dir +
'/g_{}_init.npz'.format(tl.global_flag['mode']),
network=net_g)
tl.files.load_and_assign_npz(sess=sess,
name=checkpoint_dir +
'/d_{}.npz'.format(tl.global_flag['mode']),
network=net_d)
###============================= LOAD VGG ===============================###
vgg19_npy_path = "vgg19.npy"
if not os.path.isfile(vgg19_npy_path):
print(
"Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg"
)
exit()
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
# net_vgg.print_params(False)
# net_vgg.print_layers()
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
sample_imgs_384 = tl.prepro.threading_data(sample_imgs,
fn=crop_sub_imgs_fn,
is_random=False)
print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(),
sample_imgs_384.max())
sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384,
fn=downsample_fn)
print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(),
sample_imgs_96.max())
tl.vis.save_images(sample_imgs_96, [ni, ni],
save_dir_ginit + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni],
save_dir_ginit + '/_train_sample_384.png')
tl.vis.save_images(sample_imgs_96, [ni, ni],
save_dir_gan + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni],
save_dir_gan + '/_train_sample_384.png')
###========================= initialize G ====================###
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update G
errM, _ = sess.run([mse_loss, g_optim_init], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " %
(epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni],
save_dir_ginit + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(net_g.all_params,
name=checkpoint_dir +
'/g_{}_init.npz'.format(tl.global_flag['mode']),
sess=sess)
###========================= train GAN (SRGAN) =========================###
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update D
errD, _ = sess.run([d_loss, d_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
## update G
errG, errM, errV, errA, _ = sess.run(
[g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
print(
"Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)"
% (epoch, n_epoch, n_iter, time.time() - step_time, errD, errG,
errM, errV, errA))
total_d_loss += errD
total_g_loss += errG
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni],
save_dir_gan + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(net_g.all_params,
name=checkpoint_dir +
'/g_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
tl.files.save_npz(net_d.all_params,
name=checkpoint_dir +
'/d_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.hr_img_path,
regx='.*.png',
printable=False))
train_lr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.lr_img_path,
regx='.*.png',
printable=False))
## If your machine have enough memory, please pre-load the whole train set.
print("reading images")
train_hr_imgs = [] #[None] * len(train_hr_img_list)
train_lr_imgs = [] #[None] * len(train_hr_img_list)
#sess = tf.Session()
for img__ in train_lr_img_list:
image_loaded = scipy.misc.imread(
os.path.join(config.TRAIN.hr_img_path, img__))
image_loaded = image_loaded.reshape(
(image_loaded.shape[0], image_loaded.shape[1], 1))
print(image_loaded.shape)
image_loaded = image_loaded / (np.max(image_loaded) + 1)
train_hr_imgs.append(image_loaded)
aug1, aug2, aug3 = data_augment(image_loaded, is_mask=True)
train_hr_imgs.append(aug1)
train_hr_imgs.append(aug2)
train_hr_imgs.append(aug3)
for img__ in train_lr_img_list:
image_loaded = scipy.misc.imread(
os.path.join(config.TRAIN.lr_img_path, img__))
image_loaded = image_loaded.reshape(
(image_loaded.shape[0], image_loaded.shape[1], 1))
print(image_loaded.shape)
image_loaded = image_loaded / (np.max(image_loaded) + 1)
train_lr_imgs.append(image_loaded)
aug1, aug2, aug3 = data_augment(image_loaded, is_mask=False)
train_lr_imgs.append(aug1)
train_lr_imgs.append(aug2)
train_lr_imgs.append(aug3)
#shuffle lists
random.seed(2018)
shuffle(train_hr_imgs)
random.seed(2018)
shuffle(train_lr_imgs)
###========================== DEFINE MODEL ============================###
## train inference
t_image = tf.placeholder('float32', [batch_size, 128, 128, 1],
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder(
'float32', [batch_size, 128, 128, 1], name='t_target_image'
) # may have to convert 224x224x1 into 224x224x3, with channel 1 & 2 as 0. May have to have separate place-holder ?
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
#net_g = u_net(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(
t_target_image, size=[224, 224], method=0, align_corners=False
) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(
net_g.outputs, size=[224, 224], method=0,
align_corners=False) # resize_generate_image_for_vgg
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-4 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
w_t_target_image = t_target_image
weight = 3
weights = tf.multiply(tf.cast(weight, tf.float32),
tf.cast(w_t_target_image, tf.float32)) + 1
mse_loss = tf.reduce_mean(
tf.losses.mean_squared_error(w_t_target_image,
net_g.outputs,
weights=weights))
bce_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=t_target_image,
logits=net_g.outputs))
vgg_loss = 0 #2e-3 * tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True) #-6
g_loss = mse_loss + g_gan_loss #+vgg_loss
d_loss1_summary = tf.summary.scalar('Disciminator logits_real loss',
d_loss1)
d_loss2_summary = tf.summary.scalar('Disciminator logits_fake loss',
d_loss2)
d_loss_summary = tf.summary.scalar('Disciminator total loss', d_loss)
g_gan_loss_summary = tf.summary.scalar('Generator GAN loss', g_gan_loss)
mse_loss_summary = tf.summary.scalar('Generator MSE loss', mse_loss)
vgg_loss_summary = tf.summary.scalar('Generator VGG loss', vgg_loss)
g_loss_summary = tf.summary.scalar('Generator total loss', g_loss)
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True) #SRGAN_g
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
tl.layers.initialize_global_variables(sess)
# restore model from .ckpt
saver = tf.train.Saver()
saver.restore(sess, checkpoint_dir + '/model_init_srgan_35.ckpt')
#tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_srgan_20.npz', network=net_g)
#init of generator
#tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_srgan_35_init.npz', network=net_g)
#ac dis
#tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/d_srgan_25.npz', network=net_d)
###============================= LOAD VGG ===============================###
# vgg19_npy_path = "vgg19.npy"
# if not os.path.isfile(vgg19_npy_path):
# print("Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg")
# exit()
# npz = np.load(vgg19_npy_path, encoding='latin1').item()
#
# params = []
# for val in sorted(npz.items()):
# W = np.asarray(val[1][0])
# b = np.asarray(val[1][1])
# print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
# params.extend([W, b])
# tl.files.assign_params(sess, params, net_vgg)
# # net_vgg.print_params(False)
# # net_vgg.print_layers()
#
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:batch_size]
sample_imgs1 = train_lr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
print("sample_imgs size:", len(sample_imgs), sample_imgs[0].shape)
sample_imgs_384 = sample_imgs #tl.prepro.threading_data(sample_imgs, fn=crop_sub_imgs_fn, is_random=False)
#print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(), sample_imgs_384.max())
sample_imgs_96 = sample_imgs1 #tl.prepro.threading_data(sample_imgs_384, fn=downsample_fn)
#Ankit print list of input images
print(sample_imgs_96)
###========================= initialize G ===================mse#
merged_summary_initial_G = tf.summary.merge([mse_loss_summary])
summary_intial_G_writer = tf.summary.FileWriter("./log/train/initial_G")
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
count = 0
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
intial_MSE_G_summary_per_epoch = []
train_hr = []
train_lr = []
for idx in range(0, len(train_hr_imgs) - 4, batch_size):
step_time = time.time()
train_hr = train_hr_imgs[idx:idx + batch_size]
train_lr = train_lr_imgs[idx:idx + batch_size]
## update G
errM, _, mse_summary_initial_G, out1 = sess.run(
[
mse_loss, g_optim_init, merged_summary_initial_G,
net_g.outputs
], {
t_image: train_lr,
t_target_image: train_hr
})
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " %
(epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(mse_summary_initial_G)
intial_G_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
intial_G_summaries[val.tag] = val.simple_value
intial_MSE_G_summary_per_epoch.append(
intial_G_summaries['Generator_MSE_loss'])
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_iter)
print(log)
summary_intial_G_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_Initial_MSE_loss per epoch",
simple_value=np.mean(
intial_MSE_G_summary_per_epoch)),
]), (epoch))
## quick evaluation on train set
#if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(
net_g_test.outputs,
{t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [2, 4],
save_dir_ginit + '/train_%d.png' % epoch)
tl.vis.save_images(np.asarray(sample_imgs_384), [2, 4],
save_dir_ginit + '/train_gt_%d.png' % epoch)
tl.vis.save_images(np.asarray(sample_imgs_96), [2, 4],
save_dir_ginit + '/train_in_%d.png' % epoch)
tl.vis.save_images(out1, [2, 4],
save_dir_ginit + '/train_out_%d.png' % epoch)
# save model as .ckpt
saver = tf.train.Saver(tf.global_variables())
save_path = saver.save(
sess, checkpoint_dir +
"/model_init_{}_{}.ckpt".format(tl.global_flag['mode'], epoch))
###========================= train GAN (SRGAN) =========================###
merged_summary_discriminator = tf.summary.merge(
[d_loss1_summary, d_loss2_summary, d_loss_summary])
summary_discriminator_writer = tf.summary.FileWriter(
"./log/train/discriminator")
merged_summary_generator = tf.summary.merge([
g_gan_loss_summary, mse_loss_summary, vgg_loss_summary, g_loss_summary
])
summary_generator_writer = tf.summary.FileWriter("./log/train/generator")
learning_rate_writer = tf.summary.FileWriter("./log/train/learning_rate")
count = 0
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
learning_rate_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Learning_rate per epoch",
simple_value=(lr_init * new_lr_decay)),
]), (epoch))
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
