def train_setup(self):
tf.set_random_seed(self.conf.random_seed)
# Create queue coordinator.
self.coord = tf.train.Coordinator()
# Input size
h, w = (self.conf.input_height, self.conf.input_width)
input_size = (h, w)
# Devices
gpu_list = get_available_gpus()
zip_encoder, zip_decoder_b, zip_decoder_w = [], [], []
restore_vars = []
self.loaders = []
self.im_list = []
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
# Load reader
with tf.name_scope("create_inputs"):
reader = ImageReader(self.conf.data_dir,
self.conf.data_list, input_size,
self.conf.random_scale,
self.conf.random_mirror,
self.conf.ignore_label, IMG_MEAN,
self.coord)
self.image_batch, self.label_batch, names = reader.dequeue(
self.conf.batch_size)
self.im_list.append(self.image_batch)
image_batch_075 = tf.image.resize_images(
self.image_batch,
[int(h * 0.75), int(w * 0.75)])
image_batch_05 = tf.image.resize_images(
self.image_batch,
[int(h * 0.5), int(w * 0.5)])
# Create network
with tf.variable_scope('', reuse=False):
net = Deeplab_v2(self.image_batch, self.conf.num_classes,
True)
with tf.variable_scope('', reuse=True):
net075 = Deeplab_v2(image_batch_075, self.conf.num_classes,
True)
with tf.variable_scope('', reuse=True):
net05 = Deeplab_v2(image_batch_05, self.conf.num_classes,
True)
# Variables that load from pre-trained model.
restore_var = [
v for v in tf.global_variables() if 'fc' not in v.name
]
restore_vars.append(restore_var)
# Trainable Variables
all_trainable = tf.trainable_variables()
# Fine-tune part
encoder_trainable = [
v for v in all_trainable if 'fc' not in v.name
] # lr * 1.0
# Decoder part
decoder_trainable = [
v for v in all_trainable if 'fc' in v.name
]
decoder_w_trainable = [
v for v in decoder_trainable
if 'weights' in v.name or 'gamma' in v.name
] # lr * 10.0
decoder_b_trainable = [
v for v in decoder_trainable
if 'biases' in v.name or 'beta' in v.name
] # lr * 20.0
# Check
assert (len(all_trainable) == len(decoder_trainable) +
len(encoder_trainable))
assert (len(decoder_trainable) == len(decoder_w_trainable) +
len(decoder_b_trainable))
# Network raw output
raw_output100 = net.outputs
raw_output075 = net075.outputs
raw_output05 = net05.outputs
raw_output = tf.reduce_max(tf.stack([
raw_output100,
tf.image.resize_images(raw_output075,
tf.shape(raw_output100)[1:3, ]),
tf.image.resize_images(raw_output05,
tf.shape(raw_output100)[1:3, ])
]),
axis=0)
# Groud Truth: ignoring all labels greater or equal than n_classes
label_proc = prepare_label(self.label_batch,
tf.stack(
raw_output.get_shape()[1:3]),
num_classes=self.conf.num_classes,
one_hot=False) # [batch_size, h, w]
label_proc075 = prepare_label(
self.label_batch,
tf.stack(raw_output075.get_shape()[1:3]),
num_classes=self.conf.num_classes,
one_hot=False)
label_proc05 = prepare_label(
self.label_batch,
tf.stack(raw_output05.get_shape()[1:3]),
num_classes=self.conf.num_classes,
one_hot=False)
raw_gt = tf.reshape(label_proc, [
-1,
])
raw_gt075 = tf.reshape(label_proc075, [
-1,
])
raw_gt05 = tf.reshape(label_proc05, [
-1,
])
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, self.conf.num_classes - 1)),
1)
indices075 = tf.squeeze(
tf.where(
tf.less_equal(raw_gt075, self.conf.num_classes - 1)),
1)
indices05 = tf.squeeze(
tf.where(tf.less_equal(raw_gt05,
self.conf.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
gt075 = tf.cast(tf.gather(raw_gt075, indices075), tf.int32)
gt05 = tf.cast(tf.gather(raw_gt05, indices05), tf.int32)
raw_prediction = tf.reshape(raw_output,
[-1, self.conf.num_classes])
raw_prediction100 = tf.reshape(raw_output100,
[-1, self.conf.num_classes])
raw_prediction075 = tf.reshape(raw_output075,
[-1, self.conf.num_classes])
raw_prediction05 = tf.reshape(raw_output05,
[-1, self.conf.num_classes])
prediction = tf.gather(raw_prediction, indices)
prediction100 = tf.gather(raw_prediction100, indices)
prediction075 = tf.gather(raw_prediction075, indices075)
prediction05 = tf.gather(raw_prediction05, indices05)
# Pixel-wise softmax_cross_entropy loss
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction, labels=gt)
loss100 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction100, labels=gt)
loss075 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction075, labels=gt075)
loss05 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction05, labels=gt05)
# L2 regularization
l2_losses = [
self.conf.weight_decay * tf.nn.l2_loss(v)
for v in all_trainable if 'weights' in v.name
]
# Loss function
self.reduced_loss = tf.reduce_mean(loss) + tf.reduce_mean(
loss100) + tf.reduce_mean(loss075) + tf.reduce_mean(
loss05) + tf.add_n(l2_losses)
# Define optimizers
# 'poly' learning rate
base_lr = tf.constant(self.conf.learning_rate)
self.curr_step = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(
base_lr,
tf.pow((1 - self.curr_step / self.conf.num_steps),
self.conf.power))
# We have several optimizers here in order to handle the different lr_mult
# which is a kind of parameters in Caffe. This controls the actual lr for each
# layer.
opt_encoder = tf.train.MomentumOptimizer(
learning_rate, self.conf.momentum)
opt_decoder_w = tf.train.MomentumOptimizer(
learning_rate * 10.0, self.conf.momentum)
opt_decoder_b = tf.train.MomentumOptimizer(
learning_rate * 20.0, self.conf.momentum)
# Gradient accumulation
# Define a variable to accumulate gradients.
accum_grads = [
tf.Variable(tf.zeros_like(v.initialized_value()),
trainable=False) for v in encoder_trainable +
decoder_w_trainable + decoder_b_trainable
]
# Define an operation to clear the accumulated gradients for next batch.
self.zero_op = [
v.assign(tf.zeros_like(v)) for v in accum_grads
]
# To make sure each layer gets updated by different lr's, we do not use 'minimize' here.
# Instead, we separate the steps compute_grads+update_params.
# Compute grads
grads = tf.gradients(
self.reduced_loss, encoder_trainable +
decoder_w_trainable + decoder_b_trainable)
# Accumulate and normalise the gradients.
self.accum_grads_op = [
accum_grads[i].assign_add(grad /
self.conf.grad_update_every)
for i, grad in enumerate(grads)
]
grads = tf.gradients(
self.reduced_loss, encoder_trainable +
decoder_w_trainable + decoder_b_trainable)
grads_encoder = accum_grads[:len(encoder_trainable)]
grads_decoder_w = accum_grads[len(encoder_trainable):(
len(encoder_trainable) + len(decoder_w_trainable))]
grads_decoder_b = accum_grads[(len(encoder_trainable) +
len(decoder_w_trainable)):]
zip_encoder.append(list(zip(grads_encoder, encoder_trainable)))
zip_decoder_b.append(
list(zip(grads_decoder_w, decoder_w_trainable)))
zip_decoder_w.append(
list(zip(grads_decoder_b, decoder_b_trainable)))
avg_grads_encoder = average_gradients(zip_encoder)
avg_grads_decoder_w = average_gradients(zip_decoder_w)
avg_grads_decoder_b = average_gradients(zip_decoder_b)
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
# Update params
train_op_conv = opt_encoder.apply_gradients(avg_grads_encoder)
train_op_fc_w = opt_decoder_w.apply_gradients(
avg_grads_decoder_w)
train_op_fc_b = opt_decoder_b.apply_gradients(
avg_grads_decoder_b)
# Finally, get the train_op!
update_ops = tf.get_collection(
tf.GraphKeys.UPDATE_OPS
) # for collecting moving_mean and moving_variance
with tf.control_dependencies(update_ops):
self.train_op = tf.group(train_op_conv, train_op_fc_w,
train_op_fc_b)
# Saver for storing checkpoints of the model
self.saver = tf.train.Saver(var_list=tf.global_variables(),
max_to_keep=0)
# Loader for loading the pre-trained model
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
#print(restore_var)
#print("restoring gpu ", i)
self.loaders.append(tf.train.Saver(var_list=restore_vars[i]))
#print("restored gpu ", i)
# Training summary
# Processed predictions: for visualisation.
raw_output_up = tf.image.resize_bilinear(raw_output, input_size)
raw_output_up = tf.argmax(raw_output_up, axis=3)
self.pred = tf.expand_dims(raw_output_up, axis=3)
# Image summary.
images_summary = tf.py_func(inv_preprocess,
[self.image_batch, 1, IMG_MEAN], tf.uint8)
labels_summary = tf.py_func(
decode_labels, [self.label_batch, 1, self.conf.num_classes],
tf.uint8)
preds_summary = tf.py_func(decode_labels,
[self.pred, 1, self.conf.num_classes],
tf.uint8)
self.total_summary = tf.summary.image(
'images',
tf.concat(axis=2,
values=[images_summary, labels_summary, preds_summary]),
max_outputs=1) # Concatenate row-wise.
if not os.path.exists(self.conf.logdir):
os.makedirs(self.conf.logdir)
self.summary_writer = tf.summary.FileWriter(
self.conf.logdir, graph=tf.get_default_graph())
def train(args):
## set hyparameter
img_mean = np.array((104.00698793, 116.66876762, 122.67891434),
dtype=np.float32)
tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
LAMBDA = 10
print("d_model_name:", args.d_name)
print("lambda:", args.lamb)
print("learning_rate:", args.learning_rate)
print("is_val:", args.is_val)
print("---------------------------------")
## load data
with tf.name_scope("create_inputs"):
reader = ImageReader(args.data_dir, args.img_size, args.random_scale,
args.random_mirror, args.random_crop,
args.ignore_label, args.is_val, img_mean, coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
print("Data is ready!")
## load model
g_net = choose_generator(args.g_name, image_batch)
score_map = g_net.get_output()
fk_batch = tf.nn.softmax(score_map, dim=-1)
gt_batch = tf.one_hot(label_batch, args.num_classes, dtype=tf.float32)
x_batch = tf.train.batch([
(reader.image + img_mean) / 255.,
],
args.batch_size,
dynamic_pad=True) # normalization
d_fk_net, d_gt_net = choose_discriminator(args.d_name, fk_batch, gt_batch,
x_batch)
d_fk_pred = d_fk_net.get_output() # fake segmentation result in d
d_gt_pred = d_gt_net.get_output() # ground-truth result in d
label, logits = convert_to_calculateloss(score_map, args.num_classes,
label_batch)
predict_label = tf.argmax(logits, axis=1)
predict_batch = g_net.topredict(score_map, tf.shape(image_batch)[1:3])
print("The model has been created!")
## get all kinds of variables list
g_restore_var = [
v for v in tf.global_variables() if 'discriminator' not in v.name
]
vgg_restore_var = [
v for v in tf.global_variables()
if 'discriminator' in v.name and 'image' in v.name
]
g_var = [
v for v in tf.trainable_variables() if 'discriminator' not in v.name
]
d_var = [
v for v in tf.trainable_variables()
if 'discriminator' in v.name and 'image' not in v.name
]
# g_trainable_var = [v for v in g_var if 'beta' not in v.name or 'gamma' not in v.name] #batch_norm training open
g_trainable_var = g_var
d_trainable_var = d_var
## set loss
mce_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=label,
logits=logits))
g_bce_loss = -tf.reduce_mean(d_fk_pred)
g_loss = mce_loss + args.lamb * g_bce_loss
fk_score_var = tf.reduce_mean(d_fk_pred)
gt_score_var = tf.reduce_mean(d_gt_pred)
d_loss = fk_score_var - gt_score_var
alpha = tf.random_uniform(shape=tf.shape(gt_batch), minval=0., maxval=1.)
differences = fk_batch - gt_batch
interpolates = gt_batch + (alpha * differences)
gradients = tf.gradients(
Discriminator_add_vgg({
'seg': interpolates,
'data': x_batch
},
reuse=True).get_output(), [interpolates])[0]
slopes = tf.sqrt(
tf.reduce_sum(tf.square(gradients), reduction_indices=[1, 2, 3]))
gradient_penalty = tf.reduce_mean((slopes - 1.)**2)
d_loss += gradient_penalty
mce_loss_var, mce_loss_op = tf.metrics.mean(mce_loss)
g_bce_loss_var, g_bce_loss_op = tf.metrics.mean(g_bce_loss)
g_loss_var, g_loss_op = tf.metrics.mean(g_loss)
d_loss_var, d_loss_op = tf.metrics.mean(d_loss)
iou_var, iou_op = tf.metrics.mean_iou(label, predict_label,
args.num_classes)
accuracy_var, acc_op = tf.metrics.accuracy(label, predict_label)
metrics_op = tf.group(mce_loss_op, g_bce_loss_op, g_loss_op, d_loss_op,
iou_op, acc_op)
## set optimizer
iterstep = tf.placeholder(dtype=tf.float32,
shape=[],
name='iteration_step')
base_lr = tf.constant(args.learning_rate, dtype=tf.float32, shape=[])
lr = tf.scalar_mul(base_lr,
tf.pow(
(1 - iterstep / args.num_steps),
args.power)) # learning rate reduce with the time
g_gradients = tf.train.AdamOptimizer(learning_rate=lr).compute_gradients(
g_loss, g_trainable_var)
d_gradients = tf.train.AdamOptimizer(
learning_rate=lr * 10).compute_gradients(d_loss, d_trainable_var)
grad_fk_oi = tf.gradients(d_fk_pred, fk_batch, name='grad_fk_oi')[0]
grad_gt_oi = tf.gradients(d_gt_pred, gt_batch, name='grad_gt_oi')[0]
grad_fk_img_oi = tf.gradients(d_fk_pred,
image_batch,
name='grad_fk_img_oi')[0]
grad_gt_img_oi = tf.gradients(d_gt_pred,
image_batch,
name='grad_gt_img_oi')[0]
train_g_op = tf.train.AdamOptimizer(learning_rate=lr).minimize(
g_loss, var_list=g_trainable_var)
train_d_op = tf.train.AdamOptimizer(learning_rate=lr * 10).minimize(
d_loss, var_list=d_trainable_var)
## set summary
vs_image = tf.py_func(inv_preprocess,
[image_batch, args.save_num_images, img_mean],
tf.uint8)
vs_label = tf.py_func(
decode_labels, [label_batch, args.save_num_images, args.num_classes],
tf.uint8)
vs_predict = tf.py_func(
decode_labels, [predict_batch, args.save_num_images, args.num_classes],
tf.uint8)
tf.summary.image(name='image collection_train',
tensor=tf.concat(axis=2,
values=[vs_image, vs_label, vs_predict]),
max_outputs=args.save_num_images)
tf.summary.scalar('fk_score', tf.reduce_mean(d_fk_pred))
tf.summary.scalar('gt_score', tf.reduce_mean(d_gt_pred))
tf.summary.scalar('g_loss_train', g_loss_var)
tf.summary.scalar('d_loss_train', d_loss_var)
tf.summary.scalar('mce_loss_train', mce_loss_var)
tf.summary.scalar('g_bce_loss_train', -1. * g_bce_loss_var)
tf.summary.scalar('iou_train', iou_var)
tf.summary.scalar('accuracy_train', accuracy_var)
tf.summary.scalar('grad_fk_oi', tf.reduce_mean(tf.abs(grad_fk_oi)))
tf.summary.scalar('grad_gt_oi', tf.reduce_mean(tf.abs(grad_gt_oi)))
tf.summary.scalar('grad_fk_img_oi', tf.reduce_mean(tf.abs(grad_fk_img_oi)))
tf.summary.scalar('grad_gt_img_oi', tf.reduce_mean(tf.abs(grad_gt_img_oi)))
for grad, var in g_gradients + d_gradients:
tf.summary.histogram(var.op.name + "/gradients", grad)
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name + "/values", var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.log_dir,
graph=tf.get_default_graph(),
max_queue=3)
## set session
print("GPU index:" + str(os.environ['CUDA_VISIBLE_DEVICES']))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
global_init = tf.global_variables_initializer()
local_init = tf.local_variables_initializer()
sess.run(global_init)
sess.run(local_init)
## set saver
saver_all = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=2)
trained_step = 0
if os.path.exists(args.restore_from + 'checkpoint'):
trained_step = load_weight(args.restore_from, saver_all, sess)
else:
load_weight(args.baseweight_from['d_vgg'], vgg_restore_var, sess)
saver_g = tf.train.Saver(var_list=g_restore_var, max_to_keep=2)
load_weight(args.baseweight_from['g'], saver_g, sess)
threads = tf.train.start_queue_runners(sess, coord)
print("all setting has been done,training start!")
## start training
def auto_setting_train_steps(mode):
if mode == 0:
return 5, 1
elif mode == 1:
return 1, 5
else:
return 1, 1
d_train_steps = 5
g_train_steps = 1
flags = [0 for i in range(3)]
for step in range(args.num_steps):
now_step = int(
trained_step) + step if trained_step is not None else step
feed_dict = {iterstep: now_step}
for i in range(d_train_steps):
_, _ = sess.run([train_d_op, metrics_op], feed_dict)
for i in range(g_train_steps):
g_loss_, mce_loss_, g_bce_loss_, d_loss_, _, _ = sess.run([
g_loss_var, mce_loss_var, g_bce_loss_var, d_loss_var,
train_g_op, metrics_op
], feed_dict)
########################
fk_score_, gt_score_ = sess.run([fk_score_var, gt_score_var],
feed_dict)
if fk_score_ > 0.48 and fk_score_ < 0.52:
flags[0] += 1
flags[1] = flags[2] = 0
elif gt_score_ - fk_score_ > 0.3:
flags[1] += 1
flags[0] = flags[2] = 0
else:
flags[2] += 1
flags[0] = flags[1] = 0
if max(flags) > 100:
d_train_steps, g_train_steps = auto_setting_train_steps(
flags.index(max(flags)))
########################
if step > 0 and step % args.save_pred_every == 0:
save_weight(args.restore_from, saver_all, sess, now_step)
if step % 50 == 0 or step == args.num_steps - 1:
print('step={} d_loss={} g_loss={} mce_loss={} g_bce_loss_={}'.
format(now_step, d_loss_, g_loss_, mce_loss_, g_bce_loss_))
summary_str = sess.run(summary_op, feed_dict)
summary_writer.add_summary(summary_str, now_step)
sess.run(local_init)
## end training
coord.request_stop()
coord.join(threads)
print('end....')
def train_setup(self):
tf.set_random_seed(self.conf.random_seed)
# Create queue coordinator.
self.coord = tf.train.Coordinator()
# Input size
input_size = (self.conf.input_height, self.conf.input_width)
# Load reader
with tf.name_scope("create_inputs"):
reader = ImageReader(
self.conf.data_dir,
self.conf.data_list,
input_size,
self.conf.random_scale,
self.conf.random_mirror,
self.conf.ignore_label,
IMG_MEAN,
self.coord)
self.image_batch, self.label_batch = reader.dequeue(self.conf.batch_size)
# Create network
if self.conf.encoder_name not in ['res101', 'res50', 'deeplab']:
print('encoder_name ERROR!')