learning_rate_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Learning_rate per epoch",
simple_value=lr_init),
]), (epoch))
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
loss_per_batch = []
d_loss1_summary_per_epoch = []
d_loss2_summary_per_epoch = []
d_loss_summary_per_epoch = []
g_gan_loss_summary_per_epoch = []
mse_loss_summary_per_epoch = []
vgg_loss_summary_per_epoch = []
g_loss_summary_per_epoch = []
img_target = []
img_in = []
for idx in range(0, len(train_hr_imgs) - 4, batch_size):
step_time = time.time()
## update D
img_in = train_lr_imgs[idx:idx + batch_size]
img_target = train_hr_imgs[idx:idx + batch_size]
errD, _, discriminator_summary = sess.run(
[d_loss, d_optim, merged_summary_discriminator], {
t_image: img_in,
t_target_image: img_target
})
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(discriminator_summary)
discriminator_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
discriminator_summaries[val.tag] = val.simple_value
d_loss1_summary_per_epoch.append(
discriminator_summaries['Disciminator_logits_real_loss'])
d_loss2_summary_per_epoch.append(
discriminator_summaries['Disciminator_logits_fake_loss'])
d_loss_summary_per_epoch.append(
discriminator_summaries['Disciminator_total_loss'])
## update G- GMV
errG, errM, errA, _, generator_summary = sess.run(
[
g_loss, mse_loss, g_gan_loss, g_optim,
merged_summary_generator
], {
t_image: img_in,
t_target_image: img_target
})
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(generator_summary)
#print("generator_summary", summary_pb, type(summary_pb))
generator_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
generator_summaries[val.tag] = val.simple_value
#print("generator_summaries:", generator_summaries)
g_gan_loss_summary_per_epoch.append(
generator_summaries['Generator_GAN_loss'])
mse_loss_summary_per_epoch.append(
generator_summaries['Generator_MSE_loss'])
vgg_loss_summary_per_epoch.append(
generator_summaries['Generator_VGG_loss'])
g_loss_summary_per_epoch.append(
generator_summaries['Generator_total_loss'])
print(
"Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f adv: %.6f)"
% (epoch, n_epoch, n_iter, time.time() - step_time, errD, errG,
errM, errA))
total_d_loss += errD
total_g_loss += errG
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
#####
#
# logging discriminator summary
#
######
# logging per epcoch summary of logit_real_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_logits_real_loss per epoch",
simple_value=np.mean(
d_loss1_summary_per_epoch)),
]), (epoch))
# logging per epcoch summary of logit_fake_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_logits_fake_loss per epoch",
simple_value=np.mean(
d_loss2_summary_per_epoch)),
]), (epoch))
# logging per epcoch summary of total_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_total_loss per epoch",
simple_value=np.mean(
d_loss_summary_per_epoch)),
]), (epoch))
#####
#
# logging generator summary
#
######
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_GAN_loss per epoch",
simple_value=np.mean(
g_gan_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_MSE_loss per epoch",
simple_value=np.mean(
mse_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_VGG_loss per epoch",
simple_value=np.mean(
vgg_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_total_loss per epoch",
simple_value=np.mean(
g_loss_summary_per_epoch)),
]), (epoch))
## quick evaluation on train set
#if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(
net_g_test.outputs,
{t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [2, 4], save_dir_gan + '/train_%d.png' % epoch)
tl.vis.save_images(np.asarray(sample_imgs_384), [2, 4],
save_dir_gan + '/train_gt_%d.png' % epoch)
tl.vis.save_images(np.asarray(sample_imgs_96), [2, 4],
save_dir_gan + '/train_in_%d.png' % epoch)
## save model
if (epoch % 5 == 0):
# save model as .ckpt
saver = tf.train.Saver(tf.global_variables())
save_path = saver.save(
sess, checkpoint_dir +
"/model_{}_{}.ckpt".format(tl.global_flag['mode'], epoch))
def train(train_lr_imgs, train_hr_imgs):
## create folders to save result images and trained model
checkpoint_dir = "models_checkpoints"
tl.files.exists_or_mkdir(checkpoint_dir)
###========================== DEFINE MODEL ============================###
## train inference
t_image = tf.placeholder(dtype='float32',
shape=(batch_size, 512, 512, 1),
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder(dtype='float32',
shape=(batch_size, 512, 512, 1),
name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(
t_target_image, size=[224, 224], method=0, align_corners=False
) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(
net_g.outputs, size=[224, 224], method=0,
align_corners=False) # resize_generate_image_for_vgg
net_vgg, vgg_target_emb = Vgg19_simple_api(input=(t_target_image_224 + 1) /
2,
reuse=False)
_, vgg_predict_emb = Vgg19_simple_api(input=(t_predict_image_224 + 1) / 2,
reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs,
t_target_image,
is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
tl.layers.initialize_global_variables(sess)
if tl.files.load_and_assign_npz(
sess=sess,
name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']),
network=net_g) is False:
tl.files.load_and_assign_npz(
sess=sess,
name=checkpoint_dir +
'/g_{}_init.npz'.format(tl.global_flag['mode']),
network=net_g)
tl.files.load_and_assign_npz(sess=sess,
name=checkpoint_dir +
'/d_{}.npz'.format(tl.global_flag['mode']),
network=net_d)
###============================= LOAD VGG ===============================###
vgg19_npy_path = "vgg19.npy"
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
if val[0] == 'conv1_1':
W = np.mean(W, axis=2)
W = W.reshape((3, 3, 1, 64))
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
###============================= TRAINING ===============================###
###========================= initialize G ====================###
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
start_time = time.time()
for epoch in range(0, n_epoch_init):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
step_time = None
for idx in range(0, len(train_hr_imgs), batch_size):
if idx % 1000 == 0: step_time = time.time()
b_imgs_hr = train_hr_imgs[idx:idx + batch_size]
b_imgs_lr = train_lr_imgs[idx:idx + batch_size]
b_imgs_hr = np.asarray(b_imgs_hr).reshape(
(batch_size, 512, 512, 1))
b_imgs_lr = np.asarray(b_imgs_lr).reshape(
(batch_size, 512, 512, 1))
## update G
errM, _ = sess.run([mse_loss, g_optim_init], {
t_image: b_imgs_lr,
t_target_image: b_imgs_hr
})
if idx % 1000 == 0:
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " %
(epoch, n_epoch_init, n_iter, time.time() - step_time,
errM))
tl.files.save_npz(
net_g.all_params,
name=checkpoint_dir +
'/g_{}_init.npz'.format(tl.global_flag['mode']),
sess=sess)
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_iter)
print(log)
## save model
tl.files.save_npz(net_g.all_params,
name=checkpoint_dir +
'/g_{}_init.npz'.format(tl.global_flag['mode']),
sess=sess)
print("G init took: %4.4fs" % (time.time() - start_time))
###========================= train GAN (SRGAN) =========================###
start_time = time.time()
epoch_losses = defaultdict(list)
iter_losses = defaultdict(list)
for epoch in range(0, n_epoch):
## update learning rate
if epoch != 0 and decay_every != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
step_time = None
for idx in range(0, len(train_hr_imgs), batch_size):
if idx % 1000 == 0: step_time = time.time()
b_imgs_hr = train_hr_imgs[idx:idx + batch_size]
b_imgs_lr = train_lr_imgs[idx:idx + batch_size]
b_imgs_hr = np.asarray(b_imgs_hr).reshape(
(batch_size, 512, 512, 1))
b_imgs_lr = np.asarray(b_imgs_lr).reshape(
(batch_size, 512, 512, 1))
## update D
errD, _ = sess.run([d_loss, d_optim], {
t_image: b_imgs_lr,
t_target_image: b_imgs_hr
})
## update G
errG, errM, errV, errA, _ = sess.run(
[g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {
t_image: b_imgs_lr,
t_target_image: b_imgs_hr
})
if idx % 1000 == 0:
print(
"Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)"
% (epoch, n_epoch, n_iter, time.time() - step_time, errD,
errG, errM, errV, errA))
tl.files.save_npz(net_g.all_params,
name=checkpoint_dir +
'/g_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
tl.files.save_npz(net_d.all_params,
name=checkpoint_dir +
'/d_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
total_d_loss += errD
total_g_loss += errG
n_iter += 1
iter_losses['d_loss'].append(errD)
iter_losses['g_loss'].append(errG)
iter_losses['mse_loss'].append(errM)
iter_losses['vgg_loss'].append(errV)
iter_losses['adv_loss'].append(errA)
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
epoch_losses['d_loss'].append(total_d_loss)
epoch_losses['g_loss'].append(total_g_loss)
## save model
tl.files.save_npz(net_g.all_params,
name=checkpoint_dir +
'/g_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
tl.files.save_npz(net_d.all_params,
name=checkpoint_dir +
'/d_{}.npz'.format(tl.global_flag['mode']),
sess=sess)
print("G train took: %4.4fs" % (time.time() - start_time))
## create visualizations for losses from training
plot_total_losses(epoch_losses)
plot_iterative_losses(iter_losses)
for loss, values in epoch_losses.items():
np.save(checkpoint_dir + "/epoch_" + loss + '.npy', np.asarray(values))
for loss, values in iter_losses.items():
np.save(checkpoint_dir + "/iter_" + loss + '.npy', np.asarray(values))
print("[*] saved losses")
def train():
n_epoch_init = 12
## create folders to save result images and trained model
# save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
# save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
# tl.files.exists_or_mkdir(save_dir_ginit)
# tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
log_dir = "logs" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(
get_synthia_imgs_list(config.VALID.hr_img_path,
is_train=True,
synthia_dataset=config.TRAIN.hr_img_path))
valid_hr_img_list = sorted(
get_synthia_imgs_list(config.VALID.hr_img_path,
is_train=False,
synthia_dataset=config.TRAIN.hr_img_path))
print(len(train_hr_img_list))
print(len(valid_hr_img_list))
###========================== DEFINE MODEL ============================###
## train inference
t_input = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='t_input')
# try with log?
t_input = tf.log(t_input)
d_flg = tf.placeholder(tf.bool, name='is_train')
t_image, t_target_image, t_interpolated = preprocess(t_input)
net_g_outputs = SRGAN_g(t_image,
t_interpolated,
is_train=d_flg,
reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=d_flg, reuse=False)
_, logits_fake = SRGAN_d(net_g_outputs, is_train=d_flg, reuse=True)
vgg_model_true = VGG16(vgg16_npy_path)
vgg_model_gen = VGG16(vgg16_npy_path)
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
# to 3 channels
y_true_normalized = (t_target_image - tf.reduce_min(t_target_image)) / (
tf.reduce_max(t_target_image) - tf.reduce_min(t_target_image))
gen_normalized = (net_g_outputs - tf.reduce_min(net_g_outputs)) / (
tf.reduce_max(net_g_outputs) - tf.reduce_min(net_g_outputs))
t_target_image_3ch = tf.concat([y_true_normalized] * 3, 3)
t_predict_image_3ch = tf.concat([gen_normalized] * 3, 3)
vgg_model_true.build(t_target_image_3ch)
true_features = vgg_model_true.conv3_1
vgg_model_gen.build(t_predict_image_3ch)
gen_features = vgg_model_gen.conv3_1
## test inference
net_g_test = SRGAN_g(t_image, t_interpolated, is_train=d_flg, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
# d_vgg_loss = 2e-6*tl.cost.mean_squared_error(true_features, gen_features, is_mean=True)
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-2 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g') # 1e-3 *
mse_loss = tl.cost.mean_squared_error(net_g_outputs,
t_target_image,
is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
true_features, gen_features, is_mean=True) # 2e-6 *
tv_loss = 2e-6 * tf.reduce_mean(tf.square(net_g_outputs[:, :-1, :, :] - net_g_outputs[:, 1:, :, :])) + \
tf.reduce_mean(tf.square(net_g_outputs[:, :, :-1, :] - net_g_outputs[:, :, 1:, :])) # 2e-6*
g_init_loss = mse_loss + vgg_loss # mse_loss # + vgg_loss + tv_loss
g_loss = g_gan_loss + mse_loss + vgg_loss # + mse_loss
g_vars = tl.layers.get_variables_with_name('G_Depth_SR', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
glob_step_t = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
g_init_loss, var_list=g_vars, global_step=glob_step_t)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
g_loss, var_list=g_vars, global_step=glob_step_t)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
saver = tf.train.Saver(max_to_keep=5)
saver_d = tf.train.Saver(d_vars, max_to_keep=5)
saver_g = tf.train.Saver(g_vars, max_to_keep=5)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
tl.layers.initialize_global_variables(sess)
with tf.variable_scope('summaries'):
tf.summary.scalar('d_loss', d_loss)
tf.summary.scalar('g_loss', g_loss)
tf.summary.scalar('mse_loss', mse_loss)
tf.summary.scalar('vgg_loss', vgg_loss)
tf.summary.scalar('tv_loss', tv_loss)
tf.summary.scalar('g_gan_loss', g_gan_loss)
mae = tf.reduce_mean(
tf.abs(net_g_outputs - t_target_image) /
(t_target_image + tf.constant(1e-8)))
rmse = tf.sqrt(
tf.reduce_mean(tf.square(net_g_outputs - t_target_image)))
tf.summary.scalar('MAE', mae)
tf.summary.scalar('RMSE', rmse)
tf.summary.scalar('learning_rate', lr_v)
# tf.summary.image('input', t_input , max_outputs=1)
tf.summary.image('GT', t_target_image, max_outputs=1)
tf.summary.image('input_small_size', t_image, max_outputs=1)
tf.summary.image('interpolated', t_interpolated, max_outputs=1)
tf.summary.image('result', net_g_outputs, max_outputs=1)
summary_op = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(log_dir + '/train', sess.graph)
test_writer = tf.summary.FileWriter(log_dir + '/test')
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
# sample_imgs = train_hr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
# sample_imgs = tl.prepro.threading_data(train_hr_img_list[0:batch_size], fn=get_imgs_fn) # if no pre-load train set
# print('sample images:', sample_imgs.shape, sample_imgs.min(), sample_imgs.max())
n_batches = int(len(train_hr_img_list) / batch_size)
n_batches_valid = int(len(valid_hr_img_list) / batch_size)
###========================= initialize G ====================###
if not do_init_g:
n_epoch_init = -1
try:
saver_g.restore(
sess, tf.train.latest_checkpoint(checkpoint_dir + '/g_init'))
except Exception as e:
print(
' ** You need to initialize generator: put do_init_g to True or provide a valid restore path'
)
raise e
else:
try:
#saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir+'/gan')) # 2 round
saver.restore(
sess, tf.train.latest_checkpoint(checkpoint_dir + '/g_init'))
except:
print(' ** Creating new g_init model')
pass
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
train_iter, test_iter = 0, 0
for epoch in range(0, n_epoch_init + 1):
try:
epoch_time = time.time()
val_mae, val_mse, val_g_loss = 0, 0, 0
batch_it = tqdm(SynthiaIterator(valid_hr_img_list,
batchsize=batch_size,
shuffle=True,
buffer_size=70),
total=n_batches_valid,
leave=False)
for b in batch_it:
xb = b[0]
errM, errG, mae_score = sess.run([mse_loss, g_loss, mae],
feed_dict={
t_input: xb,
d_flg: False
})
val_mae += mae_score
val_mse += errM
val_g_loss += errG
print("Validation: Epoch {0} val mae {1} val mse {2}".format(
epoch - 1, val_mae / n_batches_valid,
val_mse / n_batches_valid))
total_mse_loss, total_g_loss = 0, 0
batch_it = tqdm(SynthiaIterator(train_hr_img_list,
batchsize=batch_size,
shuffle=True,
buffer_size=70),
total=n_batches,
leave=False)
for b in batch_it:
xb = b[0]
xb = augment_imgs(xb)
glob_step, errM, errG, _ = sess.run(
[glob_step_t, mse_loss, g_loss, g_optim_init],
feed_dict={
t_input: xb,
d_flg: True
})
total_mse_loss += errM
total_g_loss += errG
if (train_iter + 1) % 200 == 0:
summary = sess.run(summary_op,
feed_dict={
t_input: xb,
d_flg: False
})
train_writer.add_summary(summary, train_iter + 1)
train_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_batches)
val_mse_summary = tf.Summary.Value(tag='g_init/val_mse_loss',
simple_value=val_mse /
n_batches_valid)
val_g_loss_summary = tf.Summary.Value(tag='g_init/val_loss',
simple_value=val_g_loss /
n_batches_valid)
train_mse_loss_summary = tf.Summary.Value(
tag='g_init/train_mse_loss',
simple_value=total_mse_loss / n_batches)
train_g_loss_summary = tf.Summary.Value(tag='g_init/train_loss',
simple_value=total_g_loss /
n_batches)
epoch_summary = tf.Summary(value=[
val_mse_summary, val_g_loss_summary, train_mse_loss_summary,
train_g_loss_summary
])
train_writer.add_summary(epoch_summary, glob_step)
print(log)
saver.save(
sess,
os.path.join(checkpoint_dir + '/g_init',
'model' + str(epoch) + '.ckpt'))
except Exception as e:
batch_it.iterable.stop()
raise e
###========================= train GAN (SRGAN) =========================###
try:
# saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir+'/g_init'))
# saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir+'/gan'))
pass
except:
print(' ** Creating new GAN model')
pass
train_iter, test_iter = 0, 0
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
try:
epoch_time = time.time()
val_mae, val_mse, val_g_loss, val_d_loss = 0, 0, 0, 0
batch_it = tqdm(SynthiaIterator(valid_hr_img_list,
batchsize=batch_size,
shuffle=True,
buffer_size=70),
total=n_batches_valid,
leave=False)
for b in batch_it:
xb = b[0]
errM, mae_score, errG, errD = sess.run(
[mse_loss, mae, g_loss, d_loss],
feed_dict={
t_input: xb,
d_flg: False
})
val_mae += mae_score
val_mse += errM
val_g_loss += errG
val_d_loss += errD
print("Validation (GAN): Epoch {0} val mae {1} val mse {2}".format(
epoch - 1, val_mae / n_batches_valid,