print("Please input: res101, res50, or deeplab")
sys.exit(-1)
elif self.conf.encoder_name == 'deeplab':
net = Deeplab_v2(self.image_batch, self.conf.num_classes, True)
# Variables that load from pre-trained model.
restore_var = [v for v in tf.global_variables() if 'fc' not in v.name]
# Trainable Variables
all_trainable = tf.trainable_variables()
# Fine-tune part
encoder_trainable = [v for v in all_trainable if 'fc' not in v.name] # lr * 1.0
# Decoder part
decoder_trainable = [v for v in all_trainable if 'fc' in v.name]
else:
net = ResNet_segmentation(self.image_batch, self.conf.num_classes, True, self.conf.encoder_name)
# Variables that load from pre-trained model.
restore_var = [v for v in tf.global_variables() if 'resnet_v1' in v.name]
# Trainable Variables
all_trainable = tf.trainable_variables()
# Fine-tune part
encoder_trainable = [v for v in all_trainable if 'resnet_v1' in v.name] # lr * 1.0
# Decoder part
decoder_trainable = [v for v in all_trainable if 'decoder' in v.name]
decoder_w_trainable = [v for v in decoder_trainable if 'weights' in v.name or 'gamma' in v.name] # lr * 10.0
decoder_b_trainable = [v for v in decoder_trainable if 'biases' in v.name or 'beta' in v.name] # lr * 20.0
# Check
assert(len(all_trainable) == len(decoder_trainable) + len(encoder_trainable))
assert(len(decoder_trainable) == len(decoder_w_trainable) + len(decoder_b_trainable))
# Network raw output
raw_output = net.outputs # [batch_size, h, w, 21]
# Output size
output_shape = tf.shape(raw_output)
output_size = (output_shape[1], output_shape[2])
# Groud Truth: ignoring all labels greater or equal than n_classes
label_proc = prepare_label(self.label_batch, output_size, num_classes=self.conf.num_classes, one_hot=False)
raw_gt = tf.reshape(label_proc, [-1,])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, self.conf.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
raw_prediction = tf.reshape(raw_output, [-1, self.conf.num_classes])
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax_cross_entropy loss
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction, labels=gt)
# L2 regularization
l2_losses = [self.conf.weight_decay * tf.nn.l2_loss(v) for v in all_trainable if 'weights' in v.name]
# Loss function
self.reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
# Define optimizers
# 'poly' learning rate
base_lr = tf.constant(self.conf.learning_rate)
self.curr_step = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - self.curr_step / self.conf.num_steps), self.conf.power))
# We have several optimizers here in order to handle the different lr_mult
# which is a kind of parameters in Caffe. This controls the actual lr for each
# layer.
opt_encoder = tf.train.MomentumOptimizer(learning_rate, self.conf.momentum)
opt_decoder_w = tf.train.MomentumOptimizer(learning_rate * 10.0, self.conf.momentum)
opt_decoder_b = tf.train.MomentumOptimizer(learning_rate * 20.0, self.conf.momentum)
# To make sure each layer gets updated by different lr's, we do not use 'minimize' here.
# Instead, we separate the steps compute_grads+update_params.
# Compute grads
grads = tf.gradients(self.reduced_loss, encoder_trainable + decoder_w_trainable + decoder_b_trainable)
grads_encoder = grads[:len(encoder_trainable)]
grads_decoder_w = grads[len(encoder_trainable) : (len(encoder_trainable) + len(decoder_w_trainable))]
grads_decoder_b = grads[(len(encoder_trainable) + len(decoder_w_trainable)):]
# Update params
train_op_conv = opt_encoder.apply_gradients(zip(grads_encoder, encoder_trainable))
train_op_fc_w = opt_decoder_w.apply_gradients(zip(grads_decoder_w, decoder_w_trainable))
train_op_fc_b = opt_decoder_b.apply_gradients(zip(grads_decoder_b, decoder_b_trainable))
# Finally, get the train_op!
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) # for collecting moving_mean and moving_variance
with tf.control_dependencies(update_ops):
self.train_op = tf.group(train_op_conv, train_op_fc_w, train_op_fc_b)
# Saver for storing checkpoints of the model
self.saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=0)
# Loader for loading the pre-trained model
self.loader = tf.train.Saver(var_list=restore_var)
# Training summary
# Processed predictions: for visualisation.
raw_output_up = tf.image.resize_bilinear(raw_output, input_size)
raw_output_up = tf.argmax(raw_output_up, axis=3)
self.pred = tf.expand_dims(raw_output_up, dim=3)
# Image summary.
images_summary = tf.py_func(inv_preprocess, [self.image_batch, 2, IMG_MEAN], tf.uint8)
labels_summary = tf.py_func(decode_labels, [self.label_batch, 2, self.conf.num_classes], tf.uint8)
preds_summary = tf.py_func(decode_labels, [self.pred, 2, self.conf.num_classes], tf.uint8)
self.total_summary = tf.summary.image('images',
tf.concat(axis=2, values=[images_summary, labels_summary, preds_summary]),
max_outputs=2) # Concatenate row-wise.
if not os.path.exists(self.conf.logdir):
os.makedirs(self.conf.logdir)
self.summary_writer = tf.summary.FileWriter(self.conf.logdir, graph=tf.get_default_graph())
def train_setup(self):
tf.set_random_seed(self.conf.random_seed)
# Create queue coordinator.
self.coord = tf.train.Coordinator()
# Input size
h, w = (self.conf.input_height, self.conf.input_width)
input_size = (h, w)
# Devices
gpu_list = get_available_gpus()
zip_encoder, zip_decoder_b, zip_decoder_w, zip_crf = [], [], [], []
previous_crf_names = []
restore_vars = []
self.loaders = []
self.im_list = []
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
# Load reader
with tf.name_scope("create_inputs"):
reader = ImageReader(self.conf.data_dir,
self.conf.data_list, input_size,
self.conf.random_scale,
self.conf.random_mirror,
self.conf.ignore_label, IMG_MEAN,
self.coord)
self.image_batch, self.label_batch, self.sp_batch = reader.dequeue(
self.conf.batch_size)
self.im_list.append(self.image_batch)
image_batch_075 = tf.image.resize_images(
self.image_batch,
[int(h * 0.75), int(w * 0.75)])
image_batch_05 = tf.image.resize_images(
self.image_batch,
[int(h * 0.5), int(w * 0.5)])
sp_batch_075 = tf.image.resize_images(
self.sp_batch,
[int(h * 0.75), int(w * 0.75)])
sp_batch_05 = tf.image.resize_images(
self.sp_batch,
[int(h * 0.5), int(w * 0.5)])
#for i in range(1):
# self.image_batch = tf.Print(self.image_batch, [self.image_batch[i]], message = 'image batch ', summarize=5)
#for i in range(1):
# self.label_batch = tf.Print(self.label_batch, [self.label_batch[i]], message = 'label batch ', summarize=5)
#for i in range(1):
# self.sp_batch = tf.Print(self.sp_batch, [self.sp_batch[i]], message = 'sp batch ', summarize=5)
# Create network
with tf.variable_scope('', reuse=False):
if self.conf.crf_type == 'crf':
net = Deeplab_v2(self.image_batch,
self.conf.num_classes,
True,
rescale075=False,
rescale05=False,
crf_type=self.conf.crf_type)
else:
net = Deeplab_v2(self.image_batch,
self.conf.num_classes,
True,
rescale075=False,
rescale05=False,
crf_type=self.conf.crf_type,
superpixels=self.sp_batch)
'''
with tf.variable_scope('', reuse=True):
if self.conf.crf_type == 'crfSP':
net075 = Deeplab_v2(image_batch_075, self.conf.num_classes, True, rescale075=True, rescale05=False, crf_type = self.conf.crf_type, superpixels=sp_batch_075)
else:
net075 = Deeplab_v2(image_batch_075, self.conf.num_classes, True, rescale075=True, rescale05=False, crf_type = self.conf.crf_type)
with tf.variable_scope('', reuse=True):
if self.conf.crf_type == 'crfSP':
net05 = Deeplab_v2(image_batch_05, self.conf.num_classes, True, rescale075=False, rescale05=True, crf_type = self.conf.crf_type, superpixels=sp_batch_05)
else:
net05 = Deeplab_v2(image_batch_05, self.conf.num_classes, True, rescale075=False, rescale05=True, crf_type = self.conf.crf_type)
'''