val_mse / n_batches_valid))
total_d_loss, total_g_loss, total_mse_loss = 0, 0, 0
batch_it = tqdm(SynthiaIterator(train_hr_img_list,
batchsize=batch_size,
shuffle=True,
buffer_size=70),
total=n_batches,
leave=False)
for b in batch_it:
xb = b[0]
xb = augment_imgs(xb)
## update D
errD, _ = sess.run([d_loss, d_optim], {
t_input: xb,
d_flg: True
})
## update G
glob_step, errG, errM, _, summary = sess.run(
[glob_step_t, g_loss, mse_loss, g_optim, summary_op], {
t_input: xb,
d_flg: True
})
total_mse_loss += errM
total_d_loss += errD
total_g_loss += errG
if (train_iter + 1) % 10 == 0:
train_writer.add_summary(summary, train_iter + 1)
train_iter += 1
except Exception as e:
batch_it.iterable.stop()
raise e
break
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f mse_loss: %.8f" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_batches,
total_g_loss / n_batches, total_mse_loss / n_batches)
val_mse_summary = tf.Summary.Value(tag='gan/val_mse_loss',
simple_value=val_mse /
n_batches_valid)
val_g_loss_summary = tf.Summary.Value(tag='gan/val_g_loss',
simple_value=val_g_loss /
n_batches_valid)
val_d_loss_summary = tf.Summary.Value(tag='gan/val_d_loss',
simple_value=val_d_loss /
n_batches_valid)
train_mse_loss_summary = tf.Summary.Value(tag='gan/train_mse_loss',
simple_value=total_mse_loss /
n_batches)
train_g_loss_summary = tf.Summary.Value(tag='gan/train_g_loss',
simple_value=total_g_loss /
n_batches)
train_d_loss_summary = tf.Summary.Value(tag='gan/train_d_loss',
simple_value=total_d_loss /
n_batches)
epoch_summary = tf.Summary(value=[
val_mse_summary, val_g_loss_summary, val_d_loss_summary,
train_mse_loss_summary, train_g_loss_summary, train_d_loss_summary
])
train_writer.add_summary(epoch_summary, glob_step)
print(log)
saver.save(
sess,
os.path.join(checkpoint_dir + '/gan',
'model' + str(n_epoch_init + epoch) + '.ckpt'))
def train():
## create folders to save result images and trained model
save_dir_gan = samples_path + "gan"
tl.files.exists_or_mkdir(save_dir_gan)
tl.files.exists_or_mkdir(checkpoint_path)
###====================== PRE-LOAD DATA ===========================###
valid_hr_img_list = sorted(
tl.files.load_file_list(path=valid_hr_img_path,
regx='.*\.(bmp|png|webp|jpg)',
printable=False))
###========================== DEFINE MODEL ============================###
## train inference
sample_t_image = tf.compat.v1.placeholder(
'float32', [sample_batch_size, 96, 96, 3],
name='sample_t_image_input_to_SRGAN_generator')
t_image = tf.compat.v1.placeholder('float32', [batch_size, 96, 96, 3],
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.compat.v1.placeholder('float32',
[batch_size, 384, 384, 3],
name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## test inference
net_g_test = SRGAN_g(sample_t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
# MAE Loss
mae_loss = tf.reduce_mean(tf.map_fn(tf.abs,
t_target_image - net_g.outputs))
# GAN Loss
d_loss = 0.5 * (
tf.reduce_mean(
tf.square(logits_real - tf.reduce_mean(logits_fake) - 1)) +
tf.reduce_mean(
tf.square(logits_fake - tf.reduce_mean(logits_real) + 1)))
g_gan_loss = 0.5 * (
tf.reduce_mean(
tf.square(logits_real - tf.reduce_mean(logits_fake) + 1)) +
tf.reduce_mean(
tf.square(logits_fake - tf.reduce_mean(logits_real) - 1)))
g_loss = 1e-1 * g_gan_loss + mae_loss
d_real = tf.reduce_mean(logits_real)
d_fake = tf.reduce_mean(logits_fake)
with tf.variable_scope('learning_rate'):
learning_rate_var = tf.Variable(learning_rate, trainable=False)
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
## SRGAN
g_optim = tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate_var).minimize(g_loss, var_list=g_vars)
d_optim = tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate_var).minimize(d_loss, var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
sess.run(tf.variables_initializer(tf.global_variables()))
tl.files.load_and_assign_npz(sess=sess,
name=checkpoint_path + 'g.npz',
network=net_g)
tl.files.load_and_assign_npz(sess=sess,
name=checkpoint_path + 'd.npz',
network=net_d)
###============================= TRAINING ===============================###
sample_imgs = tl.prepro.threading_data(
valid_hr_img_list[0:sample_batch_size],
fn=get_imgs_fn,
path=valid_hr_img_path)
sample_imgs_384 = tl.prepro.threading_data(sample_imgs,
fn=crop_sub_imgs_fn,
is_random=False)
print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(),
sample_imgs_384.max())
sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384,
fn=downsample_fn)
print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(),
sample_imgs_96.max())
save_images(sample_imgs_96, [ni, ni], save_file_format,
save_dir_gan + '/_train_sample_96')
save_images(sample_imgs_384, [ni, ni], save_file_format,
save_dir_gan + '/_train_sample_384')
###========================= train GAN =========================###
sess.run(tf.assign(learning_rate_var, learning_rate))
for epoch in range(0, n_epoch_gan + 1):
epoch_time = time.time()
total_d_loss, total_g_loss_mae, total_g_loss_gan, n_iter = 0, 0, 0, 0
train_hr_img_list = load_deep_file_list(path=train_hr_img_path,
regx='.*\.(bmp|png|webp|jpg)',
recursive=True,
printable=False)
random.shuffle(train_hr_img_list)
list_length = len(train_hr_img_list)
print("Number of images: %d" % (list_length))
if list_length % batch_size != 0:
train_hr_img_list += train_hr_img_list[0:batch_size -
list_length % batch_size:1]
list_length = len(train_hr_img_list)
print("Length of list: %d" % (list_length))
for idx in range(0, list_length, batch_size):
step_time = time.time()
b_imgs_list = train_hr_img_list[idx:idx + batch_size]
b_imgs = tl.prepro.threading_data(b_imgs_list,
fn=get_imgs_fn,
path=train_hr_img_path)
b_imgs_384 = tl.prepro.threading_data(b_imgs,
fn=crop_data_augment_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
b_imgs_384 = tl.prepro.threading_data(b_imgs_384, fn=rescale_m1p1)
## update D
errD, d_r, d_f, _ = sess.run([d_loss, d_real, d_fake, d_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
## update G
errM, errA, _, _ = sess.run(
[mae_loss, g_gan_loss, g_loss, g_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
print(
"Epoch[%2d/%2d] %4d time: %4.2fs d_loss: %.8f g_loss_mae: %.8f g_loss_gan: %.8f d_r: %.8f d_f: %.8f"
% (epoch, n_epoch_gan, n_iter, time.time() - step_time, errD,
errM, errA, d_r, d_f))
total_d_loss += errD
total_g_loss_mae += errM
total_g_loss_gan += errA
n_iter += 1
log = (
"[*] Epoch[%2d/%2d] time: %4.2fs d_loss: %.8f g_loss_mae: %.8f g_loss_gan: %.8f"
% (epoch, n_epoch_gan, time.time() - epoch_time, total_d_loss /
n_iter, total_g_loss_mae / n_iter, total_g_loss_gan / n_iter))
print(log)
## quick evaluation on train set
out = sess.run(net_g_test.outputs, {sample_t_image: sample_imgs_96})
print("[*] save images")
save_images(out, [ni, ni], save_file_format,
save_dir_gan + '/train_%d' % epoch)
## save model
tl.files.save_npz(net_g.all_params,
name=checkpoint_path + 'g.npz',
sess=sess)
tl.files.save_npz(net_d.all_params,
name=checkpoint_path + 'd.npz',
sess=sess)
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.hr_img_path, regx='.*.png', printable=False))
train_lr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.lr_img_path, regx='.*.png', printable=False))
valid_hr_img_list = sorted(tl.files.load_file_list(path=config.VALID.hr_img_path, regx='.*.png', printable=False))
valid_lr_img_list = sorted(tl.files.load_file_list(path=config.VALID.lr_img_path, regx='.*.png', printable=False))
## If your machine have enough memory, please pre-load the whole train set.
train_hr_imgs = tl.vis.read_images(train_hr_img_list, path=config.TRAIN.hr_img_path, n_threads=32)
# for im in train_hr_imgs:
# print(im.shape)
# valid_lr_imgs = tl.vis.read_images(valid_lr_img_list, path=config.VALID.lr_img_path, n_threads=32)
# for im in valid_lr_imgs:
# print(im.shape)
# valid_hr_imgs = tl.vis.read_images(valid_hr_img_list, path=config.VALID.hr_img_path, n_threads=32)
# for im in valid_hr_imgs:
# print(im.shape)
# exit()
###========================== DEFINE MODEL ============================###
## train inference
t_image = tf.placeholder('float32', [batch_size, 96, 96, 3], name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3], name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(
t_target_image, size=[224, 224], method=0,
align_corners=False) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(net_g.outputs, size=[224, 224], method=0, align_corners=False) # resize_generate_image_for_vgg
net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2, reuse=False)
_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2, reuse=True)
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real, tf.ones_like(logits_real), name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake, tf.zeros_like(logits_fake), name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs, t_target_image, is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(g_loss, var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(d_loss, var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
tl.layers.initialize_global_variables(sess)
if tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), network=net_g) is False:
tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), network=net_g)
tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), network=net_d)
###============================= LOAD VGG ===============================###
vgg19_npy_path = "vgg19.npy"
if not os.path.isfile(vgg19_npy_path):
print("Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg")
exit()
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
# net_vgg.print_params(False)
# net_vgg.print_layers()
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
sample_imgs_384 = tl.prepro.threading_data(sample_imgs, fn=crop_sub_imgs_fn, is_random=False)
print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(), sample_imgs_384.max())
sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384, fn=downsample_fn)
print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(), sample_imgs_96.max())
tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_ginit + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_ginit + '/_train_sample_384.png')
tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_gan + '/_train_sample_96.png')
tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_gan + '/_train_sample_384.png')
###========================= initialize G ====================###
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update G
errM, _ = sess.run([mse_loss, g_optim_init], {t_image: b_imgs_96, t_target_image: b_imgs_384})
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " % (epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (epoch, n_epoch_init, time.time() - epoch_time, total_mse_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {t_image: sample_imgs_96}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni], save_dir_ginit + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), sess=sess)
###========================= train GAN (SRGAN) =========================###
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init * new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update D
errD, _ = sess.run([d_loss, d_optim], {t_image: b_imgs_96, t_target_image: b_imgs_384})
## update G
errG, errM, errV, errA, _ = sess.run([g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {t_image: b_imgs_96, t_target_image: b_imgs_384})
print("Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)" %
(epoch, n_epoch, n_iter, time.time() - step_time, errD, errG, errM, errV, errA))
total_d_loss += errD
total_g_loss += errG
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {t_image: sample_imgs_96}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni], save_dir_gan + '/train_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), sess=sess)
tl.files.save_npz(net_d.all_params, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), sess=sess)
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
###====================== PRE-LOAD DATA ===========================###
train_hr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.hr_img_path,
regx='.*.png',
printable=False))
train_lr_img_list = sorted(
tl.files.load_file_list(path=config.TRAIN.lr_img_path,
regx='.*.png',
printable=False))
valid_hr_img_list = sorted(
tl.files.load_file_list(path=config.VALID.hr_img_path,
regx='.*.png',
printable=False))
valid_lr_img_list = sorted(
tl.files.load_file_list(path=config.VALID.lr_img_path,
regx='.*.png',
printable=False))
## If your machine have enough memory, please pre-load the whole train set.
print("reading images")
train_hr_imgs = [] #[None] * len(train_hr_img_list)
#sess = tf.Session()
for img__ in train_hr_img_list:
image_loaded = scipy.misc.imread(os.path.join(config.TRAIN.hr_img_path,
img__),
mode='L')
image_loaded = image_loaded.reshape(
(image_loaded.shape[0], image_loaded.shape[1], 1))
train_hr_imgs.append(image_loaded)
print(type(train_hr_imgs), len(train_hr_img_list))
###========================== DEFINE MODEL ============================###
## train inference
#t_image = tf.placeholder('float32', [batch_size, 96, 96, 3], name='t_image_input_to_SRGAN_generator')
#t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3], name='t_target_image')
t_image = tf.placeholder('float32', [batch_size, 28, 224, 1],
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder(
'float32', [batch_size, 224, 224, 1], name='t_target_image'
) # may have to convert 224x224x1 into 224x224x3, with channel 1 & 2 as 0. May have to have separate place-holder ?
print("t_image:", tf.shape(t_image))
print("t_target_image:", tf.shape(t_target_image))
net_g = SRGAN_g(t_image, is_train=True, reuse=False)
print("net_g.outputs:", tf.shape(net_g.outputs))
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## vgg inference. 0, 1, 2, 3 BILINEAR NEAREST BICUBIC AREA
t_target_image_224 = tf.image.resize_images(
t_target_image, size=[224, 224], method=0, align_corners=False
) # resize_target_image_for_vgg # http://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers.html#UpSampling2dLayer
t_predict_image_224 = tf.image.resize_images(
net_g.outputs, size=[224, 224], method=0,
align_corners=False) # resize_generate_image_for_vgg
## Added as VGG works for RGB and expects 3 channels.
t_target_image_224 = tf.image.grayscale_to_rgb(t_target_image_224)
t_predict_image_224 = tf.image.grayscale_to_rgb(t_predict_image_224)
print("net_g.outputs:", tf.shape(net_g.outputs))
print("t_predict_image_224:", tf.shape(t_predict_image_224))
net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2,
reuse=False)
_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2,
reuse=True)
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs,
t_target_image,
is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + vgg_loss + g_gan_loss
d_loss1_summary = tf.summary.scalar('Disciminator logits_real loss',
d_loss1)
d_loss2_summary = tf.summary.scalar('Disciminator logits_fake loss',
d_loss2)
d_loss_summary = tf.summary.scalar('Disciminator total loss', d_loss)
g_gan_loss_summary = tf.summary.scalar('Generator GAN loss', g_gan_loss)
mse_loss_summary = tf.summary.scalar('Generator MSE loss', mse_loss)
vgg_loss_summary = tf.summary.scalar('Generator VGG loss', vgg_loss)
g_loss_summary = tf.summary.scalar('Generator total loss', g_loss)
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
# UNCOMMENT THE LINE BELOW!!!
#g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
tl.layers.initialize_global_variables(sess)
#if tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), network=net_g) is False:
# tl.fites.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), network=net_g)
#tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), network=net_d)
###============================= LOAD VGG ===============================###
vgg19_npy_path = "vgg19.npy"
if not os.path.isfile(vgg19_npy_path):
print(
"Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg"
)
exit()
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
# net_vgg.print_params(False)
# net_vgg.print_layers()
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:batch_size]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
print("sample_imgs size:", len(sample_imgs), sample_imgs[0].shape)
sample_imgs_384 = tl.prepro.threading_data(sample_imgs,
fn=crop_sub_imgs_fn,
is_random=False)
print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(),
sample_imgs_384.max())
sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384,
fn=downsample_fn_mod)
print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(),
sample_imgs_96.max())
#tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_ginit + '/_train_sample_96.png')
#tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_ginit + '/_train_sample_384.png')
#tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_gan + '/_train_sample_96.png')
#tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_gan + '/_train_sample_384.png')
'''
###========================= initialize G ====================###
merged_summary_initial_G = tf.summary.merge([mse_loss_summary])
summary_intial_G_writer = tf.summary.FileWriter("./log/train/initial_G")
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
count = 0
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
intial_MSE_G_summary_per_epoch = []
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn_mod)
## update G
errM, _, mse_summary_initial_G = sess.run([mse_loss, g_optim_init, merged_summary_initial_G], {t_image: b_imgs_96, t_target_image: b_imgs_384})
print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " % (epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(mse_summary_initial_G)
intial_G_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
intial_G_summaries[val.tag] = val.simple_value
#print("intial_G_summaries:", intial_G_summaries)
intial_MSE_G_summary_per_epoch.append(intial_G_summaries['Generator_MSE_loss'])
#summary_intial_G_writer.add_summary(mse_summary_initial_G, (count + 1)) #(epoch + 1)*(n_iter+1))
#count += 1
total_mse_loss += errM
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (epoch, n_epoch_init, time.time() - epoch_time, total_mse_loss / n_iter)
print(log)
summary_intial_G_writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag="Generator_Initial_MSE_loss per epoch", simple_value=np.mean(intial_MSE_G_summary_per_epoch)),]), (epoch))
## quick evaluation on train set
#if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {t_image: sample_imgs_96}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
for im in range(len(out)):
if(im%4==0 or im==1197):
tl.vis.save_image(out[im], save_dir_ginit + '/train_%d_%d.png' % (epoch,im))
## save model
saver=tf.train.Saver()
if (epoch%10==0 and epoch!=0):
saver.save(sess, 'checkpoint/init_'+str(epoch)+'.ckpt')
#if (epoch != 0) and (epoch % 10 == 0):
#tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}_{}_init.npz'.format(tl.global_flag['mode'], epoch), sess=sess)
'''
###========================= train GAN (SRGAN) =========================###
saver = tf.train.Saver()
saver.restore(sess, 'checkpoint/main_10.ckpt')
print('Restored main_10, begin 11/50')
merged_summary_discriminator = tf.summary.merge(
[d_loss1_summary, d_loss2_summary, d_loss_summary])
summary_discriminator_writer = tf.summary.FileWriter(
"./log/train/discriminator")
merged_summary_generator = tf.summary.merge([
g_gan_loss_summary, mse_loss_summary, vgg_loss_summary, g_loss_summary
])
summary_generator_writer = tf.summary.FileWriter("./log/train/generator")
learning_rate_writer = tf.summary.FileWriter("./log/train/learning_rate")
count = 0
for epoch in range(11, n_epoch + 11):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
learning_rate_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Learning_rate per epoch",
simple_value=(lr_init * new_lr_decay)),
]), (epoch))
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
learning_rate_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Learning_rate per epoch",
simple_value=lr_init),
]), (epoch))
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine cannot load all images into memory, you should use