# Variables that load from pre-trained model.
restore_var = [
v for v in tf.global_variables()
if ('fc' not in v.name and 'crfrnn' not in v.name)
] # when don't want to train using previous crf weights
#restore_var = [v for v in tf.global_variables() if ('fc' not in v.name and 'superpixel' not in v.name)]
restore_vars.append(restore_var)
# Trainable Variables
all_trainable = tf.trainable_variables()
# Fine-tune part
for name in previous_crf_names:
for v in all_trainable:
if v.name == name:
all_trainable.remove(v)
crf_trainable = [
v for v in all_trainable
if ('crfrnn' in v.name and v.name not in previous_crf_names
)
]
previous_crf_names.extend(v.name for v in crf_trainable)
encoder_trainable = [
v for v in all_trainable
if 'fc' not in v.name and 'crfrnn' not in v.name
] # lr * 1.0
# Remove encoder_trainable from all_trainable
#all_trainable = [v for v in all_trainable if v not in encoder_trainable]
# Decoder part
decoder_trainable = [
v for v in all_trainable
if 'fc' in v.name and 'crfrnn' not in v.name
]
decoder_w_trainable = [
v for v in decoder_trainable
if ('weights' in v.name or 'gamma' in v.name)
and 'crfrnn' not in v.name
] # lr * 10.0
decoder_b_trainable = [
v for v in decoder_trainable
if ('biases' in v.name or 'beta' in v.name)
and 'crfrnn' not in v.name
] # lr * 20.0
# Check
assert (len(all_trainable) == len(encoder_trainable) +
len(decoder_trainable) + len(crf_trainable)
) #+ len(encoder_trainable)
assert (len(decoder_trainable) == len(decoder_w_trainable) +
len(decoder_b_trainable))
# Network raw output
raw_output100 = net.outputs
raw_output = raw_output100
'''
raw_output075 = net075.outputs
raw_output05 = net05.outputs
raw_output = tf.reduce_max(tf.stack([raw_output100,
tf.image.resize_images(raw_output075, tf.shape(raw_output100)[1:3,]),
tf.image.resize_images(raw_output05, tf.shape(raw_output100)[1:3,])]), axis=0)
'''
# Ground Truth: ignoring all labels greater or equal than n_classes
label_proc = prepare_label(self.label_batch,
tf.stack(
raw_output.get_shape()[1:3]),
num_classes=self.conf.num_classes,
one_hot=True) # [batch_size, h, w]
'''
label_proc075 = prepare_label(self.label_batch, tf.stack(raw_output075.get_shape()[1:3]), num_classes=self.conf.num_classes, one_hot=True)
label_proc05 = prepare_label(self.label_batch, tf.stack(raw_output05.get_shape()[1:3]), num_classes=self.conf.num_classes, one_hot=True)
'''
raw_gt = tf.reshape(label_proc, [
-1,
])
'''
raw_gt075 = tf.reshape(label_proc075, [-1,])
raw_gt05 = tf.reshape(label_proc05, [-1,])
'''
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, self.conf.num_classes - 1)),
1)
'''
indices075 = tf.squeeze(tf.where(tf.less_equal(raw_gt075, self.conf.num_classes - 1)), 1)
indices05 = tf.squeeze(tf.where(tf.less_equal(raw_gt05, self.conf.num_classes - 1)), 1)
'''
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
'''
gt075 = tf.cast(tf.gather(raw_gt075, indices075), tf.int32)
gt05 = tf.cast(tf.gather(raw_gt05, indices05), tf.int32)
'''
raw_prediction = tf.reshape(raw_output,
[-1, self.conf.num_classes])
raw_prediction100 = tf.reshape(raw_output100,
[-1, self.conf.num_classes])
'''
raw_prediction075 = tf.reshape(raw_output075, [-1, self.conf.num_classes])
raw_prediction05 = tf.reshape(raw_output05, [-1, self.conf.num_classes])
'''
prediction = tf.gather(raw_prediction, indices)
prediction100 = tf.gather(raw_prediction100, indices)
'''
prediction075 = tf.gather(raw_prediction075, indices075)
prediction05 = tf.gather(raw_prediction05, indices05)
'''
# Pixel-wise softmax_cross_entropy loss
#loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction, labels=gt)
loss = tf.nn.softmax_cross_entropy_with_logits(
logits=raw_prediction,
labels=tf.reshape(label_proc[0],
(h * w, self.conf.num_classes)))
# NOTE used to be loss=tf.nn.softmax_cross_entropy_with_logits_v2
'''
coefficients = [0.01460247, 1.25147725, 2.88479363, 1.20348121, 1.65261654, 1.67514772,
0.62338799, 0.7729363, 0.42038501, 0.98557268, 1.31867536, 0.85313332,
0.67227604, 1.21317965, 1. , 0.24263748, 1.80877607, 1.3082213,
0.79664027, 0.72543945, 1.27823374]
'''
#loss = weighted_loss(self.conf.num_classes, coefficients, labels=tf.reshape(label_proc[0], (h*w, self.conf.num_classes)), logits=raw_prediction)
#loss100 = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction100, labels=gt)
loss100 = tf.nn.softmax_cross_entropy_with_logits(
logits=raw_prediction100,
labels=tf.reshape(label_proc[0],
(h * w, self.conf.num_classes)))
# NOTE used to be loss=tf.nn.softmax_cross_entropy_with_logits_v2
#loss100 = weighted_loss(self.conf.num_classes, coefficients, labels=tf.reshape(label_proc[0], (h*w, self.conf.num_classes)), logits=raw_prediction100)
#loss075 = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction075, labels=gt075)
#loss075 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=raw_prediction075, labels=tf.reshape(label_proc075[0], (int(h * 0.75) * int(w * 0.75), self.conf.num_classes)))
#loss075 = weighted_loss(self.conf.num_classes, coefficients, labels=tf.reshape(label_proc075[0], (int(h * 0.75) * int(w * 0.75), self.conf.num_classes)), logits=raw_prediction075)
#loss05 = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction05, labels=gt05)
#loss05 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=raw_prediction05, labels=tf.reshape(label_proc05[0], (int(h * 0.5) * int(w * 0.5), self.conf.num_classes)))
#loss05 = weighted_loss(self.conf.num_classes, coefficients, labels=tf.reshape(label_proc05[0], (int(h * 0.5) * int(w * 0.5), self.conf.num_classes)), logits=raw_prediction05)
# L2 regularization
l2_losses = [
self.conf.weight_decay * tf.nn.l2_loss(v)
for v in all_trainable if 'weights' in v.name
]
# Loss function
self.reduced_loss = tf.reduce_mean(loss) + tf.reduce_mean(
loss100
) #+ tf.reduce_mean(loss075) + tf.reduce_mean(loss05) + tf.add_n(l2_losses)
# Define optimizers
# 'poly' learning rate
base_lr = tf.constant(self.conf.learning_rate)
self.curr_step = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(
base_lr,
tf.pow((1 - self.curr_step / self.conf.num_steps),
self.conf.power))
# We have several optimizers here in order to handle the different lr_mult
# which is a kind of parameters in Caffe. This controls the actual lr for each
# layer.
opt_encoder = tf.train.MomentumOptimizer(
learning_rate, self.conf.momentum)
opt_decoder_w = tf.train.MomentumOptimizer(
learning_rate * 10.0, self.conf.momentum)
opt_decoder_b = tf.train.MomentumOptimizer(
learning_rate * 20.0, self.conf.momentum)
opt_crf = tf.train.MomentumOptimizer(learning_rate,
self.conf.momentum)
# Gradient accumulation
# Define a variable to accumulate gradients.
accum_grads = [
tf.Variable(tf.zeros_like(v.initialized_value()),
trainable=False) for v in encoder_trainable +
decoder_w_trainable + decoder_b_trainable + crf_trainable
] #encoder_trainable +
# Define an operation to clear the accumulated gradients for next batch.
self.zero_op = [
v.assign(tf.zeros_like(v)) for v in accum_grads
]
# To make sure each layer gets updated by different lr's, we do not use 'minimize' here.
# Instead, we separate the steps compute_grads+update_params.
# Compute grads
grads = tf.gradients(self.reduced_loss,
encoder_trainable + decoder_w_trainable +
decoder_b_trainable +
crf_trainable) #encoder_trainable +
# Accumulate and normalise the gradients.
self.accum_grads_op = [
accum_grads[i].assign_add(grad /
self.conf.grad_update_every)
for i, grad in enumerate(grads)
]
#'''
grads_encoder = accum_grads[:len(encoder_trainable)]
grads_decoder_w = accum_grads[len(encoder_trainable
):len(encoder_trainable) +
len(decoder_w_trainable)]
grads_decoder_b = accum_grads[(
len(encoder_trainable) +
len(decoder_w_trainable)):(len(encoder_trainable) +
len(decoder_w_trainable) +
len(decoder_b_trainable))]
grads_crf = accum_grads[
len(encoder_trainable) + len(decoder_w_trainable) +
len(decoder_b_trainable
):] # assuming crf gradients are appended to the end
#'''
'''
grads_decoder_w = accum_grads[: len(decoder_w_trainable)]
grads_decoder_b = accum_grads[(len(decoder_w_trainable)):(len(decoder_w_trainable)+len(decoder_b_trainable))]
grads_crf = accum_grads[len(decoder_w_trainable)+len(decoder_b_trainable):] # assuming crf gradients are appended to the end
'''
zip_encoder.append(list(zip(grads_encoder, encoder_trainable)))
zip_decoder_b.append(
list(zip(grads_decoder_b, decoder_b_trainable)))
zip_decoder_w.append(
list(zip(grads_decoder_w, decoder_w_trainable)))
zip_crf.append(list(zip(grads_crf, crf_trainable)))
avg_grads_encoder = average_gradients(zip_encoder)
avg_grads_decoder_w = average_gradients(zip_decoder_w)
avg_grads_decoder_b = average_gradients(zip_decoder_b)
avg_grads_crf = average_gradients(zip_crf)
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
# Update params
train_op_conv = opt_encoder.apply_gradients(avg_grads_encoder)
train_op_fc_w = opt_decoder_w.apply_gradients(
avg_grads_decoder_w)
train_op_fc_b = opt_decoder_b.apply_gradients(
avg_grads_decoder_b)
train_op_crf = opt_crf.apply_gradients(avg_grads_crf)
# Finally, get the train_op!
update_ops = tf.get_collection(
tf.GraphKeys.UPDATE_OPS
) # for collecting moving_mean and moving_variance
with tf.control_dependencies(update_ops):
self.train_op = tf.group(train_op_conv, train_op_fc_w,
train_op_fc_b,
train_op_crf) # train_op_conv
# Saver for storing checkpoints of the model
self.saver = tf.train.Saver(var_list=tf.global_variables(),
max_to_keep=0)
# Loader for loading the pre-trained model
for i in range(len(gpu_list)):
with tf.device(gpu_list[i]):
self.loaders.append(tf.train.Saver(var_list=restore_vars[i]))
#self.loaders.append(tf.train.Saver(var_list=tf.global_variables()))
# Training summary
# Processed predictions: for visualisation.
raw_output_up = tf.image.resize_bilinear(raw_output, input_size)
raw_output_up = tf.argmax(raw_output_up, axis=3)
self.pred = tf.expand_dims(raw_output_up, axis=3)
# Image summary.
images_summary = tf.py_func(inv_preprocess,
[self.image_batch, 1, IMG_MEAN], tf.uint8)
labels_summary = tf.py_func(
decode_labels, [self.label_batch, 1, self.conf.num_classes],
tf.uint8)
preds_summary = tf.py_func(decode_labels,
[self.pred, 1, self.conf.num_classes],
tf.uint8)
self.total_summary = tf.summary.image(
'images',
tf.concat(axis=2,
values=[images_summary, labels_summary, preds_summary]),
max_outputs=1) # Concatenate row-wise.
if not os.path.exists(self.conf.logdir):
os.makedirs(self.conf.logdir)
self.summary_writer = tf.summary.FileWriter(
self.conf.logdir, graph=tf.get_default_graph())
random_scale = False
random_mirror = True
random_crop = True
batch_size = 8
learning_rate = 0.00001
power = 0.9
num_steps = 300000
restore_from = './weights/dvn/20171119/'
g_weight_from = ''
d_weight_from = ''
data_dir = '/data/rui.wu/irfan/gan_seg/dvn/data/'
is_train = True
with tf.name_scope("create_inputs"):
reader = ImageReader(data_dir, img_size, crop_size, random_scale,
random_mirror, random_crop, is_train, coord)
image_batch, label_batch = reader.dequeue(batch_size)
print("Data is ready!")