## this one to load batch of images while training.
# random.shuffle(train_hr_img_list)
# for idx in range(0, len(train_hr_img_list), batch_size):
# step_time = time.time()
# b_imgs_list = train_hr_img_list[idx : idx + batch_size]
# b_imgs = tl.prepro.threading_data(b_imgs_list, fn=get_imgs_fn, path=config.TRAIN.hr_img_path)
# b_imgs_384 = tl.prepro.threading_data(b_imgs, fn=crop_sub_imgs_fn, is_random=True)
# b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## If your machine have enough memory, please pre-load the whole train set.
loss_per_batch = []
d_loss1_summary_per_epoch = []
d_loss2_summary_per_epoch = []
d_loss_summary_per_epoch = []
g_gan_loss_summary_per_epoch = []
mse_loss_summary_per_epoch = []
vgg_loss_summary_per_epoch = []
g_loss_summary_per_epoch = []
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384,
fn=downsample_fn_mod)
## update D
errD, _, discriminator_summary = sess.run(
[d_loss, d_optim, merged_summary_discriminator], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(discriminator_summary)
#print("discriminator_summary", summary_pb, type(summary_pb))
discriminator_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
discriminator_summaries[val.tag] = val.simple_value
d_loss1_summary_per_epoch.append(
discriminator_summaries['Disciminator_logits_real_loss'])
d_loss2_summary_per_epoch.append(
discriminator_summaries['Disciminator_logits_fake_loss'])
d_loss_summary_per_epoch.append(
discriminator_summaries['Disciminator_total_loss'])
## update G
errG, errM, errV, errA, _, generator_summary = sess.run(
[
g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim,
merged_summary_generator
], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
summary_pb = tf.summary.Summary()
summary_pb.ParseFromString(generator_summary)
#print("generator_summary", summary_pb, type(summary_pb))
generator_summaries = {}
for val in summary_pb.value:
# Assuming all summaries are scalars.
generator_summaries[val.tag] = val.simple_value
#print("generator_summaries:", generator_summaries)
g_gan_loss_summary_per_epoch.append(
generator_summaries['Generator_GAN_loss'])
mse_loss_summary_per_epoch.append(
generator_summaries['Generator_MSE_loss'])
vgg_loss_summary_per_epoch.append(
generator_summaries['Generator_VGG_loss'])
g_loss_summary_per_epoch.append(
generator_summaries['Generator_total_loss'])
#summary_generator_writer.add_summary(generator_summary, (count + 1))
#summary_total = sess.run(summary_total_merged, {t_image: b_imgs_96, t_target_image: b_imgs_384})
#summary_total_merged_writer.add_summary(summary_total, (count + 1))
#count += 1
tot_epoch = n_epoch + 10
print(
"Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)"
% (epoch, tot_epoch, n_iter, time.time() - step_time, errD,
errG, errM, errV, errA))
total_d_loss += errD
total_g_loss += errG
n_iter += 1
#remove this for normal running:
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (
epoch, tot_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
#####
#
# logging discriminator summary
#
######
# logging per epcoch summary of logit_real_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_logits_real_loss per epoch",
simple_value=np.mean(
d_loss1_summary_per_epoch)),
]), (epoch))
# logging per epcoch summary of logit_fake_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_logits_fake_loss per epoch",
simple_value=np.mean(
d_loss2_summary_per_epoch)),
]), (epoch))
# logging per epcoch summary of total_loss per epoch. Value logged is averaged across batches used per epoch.
summary_discriminator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Disciminator_total_loss per epoch",
simple_value=np.mean(
d_loss_summary_per_epoch)),
]), (epoch))
#####
#
# logging generator summary
#
######
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_GAN_loss per epoch",
simple_value=np.mean(
g_gan_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_MSE_loss per epoch",
simple_value=np.mean(
mse_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_VGG_loss per epoch",
simple_value=np.mean(
vgg_loss_summary_per_epoch)),
]), (epoch))
summary_generator_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="Generator_total_loss per epoch",
simple_value=np.mean(
g_loss_summary_per_epoch)),
]), (epoch))
## quick evaluation on train set
#if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(
net_g_test.outputs,
{t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
## save model
if (epoch % 10 == 0 and epoch != 0):
saver.save(sess, 'checkpoint/main_' + str(epoch) + '.ckpt')
print("[*] save images")
for im in range(len(out)):
tl.vis.save_image(
out[im], save_dir_gan + '/train_%d_%d.png' % (epoch, im))
def train():
## create folders to save result images and trained model
save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])
save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])
save_dir_valid = "samples/{}_valid".format(tl.global_flag['mode'])
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
tl.files.exists_or_mkdir(save_dir_valid)
checkpoint_dir = "checkpoint" # checkpoint_resize_conv
tl.files.exists_or_mkdir(checkpoint_dir)
train_hr_imgs = read_csv_data(config.TRAIN.hr_img_path,
width=48,
height=48,
channel=1)
valid_hr_imgs = read_csv_data(config.VALID.hr_img_path,
width=48,
height=48,
channel=1)
###========================== DEFINE MODEL ============================###
## train inference ## t = train
t_image = tf.placeholder('float32', [None, 16, 16, 1],
name='t_image_input_to_SRGAN_generator')
t_target_image = tf.placeholder('float32', [None, 48, 48, 1],
name='t_target_image')
net_g = SRGAN_g(t_image, is_train=True, reuse=False, nb_block=16)
net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
net_g.print_params(False)
net_g.print_layers()
net_d.print_params(False)
net_d.print_layers()
## test inference
net_g_test = SRGAN_g(t_image, is_train=False, reuse=True)
# ###========================== DEFINE TRAIN OPS ==========================###
# d_loss: for discriminator
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
# g_loss: for generator
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs,
t_target_image,
is_mean=True)
# vgg_loss = 2e-6 * tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + g_gan_loss
# g_loss = mse_loss + vgg_loss + g_gan_loss
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###============================= TRAINING ===============================###
## use first `batch_size` of train set to have a quick test during training
sample_imgs = train_hr_imgs[0:9]
valid_imgs = valid_hr_imgs[44:53]
# sample_imgs = tl.vis.read_images(train_hr_img_list[0:batch_size], path=config.TRAIN.hr_img_path, n_threads=32) # if no pre-load train set
sample_imgs_384 = tl.prepro.threading_data(sample_imgs,
fn=crop_sub_imgs_fn,
is_random=False)
valid_imgs_48 = tl.prepro.threading_data(valid_imgs,
fn=crop_sub_imgs_fn,
is_random=False)
print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(),
sample_imgs_384.max())
sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384,
fn=downsample_fn,
down_rate=3)
valid_imgs_16 = tl.prepro.threading_data(valid_imgs_48,
fn=downsample_fn,
down_rate=3)
print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(),
sample_imgs_96.max())
tl.vis.save_images(sample_imgs_96, [ni, ni],
save_dir_ginit + '/_train_sample_16.png')
tl.vis.save_images(sample_imgs_384, [ni, ni],
save_dir_ginit + '/_train_sample_48.png')
tl.vis.save_images(sample_imgs_96, [ni, ni],
save_dir_gan + '/_train_sample_16.png')
tl.vis.save_images(sample_imgs_384, [ni, ni],
save_dir_gan + '/_train_sample_48.png')
tl.vis.save_images(valid_imgs_48, [ni, ni],
save_dir_valid + '/_valid_sample_48.png')
tl.vis.save_images(valid_imgs_16, [ni, ni],
save_dir_valid + '/_valid_sample_16.png')
sample_hr_imgs_bicubic = tl.prepro.threading_data(sample_imgs_96,
fn=upsample_fn,
up_rate=3)
valid_hr_imgs_bicubic = tl.prepro.threading_data(valid_imgs_16,
fn=upsample_fn,
up_rate=3)
tl.vis.save_images(sample_hr_imgs_bicubic, [ni, ni],
save_dir_ginit + '/_sample_bicubic_48.png')
tl.vis.save_images(valid_hr_imgs_bicubic, [ni, ni],
save_dir_valid + '/_valid_sample_bicubic_48.png')
###========================= initialize G ====================###
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
train_writer_path = "./log/train"
tl.files.exists_or_mkdir(train_writer_path)
train_writer = tf.summary.FileWriter(train_writer_path,
graph=tf.get_default_graph())
print(" ** fixed learning rate: %f (for init G)" % lr_init)
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, n_iter = 0, 0
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
if idx + batch_size > len(train_hr_imgs):
break
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update G
errM, _ = sess.run([mse_loss, g_optim_init], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
sys.stdout.write(
"Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f \r" %
(epoch, n_epoch_init, n_iter, time.time() - step_time, errM))
sys.stdout.flush()
total_mse_loss += errM
n_iter += 1
log = "\n[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni],
save_dir_ginit + '/train_%d.png' % epoch)
## quick evaluation on validation set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: valid_imgs_16
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
tl.vis.save_images(out, [ni, ni],
save_dir_valid + '/valid_ganit_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(
net_g.all_params,
name=checkpoint_dir +
'/g_{}_init_{}.npz'.format(tl.global_flag['mode'], epoch),
sess=sess)
###========================= train GAN (SRGAN) =========================###
for epoch in range(0, n_epoch + 1):
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, n_iter = 0, 0, 0
## If your machine have enough memory, please pre-load the whole train set.
for idx in range(0, len(train_hr_imgs), batch_size):
step_time = time.time()
b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)
## update D
errD, _ = sess.run([d_loss, d_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
## update G
# errG, errM, errV, errA, _ = sess.run([g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {t_image: b_imgs_96, t_target_image: b_imgs_384})
errG, errM, errA, _ = sess.run(
[g_loss, mse_loss, g_gan_loss, g_optim], {
t_image: b_imgs_96,
t_target_image: b_imgs_384
})
# sys.stdout.write("Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)\n" %
# (epoch, n_epoch, n_iter, time.time() - step_time, errD, errG, errM, errV, errA))
sys.stdout.write(
"Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f adv: %.6f) \r"
% (epoch, n_epoch, n_iter, time.time() - step_time, errD, errG,
errM, errA))
sys.stdout.flush()
total_d_loss += errD
total_g_loss += errG
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter)
print(log)
## quick evaluation on train set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: sample_imgs_96
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
print("[*] save images")
tl.vis.save_images(out, [ni, ni],
save_dir_gan + '/train_%d.png' % epoch)
## quick evaluation on validation set
if (epoch != 0) and (epoch % 10 == 0):
out = sess.run(net_g_test.outputs, {
t_image: valid_imgs_16
}) #; print('gen sub-image:', out.shape, out.min(), out.max())
tl.vis.save_images(out, [ni, ni],
save_dir_valid + '/valid_gan_%d.png' % epoch)
## save model
if (epoch != 0) and (epoch % 10 == 0):
tl.files.save_npz(
net_g.all_params,
name=checkpoint_dir +
'/g_{}_{}.npz'.format(tl.global_flag['mode'], epoch),
sess=sess)
tl.files.save_npz(
net_d.all_params,
name=checkpoint_dir +
'/d_{}_{}.npz'.format(tl.global_flag['mode'], epoch),
sess=sess)
def train(train_lr_path,
train_hr_path,
save_path,
save_every_epoch=2,
validation=True,
ratio=0.9,
batch_size=16,
lr_init=1e-4,
beta1=0.9,
n_epoch_init=10,
n_epoch=20,
lr_decay=0.1):
'''
Parameters:
data:
train_lr_path/train_hr_path: path of data
save_path: the parent folder to save model result
validation: whether to split data into train set and validation set
save_every_epoch: how frequent to save the checkpoints and sample images
Adam:
batch_size
lr_init
beta1
Generator Initialization
n_epoch_init
Adversarial Net
n_epoch
lr_decay
'''
## Folders to save results
save_dir_ginit = os.path.join(save_path, 'srgan_ginit')
save_dir_gan = os.path.join(save_path, 'srgan_gan')
checkpoint_dir = os.path.join(save_path, 'checkpoint')
tl.files.exists_or_mkdir(save_dir_ginit)
tl.files.exists_or_mkdir(save_dir_gan)
tl.files.exists_or_mkdir(checkpoint_dir)
###======LOAD DATA======###
train_lr_img_list = sorted(
tl.files.load_file_list(path=train_lr_path,
regx='.*.jpg',
printable=False))
train_hr_img_list = sorted(
tl.files.load_file_list(path=train_hr_path,
regx='.*.jpg',
printable=False))
if validation:
idx = np.random.choice(len(train_lr_img_list),
int(len(train_lr_img_list) * ratio),
replace=False)
valid_lr_img_list = sorted(
[x for i, x in enumerate(train_lr_img_list) if i not in idx])
valid_hr_img_list = sorted(
[x for i, x in enumerate(train_hr_img_list) if i not in idx])
train_lr_img_list = sorted(
[x for i, x in enumerate(train_lr_img_list) if i in idx])
train_hr_img_list = sorted(
[x for i, x in enumerate(train_hr_img_list) if i in idx])
valid_lr_imgs = tl.vis.read_images(valid_lr_img_list,
path=train_lr_path,
n_threads=32)
valid_hr_imgs = tl.vis.read_images(valid_hr_img_list,
path=train_hr_path,
n_threads=32)
train_lr_imgs = tl.vis.read_images(train_lr_img_list,
path=train_lr_path,
n_threads=32)
train_hr_imgs = tl.vis.read_images(train_hr_img_list,
path=train_hr_path,
n_threads=32)
###======DEFINE MODEL======###
## train inference
lr_image = tf.placeholder('float32', [None, 96, 96, 3], name='lr_image')
hr_image = tf.placeholder('float32', [None, 192, 192, 3], name='hr_image')
net_g = SRGAN_g(lr_image, is_train=True, reuse=False)
net_d, logits_real = SRGAN_d(hr_image, is_train=True, reuse=False)
_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True)
# net_g.print_params(False)
# net_g.print_layers()
# net_d.print_params(False)
# net_d.print_layers()
## resize original hr images for VGG19
hr_image_224 = tf.image.resize_images(
hr_image,
size=[224, 224],
method=0, # BICUBIC
align_corners=False)
## generated hr image for VGG19
generated_image_224 = tf.image.resize_images(
net_g.outputs,
size=[224, 224],
method=0, #BICUBIC
align_corners=False)
## scale image to [0,1] and get conv characteristics
net_vgg, vgg_target_emb = Vgg19_simple_api((hr_image_224 + 1) / 2,
reuse=False)
_, vgg_predict_emb = Vgg19_simple_api((generated_image_224 + 1) / 2,
reuse=True)
## test inference
net_g_test = SRGAN_g(lr_image, is_train=False, reuse=True)
###======DEFINE TRAIN PROCESS======###
d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real,
tf.ones_like(logits_real),
name='d1')
d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake,
tf.zeros_like(logits_fake),
name='d2')
d_loss = d_loss1 + d_loss2
prediction1 = tf.greater(logits_real, tf.fill(tf.shape(logits_real), 0.5))
acc_metric1 = tf.reduce_mean(tf.cast(prediction1, tf.float32))
prediction2 = tf.less(logits_fake, tf.fill(tf.shape(logits_fake), 0.5))
acc_metric2 = tf.reduce_mean(tf.cast(prediction2, tf.float32))
acc_metric = acc_metric1 + acc_metric2
g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(
logits_fake, tf.ones_like(logits_fake), name='g')
mse_loss = tl.cost.mean_squared_error(net_g.outputs,
hr_image,
is_mean=True)
vgg_loss = 2e-6 * tl.cost.mean_squared_error(
vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)
g_loss = mse_loss + g_gan_loss + vgg_loss
psnr_metric = tf.image.psnr(net_g.outputs,
hr_image,
max_val=2.0,
name='psnr')
g_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)
d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr_init, trainable=False)
## Pretrain
g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(
mse_loss, var_list=g_vars)
## SRGAN
g_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(g_loss,
var_list=g_vars)
d_optim = tf.train.AdamOptimizer(lr_v,
beta1=beta1).minimize(d_loss,
var_list=d_vars)
###========================== RESTORE MODEL =============================###
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
sess.run(tf.global_variables_initializer())
if tl.files.file_exists(os.path.join(checkpoint_dir, 'g_srgan.npz')):
tl.files.load_and_assign_npz(sess=sess,
name=os.path.join(checkpoint_dir,