## load model
label_batch = tf.cast(label_batch, tf.uint8)
image_batch = tf.cast(image_batch, tf.float32)
# b = tf.zeros(label_batch.get_shape())
# a = tf.ones(label_batch.get_shape())
# label_batch_b = tf.where(tf.greater(label_batch, 0.5), a, b)
real_iou = tf.placeholder(tf.float32, [batch_size, 1])
train_seg = tf.placeholder(tf.float32, [batch_size, 128, 128, 1])
train_image = tf.placeholder(tf.float32, [batch_size, 128, 128, 3])
train_seg_new = tf.cast(train_seg, tf.uint8)
train_seg_new = tf.squeeze(train_seg_new, squeeze_dims=[3])
train_seg_new = tf.one_hot(train_seg_new, 2)
def build(self):
config = self.__dict__.copy()
num_labels = self.num_labels #for segmentation (pixel labels)
ignore_label = 255 #for segmentation (pixel labels)
random_seed = 1234
generator = self.resnetG
discriminator = self.resnetD
GEN_A2B_NAME = 'GEN_A2B'
GEN_B2A_NAME = 'GEN_B2A'
DIS_A_NAME = 'DIS_A'
DIS_B_NAME = 'DIS_B'
global_step = tf.train.get_or_create_global_step()
slim.add_model_variable(global_step)
global_step_update = tf.assign_add(global_step, 1, name='global_step_update')
def resize_and_onehot(tensor, shape, depth):
with tf.device('/device:CPU:0'):
onehot_tensor = tf.one_hot(tf.squeeze(
tf.image.resize_nearest_neighbor(
tf.cast(tensor, tf.int32), shape), -1), depth=depth)
return onehot_tensor
def convert_to_labels(onehot_seg, crop_size=None):
fake_segments_output = onehot_seg
print ('%s | ' % fake_segments_output.device, fake_segments_output)
if crop_size:
fake_segments_output = tf.image.resize_bilinear(fake_segments_output, crop_size) #tf.shape(source_segments_batch)[1:3])
fake_segments_output = tf.argmax(fake_segments_output, axis=-1) # generate segment indices matrix
fake_segments_output = tf.expand_dims(fake_segments_output, dim=-1) # Create 4-d tensor.
return fake_segments_output
target_data_queue = []
tf.set_random_seed(random_seed)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
for i, data in enumerate([config['source_data']] + config['target_data']):
reader = ImageReader(
data['data_dir'],
data['data_list'],
config['crop_size'], # Original size: [1024, 2048]
random_scale=config['random_scale'],
random_mirror=True,
ignore_label=ignore_label,
img_mean=0, # set IMG_MEAN to centralize image pixels (set NONE for automatic choosing)
img_channel_format='RGB', # Default: BGR in deeplab_v2. See here: https://github.com/zhengyang-wang/Deeplab-v2--ResNet-101--Tensorflow/issues/30
coord=coord,
rgb_label=False)
data_queue = reader.dequeue(config['batch_size'])
if i == 0:
# ---[ source: training data
source_images_batch = data_queue[0] #A: 3 chaanels
source_segments_batch = data_queue[1] #B: 1-label channels
source_images_batch = tf.cast(source_images_batch, tf.float32) / 127.5 - 1.
source_images_batch = tf.image.resize_bilinear(source_images_batch, config['resize']) #A: 3 chaanels
source_segments_batch = tf.image.resize_nearest_neighbor(source_segments_batch, config['resize']) #B: 1-label channels
source_segments_batch = tf.cast(tf.one_hot(tf.squeeze(source_segments_batch, -1), depth=num_labels), tf.float32) - 0.5 #B: 19 channels
else:
# ---[ target: validation data / testing data
target_images_batch = data_queue[0] #A: 3 chaanels
target_segments_batch = data_queue[1] #B: 1-label channels
target_images_batch = tf.cast(target_images_batch, tf.float32) / 127.5 - 1.
target_images_batch = tf.image.resize_bilinear(target_images_batch, config['resize']) #A: 3 chaanels
target_segments_batch = tf.image.resize_nearest_neighbor(target_segments_batch, config['resize']) #B: 1-label channels
target_segments_batch = tf.cast(tf.one_hot(tf.squeeze(target_segments_batch, -1), depth=num_labels), tf.float32) - 0.5 #B: 19 channels
target_data_queue.append([target_images_batch, target_segments_batch])
size_list = cuttool(config['batch_size'], config['gpus'])
source_images_batches = tf.split(source_images_batch, size_list)
source_segments_batches = tf.split(source_segments_batch, size_list)
fake_1_segments_output = [None] * len(size_list)
fake_2_segments_output = [None] * len(size_list)
fake_1_images_output = [None] * len(size_list)
fake_2_images_output = [None] * len(size_list)
d_real_img_output = [None] * len(size_list)
d_fake_img_output = [None] * len(size_list)
d_real_seg_output = [None] * len(size_list)
d_fake_seg_output = [None] * len(size_list)
for gid, (source_images_batch, source_segments_batch) in \
enumerate(zip(source_images_batches, source_segments_batches)):
# ---[ Generator A2B & B2A
with tf.device('/device:GPU:{}'.format((gid-1) % config['gpus'])):
fake_seg = generator(source_images_batch, output_channel=num_labels, reuse=tf.AUTO_REUSE, phase_train=True, scope=GEN_A2B_NAME)
fake_seg = tf.nn.softmax(fake_seg) - 0.5
fake_img_ = generator(fake_seg, output_channel=3, reuse=tf.AUTO_REUSE, phase_train=True, scope=GEN_B2A_NAME)
fake_img_ = tf.nn.tanh(fake_img_)
fake_img = generator(source_segments_batch, output_channel=3, reuse=tf.AUTO_REUSE, phase_train=True, scope=GEN_B2A_NAME)
fake_img = tf.nn.tanh(fake_img)
fake_seg_ = generator(fake_img, output_channel=num_labels, reuse=tf.AUTO_REUSE, phase_train=True, scope=GEN_A2B_NAME)
fake_seg_ = tf.nn.softmax(fake_seg_) - 0.5
# ---[ Discriminator A & B
with tf.device('/device:GPU:{}'.format((gid-1) % config['gpus'])):
d_real_img = discriminator(source_images_batch, reuse=tf.AUTO_REUSE, phase_train=True, scope=DIS_A_NAME)
d_fake_img = discriminator(fake_img, reuse=tf.AUTO_REUSE, phase_train=True, scope=DIS_A_NAME)
d_real_seg = discriminator(source_segments_batch, reuse=tf.AUTO_REUSE, phase_train=True, scope=DIS_B_NAME)
d_fake_seg = discriminator(fake_seg, reuse=tf.AUTO_REUSE, phase_train=True, scope=DIS_B_NAME)
#d_fake_img_val = discriminator(fake_img_val, reuse=tf.AUTO_REUSE, phase_train=False, scope=DIS_A_NAME)
#d_fake_seg_val = discriminator(fake_seg_val, reuse=tf.AUTO_REUSE, phase_train=False, scope=DIS_B_NAME)
fake_1_segments_output [gid] = fake_seg
fake_2_segments_output [gid] = fake_seg_
fake_1_images_output [gid] = fake_img
fake_2_images_output [gid] = fake_img_
d_real_img_output [gid] = d_real_img
d_fake_img_output [gid] = d_fake_img
d_real_seg_output [gid] = d_real_seg
d_fake_seg_output [gid] = d_fake_seg
source_images_batch = tf.concat(source_images_batches, axis=0) #-1~1
source_segments_batch = tf.concat(source_segments_batches, axis=0) #onehot: -0.5~+0.5
fake_1_segments_output = tf.concat(fake_1_segments_output, axis=0) ; print('fake_1_segments_output', fake_1_segments_output)
fake_2_segments_output = tf.concat(fake_2_segments_output, axis=0) ; print('fake_2_segments_output', fake_2_segments_output)
fake_1_images_output = tf.concat(fake_1_images_output , axis=0) ; print('fake_1_images_output ', fake_1_images_output )
fake_2_images_output = tf.concat(fake_2_images_output , axis=0) ; print('fake_2_images_output ', fake_2_images_output )
d_real_img_output = tf.concat(d_real_img_output , axis=0)
d_fake_img_output = tf.concat(d_fake_img_output , axis=0)
d_real_seg_output = tf.concat(d_real_seg_output , axis=0)
d_fake_seg_output = tf.concat(d_fake_seg_output , axis=0)
source_data_color = [
(1.+source_images_batch ) / 2. , # source_images_batch_color
sgtools.decode_labels(tf.cast(convert_to_labels(source_segments_batch + 0.5), tf.int32), num_labels), # source_segments_batch_colo
sgtools.decode_labels(tf.cast(convert_to_labels(fake_1_segments_output + 0.5), tf.int32), num_labels), # fake_1_segments_output_col
sgtools.decode_labels(tf.cast(convert_to_labels(fake_2_segments_output + 0.5), tf.int32), num_labels), # fake_2_segments_output_col
(1.+fake_1_images_output ) / 2. , # fake_1_images_output_color
(1.+fake_2_images_output ) / 2. , # fake_2_images_output_color
]
# ---[ Validation Model
target_data_color_queue = []
for target_data in target_data_queue:
with tf.device('/device:GPU:{}'.format((2) % config['gpus'])):
fake_seg = generator(val_images_holder, output_channel=num_labels, reuse=tf.AUTO_REUSE, phase_train=False, scope=GEN_A2B_NAME)
fake_seg = tf.nn.softmax(fake_seg) - 0.5
fake_img_ = generator(fake_seg, output_channel=3, reuse=tf.AUTO_REUSE, phase_train=False, scope=GEN_B2A_NAME)
fake_img_ = tf.nn.tanh(fake_img_)
fake_img = generator(val_segments_holder, output_channel=3, reuse=tf.AUTO_REUSE, phase_train=False, scope=GEN_B2A_NAME)
fake_img = tf.nn.tanh(fake_img)
fake_seg_ = generator(fake_img, output_channel=num_labels, reuse=tf.AUTO_REUSE, phase_train=False, scope=GEN_A2B_NAME)
fake_seg_ = tf.nn.softmax(fake_seg) - 0.5
target_data_color_queue.append([
(1.+target_images_batch ) / 2. , # target_images_batch_color
sgtools.decode_labels(tf.cast(convert_to_labels(target_segments_batch + 0.5), tf.int32), num_labels) , # target_segments_batch_color
sgtools.decode_labels(tf.cast(convert_to_labels(fake_seg + 0.5), tf.int32), num_labels) , # val_fake_1_segments_output_color
sgtools.decode_labels(tf.cast(convert_to_labels(fake_seg_ + 0.5), tf.int32), num_labels) , # val_fake_2_segments_output_color
(1.+val_fake_1_images_output ) / 2. , # val_fake_1_images_output_color
(1.+val_fake_2_images_output ) / 2. , # val_fake_2_images_output_color
])
# ---[ Segment-level loss: pixelwise loss
# d_seg_batch = tf.image.resize_nearest_neighbor(seg_gt, tf.shape(_d_real['segment'])[1:3])
# d_seg_batch = tf.squeeze(d_seg_batch, -1)
# d_seg_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=d_seg_batch, logits=_d_real['segment'], name='segment_pixelwise_loss') # pixel-wise loss
# d_seg_loss = tf.reduce_mean(d_seg_loss)
# d_seg_loss = tf.identity(d_seg_loss, name='d_seg_loss')
# ---[ GAN Loss: crite loss
#d_loss_old = - (tf.reduce_mean(d_source_output['critic']) - tf.reduce_mean(d_target_output['critic']))
#g_loss = - (tf.reduce_mean(d_target_output['critic']))
## gradient penalty
#LAMBDA = 10
##alpha = tf.placeholder(tf.float32, shape=[None], name='alpha')
#alpha = tf.random_uniform([config['batch_size']], 0.0, 1.0, dtype=tf.float32)
#for _ in source_segments_batch.shape[1:]:
#alpha = tf.expand_dims(alpha, axis=1) #shape=[None,1,1,1]
#interpolates = alpha * source_segments_batch + (1.-alpha) * target_segments_output
#print ('source_segments_batch:', source_segments_batch)
#print ('target_segments_output:',target_segments_output)
#print ('interpolates:', interpolates)
#interpolates = resize_and_onehot(interpolates, target_raw_segments_output.shape.as_list()[1:3], num_labels)
#print ('interpolates:', interpolates)
#_d_intp = discriminator(interpolates, reuse=True, phase_train=True, scope=DIS_NAME)
#intp_grads = tf.gradients(_d_intp['critic'], [interpolates])[0]
#slopes = tf.sqrt(tf.reduce_sum(tf.square(intp_grads), reduction_indices=[1])) #L2-distance
#grads_penalty = tf.reduce_mean(tf.square(slopes-1), name='grads_penalty')
#d_loss = d_loss_old + LAMBDA * grads_penalty
def sigmoid_cross_entropy(labels, logits):
return tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=logits) )
def least_square(labels, logits):
return tf.reduce_mean( (labels - logits) ** 2 )
if config['loss_mode'] == 'lsgan':
# ---[ GAN loss: LSGAN loss (chi-square, or called least-square)
loss_func = least_square
else:
# ---[ GAN loss: sigmoid BCE loss
loss_func = sigmoid_cross_entropy
# ---[ LOSS
_img_recovery = config['L1_lambda'] * tf.reduce_mean( tf.abs(source_images_batch - fake_2_images_output))
#_seg_recovery = config['L1_lambda'] * tf.reduce_mean( tf.abs(source_segments_batch - fake_1_segments_output)) #r1.0: error
#_seg_recovery = config['L1_lambda'] * tf.reduce_mean( tf.abs(source_segments_batch - fake_2_segments_output)) #r2.0
_seg_recovery = config['L1_lambda'] * tf.reduce_mean( tf.abs(source_segments_batch_color - fake_2_segments_output_color)) #r2.0.5: not sure because, in theory, no gradient if using decode_labels()
g_loss_a2b = \
loss_func( labels=tf.ones_like(d_fake_seg_output), logits=d_fake_seg_output ) + \
_img_recovery + _seg_recovery
g_loss_b2a = \
loss_func( labels=tf.ones_like(d_fake_img_output), logits=d_fake_img_output ) + \
_img_recovery + _seg_recovery
g_loss = \
loss_func( labels=tf.ones_like(d_fake_seg_output), logits=d_fake_seg_output ) + \
loss_func( labels=tf.ones_like(d_fake_img_output), logits=d_fake_img_output ) + \
_img_recovery + _seg_recovery
da_loss = \
loss_func( labels=tf.ones_like(d_real_img_output), logits=d_real_img_output ) + \
loss_func( labels=tf.zeros_like(d_fake_img_output), logits=d_fake_img_output )
db_loss = \
loss_func( labels=tf.ones_like(d_real_seg_output), logits=d_real_seg_output ) + \
loss_func( labels=tf.zeros_like(d_fake_seg_output), logits=d_fake_seg_output )
d_loss = \
(da_loss + db_loss) / 2.
# D will output [BATCH_SIZE, 32, 32, 1]
num_da_real_img_acc = tf.size( tf.where(tf.reduce_mean(tf.nn.sigmoid(d_real_img_output), axis=[1,2,3]) > 0.5)[:,0], name='num_da_real_img_acc' )
num_da_fake_img_acc = tf.size( tf.where(tf.reduce_mean(tf.nn.sigmoid(d_fake_img_output), axis=[1,2,3]) < 0.5)[:,0], name='num_da_fake_img_acc' )
num_db_real_seg_acc = tf.size( tf.where(tf.reduce_mean(tf.nn.sigmoid(d_real_seg_output), axis=[1,2,3]) > 0.5)[:,0], name='num_db_real_seg_acc' )
num_db_fake_seg_acc = tf.size( tf.where(tf.reduce_mean(tf.nn.sigmoid(d_fake_seg_output), axis=[1,2,3]) < 0.5)[:,0], name='num_db_fake_seg_acc' )
## limit weights to 0
#g_weight_regularizer = [0.0001 * tf.nn.l2_loss(v) for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, GEN_NAME) if 'weight' in v.name]
#g_weight_regularizer = tf.add_n(g_weight_regularizer, name='g_weight_regularizer_loss')
#g_loss += g_weight_regularizer
#d_weight_regularizer = [0.0001 * tf.nn.l2_loss(v) for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, DIS_NAME) if 'weight' in v.name]
#d_weight_regularizer = tf.add_n(d_weight_regularizer, name='d_weight_regularizer_loss')
#d_loss += d_weight_regularizer
d_loss = tf.identity(d_loss, name='d_loss')
g_loss = tf.identity(g_loss, name='g_loss')
## --- Training Set Validation ---
# Predictions.
#pred_gt = tf.reshape(target_segments_batch, [-1,])
#pred = tf.reshape(target_segments_output, [-1,])
#indices = tf.squeeze(tf.where(tf.not_equal(pred_gt, ignore_label)), 1)
#pred_gt = tf.cast(tf.gather(pred_gt, indices), tf.int32)
#pred = tf.cast(tf.gather(pred, indices), tf.int32)
## mIoU
### Allowing to use indices matrices in mean_iou() with `num_classes=indices.max()`
#weights = tf.cast(tf.less_equal(pred_gt, num_labels), tf.int32) # Ignoring all labels greater than or equal to n_classes.