'g_srgan.npz'),
network=net_g)
else:
tl.files.load_and_assign_npz(sess=sess,
name=os.path.join(checkpoint_dir,
'g_srgan_init.npz'),
network=net_g)
tl.files.load_and_assign_npz(sess=sess,
name=os.path.join(checkpoint_dir,
'd_srgan.npz'),
network=net_d)
###======LOAD VGG======###
vgg19_npy_path = '../lib/SRGAN/vgg19.npy'
npz = np.load(vgg19_npy_path, encoding='latin1').item()
params = []
for val in sorted(npz.items()):
W = np.asarray(val[1][0])
b = np.asarray(val[1][1])
print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))
params.extend([W, b])
tl.files.assign_params(sess, params, net_vgg)
# net_vgg.print_params(False)
# net_vgg.print_layers()
###======TRAINING======###
## use train set to have a quick test during training
ni = 4
num_sample = ni * ni
idx = np.random.choice(len(train_lr_imgs), num_sample, replace=False)
sample_imgs_lr = tl.prepro.threading_data(
[img for i, img in enumerate(train_lr_imgs) if i in idx],
fn=crop_sub_imgs_fn,
size=(96, 96),
is_random=False)
sample_imgs_hr = tl.prepro.threading_data(
[img for i, img in enumerate(train_hr_imgs) if i in idx],
fn=crop_sub_imgs_fn,
size=(192, 192),
is_random=False)
print('sample LR sub-image:', sample_imgs_lr.shape, sample_imgs_lr.min(),
sample_imgs_lr.max())
print('sample HR sub-image:', sample_imgs_hr.shape, sample_imgs_hr.min(),
sample_imgs_hr.max())
## save the images
tl.vis.save_images(sample_imgs_lr, [ni, ni],
os.path.join(save_dir_ginit, '_train_sample_96.jpg'))
tl.vis.save_images(sample_imgs_hr, [ni, ni],
os.path.join(save_dir_ginit, '_train_sample_192.jpg'))
tl.vis.save_images(sample_imgs_lr, [ni, ni],
os.path.join(save_dir_gan, '_train_sample_96.jpg'))
tl.vis.save_images(sample_imgs_hr, [ni, ni],
os.path.join(save_dir_gan, '_train_sample_192.jpg'))
print('finish saving sample images')
###====== initialize G ======###
## fixed learning rate
sess.run(tf.assign(lr_v, lr_init))
print(" ** fixed learning rate: %f (for init G)" % lr_init)
for epoch in range(0, n_epoch_init + 1):
epoch_time = time.time()
total_mse_loss, total_psnr, n_iter = 0, 0, 0
# random shuffle the train set for each epoch
random.shuffle(train_hr_imgs)
for idx in range(0, len(train_lr_imgs), batch_size):
step_time = time.time()
b_imgs_192 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
size=(192, 192),
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_192,
fn=downsample_fn,
size=(96, 96))
## update G
errM, metricP, _ = sess.run([mse_loss, psnr_metric, g_optim_init],
{
lr_image: b_imgs_96,
hr_image: b_imgs_192
})
print("Epoch [%2d/%2d] %4d time: %4.2fs, mse: %.4f, psnr: %.4f " %
(epoch, n_epoch_init, n_iter, time.time() - step_time, errM,
metricP.mean()))
total_mse_loss += errM
total_psnr += metricP.mean()
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.2fs, mse: %.4f, psnr: %.4f" % (
epoch, n_epoch_init, time.time() - epoch_time,
total_mse_loss / n_iter, total_psnr / n_iter)
print(log)
if validation:
b_imgs_192_V = tl.prepro.threading_data(valid_hr_imgs,
fn=crop_sub_imgs_fn,
size=(192, 192),
is_random=True)
b_imgs_96_V = tl.prepro.threading_data(b_imgs_192_V,
fn=downsample_fn,
size=(96, 96))
errM_V, metricP_V, _ = sess.run(
[mse_loss, psnr_metric, g_optim_init], {
lr_image: b_imgs_96_V,
hr_image: b_imgs_192_V
})
print("Validation | mse: %.4f, psnr: %.4f" %
(errM_V, metricP_V.mean()))
## quick evaluation on train set
if (epoch != 0) and (epoch % save_every_epoch == 0):
out = sess.run(net_g_test.outputs, {lr_image: sample_imgs_lr})
print("[*] save sample images")
tl.vis.save_images(
out, [ni, ni],
os.path.join(save_dir_ginit, 'train_{}.jpg'.format(epoch)))
## save model
if (epoch != 0) and (epoch % save_every_epoch == 0):
tl.files.save_npz(net_g.all_params,
name=os.path.join(checkpoint_dir,
'g_srgan_init.npz'),
sess=sess)
###========================= train GAN (SRGAN) =========================###
## Learning rate decay
decay_every = int(n_epoch / 2) if int(n_epoch / 2) > 0 else 1
for epoch in range(0, n_epoch + 1):
# random shuffle the train set for each epoch
random.shuffle(train_hr_imgs)
## update learning rate
if epoch != 0 and (epoch % decay_every == 0):
new_lr_decay = lr_decay**(epoch // decay_every)
sess.run(tf.assign(lr_v, lr_init * new_lr_decay))
log = " ** new learning rate: %f (for GAN)" % (lr_init *
new_lr_decay)
print(log)
elif epoch == 0:
sess.run(tf.assign(lr_v, lr_init))
log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (
lr_init, decay_every, lr_decay)
print(log)
epoch_time = time.time()
total_d_loss, total_g_loss, total_mse_loss, total_psnr, total_acc, n_iter = 0, 0, 0, 0, 0, 0
for idx in range(0, len(train_lr_imgs), batch_size):
step_time = time.time()
b_imgs_192 = tl.prepro.threading_data(train_hr_imgs[idx:idx +
batch_size],
fn=crop_sub_imgs_fn,
size=(192, 192),
is_random=True)
b_imgs_96 = tl.prepro.threading_data(b_imgs_192,
fn=downsample_fn,
size=(96, 96))
## update D
errD, metricA, _ = sess.run([d_loss, acc_metric, d_optim], {
lr_image: b_imgs_96,
hr_image: b_imgs_192
})
## update G
errG, errM, metricP, _ = sess.run(
[g_loss, mse_loss, psnr_metric, g_optim], {
lr_image: b_imgs_96,
hr_image: b_imgs_192
})
print(
"Epoch [%2d/%2d] %4d time: %4.2fs, d_loss: %.4f g_loss: %.4f (mse: %.4f, psnr: %.4f, accuracy: %.4f)"
% (epoch, n_epoch, n_iter, time.time() - step_time, errD, errG,
errM, metricP.mean(), metricA / 2))
total_d_loss += errD
total_g_loss += errG
total_mse_loss += errM
total_psnr += metricP.mean()
total_acc += metricA / 2
n_iter += 1
log = "[*] Epoch: [%2d/%2d] time: %4.2fs, d_loss: %.4f g_loss: %.4f (mse: %4f, psnr: %.4f, accuracy: %.4f)" % (
epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,
total_g_loss / n_iter, total_mse_loss / n_iter,
total_psnr / n_iter, total_acc / n_iter)
print(log)
if validation:
b_imgs_192_V = tl.prepro.threading_data(valid_hr_imgs,
fn=crop_sub_imgs_fn,
size=(192, 192),
is_random=True)
b_imgs_96_V = tl.prepro.threading_data(b_imgs_192_V,
fn=downsample_fn,
size=(96, 96))
errM_V, metricP_V, _ = sess.run([mse_loss, psnr_metric, g_optim], {
lr_image: b_imgs_96_V,
hr_image: b_imgs_192_V
})
print("Validation | mse: %.4f, psnr: %.4f" %
(errM_V, metricP_V.mean()))
## quick evaluation on train set
if (epoch != 0) and (epoch % save_every_epoch == 0):
out = sess.run(net_g_test.outputs, {lr_image: sample_imgs_lr})
print("[*] save images")
tl.vis.save_images(
out, [ni, ni],
os.path.join(save_dir_gan, 'train_{}.jpg'.format(epoch)))
## save model
if (epoch != 0) and (epoch % save_every_epoch == 0):
tl.files.save_npz(net_g.all_params,
name=os.path.join(checkpoint_dir, 'g_srgan.npz'),
sess=sess)
tl.files.save_npz(net_d.all_params,
name=os.path.join(checkpoint_dir, 'd_srgan.npz'),
sess=sess)