#mIoU, mIoU_update_op = tf.metrics.mean_iou(pred, pred_gt, num_classes=num_labels, weights=weights)
# ---[ Variables
g_a2b_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, GEN_A2B_NAME)
g_b2a_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, GEN_B2A_NAME)
d_a_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, DIS_A_NAME)
d_b_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, DIS_B_NAME)
g_vars = g_a2b_vars + g_b2a_vars
d_vars = d_a_vars + d_b_vars
print_list(g_a2b_vars, GEN_A2B_NAME)
print_list(g_b2a_vars, GEN_B2A_NAME)
print_list(d_a_vars, DIS_A_NAME)
print_list(d_b_vars, DIS_B_NAME)
# ---[ Optimizer
## `colocate_gradients_with_ops = True` to reduce GPU MEM utils, and fasten training speed
OPT_NAME = 'Optimizer'
g_opts = []; d_opts = []
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
with tf.variable_scope(OPT_NAME):
#with tf.device('/device:GPU:{}'.format(config['gpus']-1)):
if True:
if len(g_vars) > 0:
g_opt = tf.train.AdamOptimizer(learning_rate=config['g_lr'], beta1=0.5, beta2=0.9).minimize(g_loss,
var_list=g_vars, colocate_gradients_with_ops=True)
g_opts.append(g_opt)
if len(d_vars) > 0:
d_opt = tf.train.AdamOptimizer(learning_rate=config['d_lr'], beta1=0.5, beta2=0.9).minimize(d_loss,
var_list=d_vars, colocate_gradients_with_ops=True)
d_opts.append(d_opt)
g_opt = tf.group(*g_opts)
d_opt = tf.group(*d_opts)
opt_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, OPT_NAME)
print_list(opt_vars, OPT_NAME)
# --- [ Summary
scalars = [d_loss, g_loss]
#scalars += [mIoU]
scalars += [num_da_real_img_acc, num_da_fake_img_acc, num_db_real_seg_acc, num_db_fake_seg_acc]
scalars += [g_loss_a2b, g_loss_b2a, da_loss, db_loss]
writer, summarys = create_summary(summary_dir=config['summary_dir'], name=config['suffix'],
scalar = scalars,
)
'''
Training
'''
with tf.Session(config=GpuConfig) as sess:
sess.run(tf.global_variables_initializer()) #DONOT put it after ``saver.restore``
sess.run(tf.local_variables_initializer()) #DONOT put it after ``saver.restore``
saver = tf.train.Saver(g_vars + d_vars, max_to_keep=1)
#g_saver = tf.train.Saver(g_vars, max_to_keep=1)
#d_saver = tf.train.Saver(d_vars, max_to_keep=1)
#if self.ckpt:
#saver.restore(sess, self.ckpt)
#print ("Training starts at %d iteration..." % sess.run(global_step))
feeds = {}
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
inside_epoch = int(config['print_epoch']) if config['print_epoch'] < config['max_epoch'] else int(config['max_epoch'] / 1)
outside_epoch = int(config['max_epoch'] / inside_epoch)
start = int(sess.run(global_step) / inside_epoch)
if start >= outside_epoch:
raise ValueError("initial iteration:%d >= max iteration:%d. please reset '--max_epoch' value." % (sess.run(global_step), config['max_epoch']))
start_time = time.time()
for epo in range(start, outside_epoch):
bar = IncrementalBar('[epoch {:<4d}/{:<4d}]'.format(epo, outside_epoch), max=inside_epoch)
for epi in range(inside_epoch):
iters = sess.run(global_step)
# save summary
if epo == 0:
save_summarys = sess.run(summarys, feed_dict=feeds)
writer.add_summary(save_summarys, iters)
for _ in range(config['d_epoch']):
sess.run(d_opt, feed_dict=feeds)
if iters > self.pretrain_D_epoch:
for _ in range(config['g_epoch']):
sess.run(g_opt, feed_dict=feeds)
sess.run(global_step_update)
bar.next()
duration = time.time() - start_time
disc_loss, gen_loss = \
sess.run([d_loss, g_loss], feed_dict=feeds)
na_real, na_fake, nb_real, nb_fake = \
sess.run([num_da_real_img_acc, num_da_fake_img_acc, num_db_real_seg_acc, num_db_fake_seg_acc], feed_dict=feeds)
#sess.run(mIoU_update_op, feed_dict=feeds)
#miou = sess.run(mIoU, feed_dict=feeds)
print (' -',
'DLoss: %-8.2e' % disc_loss,
#'(W: %-8.2e)' % disc_wloss,
'GLoss: %-8.2e' % gen_loss,
#'(W: %-8.2e)' % gen_wloss,
'|',
'[Da_img] #real: %d, #fake: %d' % (na_real, na_fake),
'[Db_seg] #real: %d, #fake: %d' % (nb_real, nb_fake),
'|',
#'[train_mIoU] %.2f' % miou,
'[ETA] %s' % format_time(duration)
)
bar.finish()
iters = sess.run(global_step)
# save checkpoint
if epo % 2 == 0:
saver_path = os.path.join(config['ckpt_dir'], '{}.ckpt'.format(config['name']))
saver.save(sess, save_path=saver_path, global_step=global_step)
# save summary
if epo % 1 == 0:
save_summarys = sess.run(summarys, feed_dict=feeds)
writer.add_summary(save_summarys, iters)
# output samples
if epo % 5 == 0:
img_gt, seg_gt, seg_1, seg_2, img_1, img_2 = sess.run(source_data_color)
print ("Range %10s:" % "seg_gt", seg_gt.min(), seg_gt.max())
print ("Range %10s:" % "seg_1", seg_1.min(), seg_1.max())
print ("Range %10s:" % "seg_2", seg_2.min(), seg_2.max())
print ("Range %10s:" % "img_gt", img_gt.min(), img_gt.max())
print ("Range %10s:" % "img_1", img_1.min(), img_1.max())
print ("Range %10s:" % "img_2", img_2.min(), img_2.max())
_output = np.concatenate([img_gt, seg_gt, seg_1, img_1, img_2, seg_2], axis=0)
save_visualization(_output, save_path=os.path.join(config['result_dir'], 'tr-{}.jpg'.format(iters)), size=[3, 2*config['batch_size']])
#seg_output = np.concatenate([seg_gt, seg_2, seg_1], axis=0)
#img_output = np.concatenate([img_gt, img_2, img_1], axis=0)
#save_visualization(seg_output, save_path=os.path.join(config['result_dir'], 'tr-seg-1gt_2mapback_3map-{}.jpg'.format(iters)), size=[3, config['batch_size']])
#save_visualization(img_output, save_path=os.path.join(config['result_dir'], 'tr-img-1gt_2mapback_3map-{}.jpg'.format(iters)), size=[3, config['batch_size']])
for i,target_data_color in enumerate(target_data_color_queue):
val_img_gt, val_seg_gt, val_seg_1, val_seg_2, val_img_1, val_img_2 = sess.run(target_data_color)
print ("Val Range %10s:" % "seg_gt", val_seg_gt.min(), val_seg_gt.max())
print ("Val Range %10s:" % "seg_1", val_seg_1.min(), val_seg_1.max())
print ("Val Range %10s:" % "seg_2", val_seg_2.min(), val_seg_2.max())
print ("Val Range %10s:" % "img_gt", val_img_gt.min(), val_img_gt.max())
print ("Val Range %10s:" % "img_1", val_img_1.min(), val_img_1.max())
print ("Val Range %10s:" % "img_2", val_img_2.min(), val_img_2.max())
_output = np.concatenate([val_img_gt, val_seg_gt, val_seg_1, val_img_1, val_img_2, val_seg_2], axis=0)
save_visualization(_output, save_path=os.path.join(config['result_dir'], 'val{}-{}.jpg'.format(i,iters)), size=[3, 2*config['batch_size']])
#val_seg_output = np.concatenate([val_seg_gt, val_seg_2, val_seg_1], axis=0)
#val_img_output = np.concatenate([val_img_gt, val_img_2, val_img_1], axis=0)
#save_visualization(seg_output, save_path=os.path.join(config['result_dir'], 'val{}-seg-1gt_2mapback_3map-{}.jpg'.format(i,iters)), size=[3, config['batch_size']])
#save_visualization(img_output, save_path=os.path.join(config['result_dir'], 'val{}-img-1gt_2mapback_3map-{}.jpg'.format(i,iters)), size=[3, config['batch_size']])
writer.flush()
writer.close()
def train(args):
## set hyparameter
img_mean = np.array((104.00698793, 116.66876762, 122.67891434),
dtype=np.float32)
tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
print("g_model_name:", args.g_name)
print("lambda:", args.lambd)
print("learning_rate:", args.learning_rate)
print("is_val:", args.is_val)
print("---------------------------------")
## load data
with tf.name_scope("create_inputs"):
reader = ImageReader(args.data_dir, args.img_size, args.random_scale,
args.random_mirror, args.random_crop,
args.ignore_label, args.is_val, img_mean, coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
print("Data is ready!")
## load model
g_net = choose_generator(args.g_name, image_batch)
score_map = g_net.get_output() # [batch_size, h, w, num_classes]
label, logits = convert_to_calculateloss(score_map, args.num_classes,
label_batch)
predict_label = tf.argmax(logits, axis=1)
predict_batch = g_net.topredict(score_map, tf.shape(image_batch)[1:3])
print("The model has been created!")
## get all kinds of variables list
if '50' not in args.g_name: # aim at vgg16
g_restore_var = [
v for v in tf.global_variables()
if 'generator' in v.name and 'image' in v.name
]
g_trainable_var = [
v for v in tf.trainable_variables()
if 'generator' in v.name and 'upscore' not in v.name
]
else: # aim at resnet50
g_restore_var = [
v for v in tf.global_variables() if 'fc' not in v.name
]
g_trainable_var = [
v for v in tf.trainable_variables()
if 'beta' not in v.name or 'gamma' not in v.name
]
## set loss
mce_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=label,
logits=logits))
# l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
# g_loss = tf.reduce_mean(mce_loss) + tf.add_n(l2_losses)
g_loss = mce_loss # don't add the penalization
g_loss_var, g_loss_op = tf.metrics.mean(g_loss)
iou_var, iou_op = tf.metrics.mean_iou(label, predict_label,
args.num_classes)
accuracy_var, acc_op = tf.metrics.accuracy(label, predict_label)
metrics_op = tf.group(g_loss_op, iou_op, acc_op)
## set optimizer
iterstep = tf.placeholder(dtype=tf.float32,
shape=[],
name='iteration_step')
base_lr = tf.constant(args.learning_rate, dtype=tf.float32, shape=[])
lr = tf.scalar_mul(base_lr,
tf.pow(
(1 - iterstep / args.num_steps),
args.power)) # learning rate reduce with the time
# g_gradients = tf.train.MomentumOptimizer(learning_rate=lr, momentum=args.momentum).compute_gradients(g_loss,
# g_trainable_var)
train_g_op = tf.train.MomentumOptimizer(learning_rate=lr,
momentum=args.momentum).minimize(
g_loss,
var_list=g_trainable_var)
train_all_op = train_g_op
## set summary
vs_image = tf.py_func(inv_preprocess,
[image_batch, args.save_num_images, img_mean],
tf.uint8)
vs_label = tf.py_func(
decode_labels, [label_batch, args.save_num_images, args.num_classes],
tf.uint8)
vs_predict = tf.py_func(
decode_labels, [predict_batch, args.save_num_images, args.num_classes],
tf.uint8)
tf.summary.image(name='image collection_train',
tensor=tf.concat(axis=2,
values=[vs_image, vs_label, vs_predict]),
max_outputs=args.save_num_images)
tf.summary.scalar('g_loss_train', g_loss_var)
tf.summary.scalar('iou_train', iou_var)
tf.summary.scalar('accuracy_train', accuracy_var)
# for grad, var in g_gradients:
# tf.summary.histogram(var.op.name + "/gradients", grad)
#
# for var in tf.trainable_variables():
# tf.summary.histogram(var.op.name + "/values", var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.log_dir,
graph=tf.get_default_graph(),
max_queue=10)
## set session
print("GPU index:" + str(os.environ['CUDA_VISIBLE_DEVICES']))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
global_init = tf.global_variables_initializer()
local_init = tf.local_variables_initializer()
sess.run(global_init)
sess.run(local_init)
## set saver
saver_all = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=5)
trained_step = 0
if os.path.exists(args.restore_from + 'checkpoint'):
trained_step = load_weight(args.restore_from, saver_all, sess)
else:
if '50' in args.g_name:
saver_g = tf.train.Saver(var_list=g_restore_var)
load_weight(args.baseweight_from['res50'], saver_g, sess)
elif 'vgg' in args.g_name:
load_weight(args.baseweight_from['vgg16'], g_restore_var, sess)
threads = tf.train.start_queue_runners(sess, coord)
print("all setting has been done,training start!")
## start training
for step in range(args.num_steps):
now_step = int(
trained_step) + step if trained_step is not None else step
feed_dict = {iterstep: now_step}
_, _, g_loss_ = sess.run([train_all_op, metrics_op, g_loss], feed_dict)
if step > 0 and step % args.save_pred_every == 0:
save_weight(args.restore_from, saver_all, sess, now_step)
if step % 50 == 0 or step == args.num_steps - 1:
print('step={} g_loss={}'.format(now_step, g_loss_))
summary_str = sess.run(summary_op, feed_dict)
summary_writer.add_summary(summary_str, now_step)
sess.run(local_init)
## end training
coord.request_stop()
coord.join(threads)
print('end....')
def train(args):
## set hyparameter
img_mean = np.array((104.00698793, 116.66876762, 122.67891434),
dtype=np.float32)
tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
print("g_name:", args.g_name)
print("d_name:", args.d_name)
print("lambda:", args.lambd)
print("learning_rate:", args.learning_rate)
print("is_val:", args.is_val)
print("---------------------------------")
## load data
with tf.name_scope("create_inputs"):
reader = ImageReader(args.data_dir, args.img_size, args.random_scale,
args.random_mirror, args.random_crop,
args.ignore_label, args.is_val, img_mean, coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
print("Data is ready!")
## load model
image_normal_batch = tf.train.batch([
(reader.image + img_mean) / 255.,
],
args.batch_size,
dynamic_pad=True)
g_net, g_net_x = choose_generator(args.g_name, image_batch,
image_normal_batch)
score_map = g_net.get_output()
fk_batch = tf.nn.softmax(score_map, dim=-1)
pre_batch = tf.expand_dims(tf.cast(tf.argmax(fk_batch, axis=-1), tf.uint8),
axis=-1)
gt_batch = tf.image.resize_nearest_neighbor(label_batch,
tf.shape(score_map)[1:3])
gt_batch = tf.where(tf.equal(gt_batch, args.ignore_label), pre_batch,
gt_batch)
gt_batch = convert_to_scaling(fk_batch, args.num_classes, gt_batch)
x_batch = g_net_x.get_appointed_layer('generator/image_conv5_3')
d_fk_net, d_gt_net = choose_discriminator(args.d_name, fk_batch, gt_batch,
x_batch)
d_fk_pred = d_fk_net.get_output() # fake segmentation result in d
d_gt_pred = d_gt_net.get_output() # ground-truth result in d
label, logits = convert_to_calculateloss(score_map, args.num_classes,
label_batch)
predict_label = tf.argmax(logits, axis=1)
predict_batch = g_net.topredict(score_map, tf.shape(image_batch)[1:3])
print("The model has been created!")
## get all kinds of variables list
g_restore_var = [
v for v in tf.global_variables() if 'discriminator' not in v.name
]
g_var = [
v for v in tf.trainable_variables()
if 'generator' in v.name and 'deconv' not in v.name
]
d_var = [v for v in tf.trainable_variables() if 'discriminator' in v.name]
# g_trainable_var = [v for v in g_var if 'beta' not in v.name or 'gamma' not in v.name] # batch_norm training open
g_trainable_var = g_var
d_trainable_var = d_var
## set loss
mce_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=label,
logits=logits))
# l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
# mce_loss = tf.reduce_mean(mce_loss) + tf.add_n(l2_losses)
# g_bce_loss = tf.reduce_mean(tf.log(d_fk_pred + eps))
g_bce_loss = args.lambd * tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(d_fk_pred),
logits=d_fk_pred))
g_loss = mce_loss + g_bce_loss
# d_loss = tf.reduce_mean(tf.constant(-1.0) * [tf.log(d_gt_pred + eps) + tf.log(1. - d_fk_pred + eps)])
d_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(d_gt_pred), logits=d_gt_pred) \
+ tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(d_fk_pred),
logits=d_fk_pred))
fk_score_var = tf.reduce_mean(tf.sigmoid(d_fk_pred))
gt_score_var = tf.reduce_mean(tf.sigmoid(d_gt_pred))
mce_loss_var, mce_loss_op = tf.metrics.mean(mce_loss)
g_bce_loss_var, g_bce_loss_op = tf.metrics.mean(g_bce_loss)
g_loss_var, g_loss_op = tf.metrics.mean(g_loss)
d_loss_var, d_loss_op = tf.metrics.mean(d_loss)
iou_var, iou_op = tf.metrics.mean_iou(label, predict_label,
args.num_classes)
accuracy_var, acc_op = tf.metrics.accuracy(label, predict_label)
metrics_op = tf.group(mce_loss_op, g_bce_loss_op, g_loss_op, d_loss_op,
iou_op, acc_op)
## set optimizer
iterstep = tf.placeholder(dtype=tf.float32,
shape=[],
name='iteration_step')
base_lr = tf.constant(args.learning_rate, dtype=tf.float32, shape=[])
lr = tf.scalar_mul(base_lr,
tf.pow(
(1 - iterstep / args.num_steps),
args.power)) # learning rate reduce with the time
# g_gradients = tf.train.MomentumOptimizer(learning_rate=lr,
# momentum=args.momentum).compute_gradients(g_loss,
# var_list=g_trainable_var)
g_gradients = tf.train.AdamOptimizer(learning_rate=lr).compute_gradients(
g_loss, var_list=g_trainable_var)
d_gradients = tf.train.MomentumOptimizer(
learning_rate=lr * 10,
momentum=args.momentum).compute_gradients(d_loss,
var_list=d_trainable_var)
grad_fk_oi = tf.gradients(d_fk_pred, fk_batch, name='grad_fk_oi')[0]
grad_gt_oi = tf.gradients(d_gt_pred, gt_batch, name='grad_gt_oi')[0]
grad_fk_img_oi = tf.gradients(d_fk_pred,
image_batch,
name='grad_fk_img_oi')[0]
grad_gt_img_oi = tf.gradients(d_gt_pred,
image_batch,
name='grad_gt_img_oi')[0]
train_g_op = tf.train.AdamOptimizer(learning_rate=lr).minimize(
g_loss, var_list=g_trainable_var)
train_d_op = tf.train.MomentumOptimizer(learning_rate=lr * 10,
momentum=args.momentum).minimize(
d_loss,
var_list=d_trainable_var)
## set summary
vs_image = tf.py_func(inv_preprocess,
[image_batch, args.save_num_images, img_mean],
tf.uint8)
vs_label = tf.py_func(
decode_labels, [label_batch, args.save_num_images, args.num_classes],
tf.uint8)
vs_predict = tf.py_func(
decode_labels, [predict_batch, args.save_num_images, args.num_classes],
tf.uint8)
tf.summary.image(name='image collection_train',
tensor=tf.concat(axis=2,
values=[vs_image, vs_label, vs_predict]),
max_outputs=args.save_num_images)
tf.summary.scalar('fk_score', fk_score_var)
tf.summary.scalar('gt_score', gt_score_var)
tf.summary.scalar('g_loss_train', g_loss_var)
tf.summary.scalar('d_loss_train', d_loss_var)
tf.summary.scalar('mce_loss_train', mce_loss_var)
tf.summary.scalar('g_bce_loss_train', g_bce_loss_var)
tf.summary.scalar('iou_train', iou_var)
tf.summary.scalar('accuracy_train', accuracy_var)
tf.summary.scalar('grad_fk_oi', tf.reduce_mean(tf.abs(grad_fk_oi)))
tf.summary.scalar('grad_gt_oi', tf.reduce_mean(tf.abs(grad_gt_oi)))
tf.summary.scalar('grad_fk_img_oi', tf.reduce_mean(tf.abs(grad_fk_img_oi)))
tf.summary.scalar('grad_gt_img_oi', tf.reduce_mean(tf.abs(grad_gt_img_oi)))
for grad, var in g_gradients + d_gradients:
tf.summary.histogram(var.op.name + "/gradients", grad)
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name + "/values", var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.log_dir,
graph=tf.get_default_graph(),
max_queue=3)
## set session
print("GPU index:" + str(os.environ['CUDA_VISIBLE_DEVICES']))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
global_init = tf.global_variables_initializer()
local_init = tf.local_variables_initializer()
sess.run(global_init)
sess.run(local_init)
## set saver
saver_all = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=2)
trained_step = 0
if os.path.exists(args.restore_from + 'checkpoint'):
trained_step = load_weight(args.restore_from, saver_all, sess)
else:
saver_g = tf.train.Saver(var_list=g_restore_var, max_to_keep=2)
load_weight(args.baseweight_from['g'], saver_g,
sess) # the weight is the completely g model
threads = tf.train.start_queue_runners(sess, coord)
print("all setting has been done,training start!")
## start training
# def auto_setting_train_steps(mode):
# if mode == 0:
# return 5, 1
# elif mode == 1:
# return 1, 5
# else:
# return 1, 1
d_train_steps = 10
g_train_steps = 1
# flags = [0 for i in range(3)]
for step in range(args.num_steps):
now_step = int(
trained_step) + step if trained_step is not None else step
feed_dict = {iterstep: step}
for i in range(d_train_steps):
_, _ = sess.run([train_d_op, metrics_op], feed_dict)
for i in range(g_train_steps):
g_loss_, mce_loss_, g_bce_loss_, d_loss_, _, _ = sess.run([
g_loss_var, mce_loss_var, g_bce_loss_var, d_loss_var,
train_g_op, metrics_op
], feed_dict)
########################
# fk_score_, gt_score_ = sess.run([fk_score_var, gt_score_var], feed_dict)
# if fk_score_ > 0.48 and fk_score_ < 0.52:
# flags[0] += 1
# flags[1] = flags[2] = 0
# elif gt_score_ - fk_score_ > 0.3:
# flags[1] += 1
# flags[0] = flags[2] = 0
# else:
# flags[2] += 1
# flags[0] = flags[1] = 0
# if max(flags) > 100:
# d_train_steps, g_train_steps = auto_setting_train_steps(flags.index(max(flags)))
########################
if step > 0 and step % args.save_pred_every == 0:
save_weight(args.restore_from, saver_all, sess, now_step)
if step % 50 == 0 or step == args.num_steps - 1:
print('step={} d_loss={} g_loss={} mce_loss={} g_bce_loss_={}'.
format(now_step, d_loss_, g_loss_, mce_loss_, g_bce_loss_))
summary_str = sess.run(summary_op, feed_dict)
summary_writer.add_summary(summary_str, now_step)
sess.run(local_init)
## end training
coord.request_stop()
coord.join(threads)
print('end....')
