import time
import sys
try:
serial_interface = sys.argv[1]
image_file = sys.argv[2]
colorduinos_horizontal = int(sys.argv[3])
colorduinos_vertical = int(sys.argv[4])
colorduino_x = int(sys.argv[5])
colorduino_y = int(sys.argv[6])
except Exception:
if(len(sys.argv) != 7):
print "Usage: python main.py SERIAL_INTERFACE IMAGE_FILE COLORDUINOS_HORIZONTAL COLORDUINOS_VERTICAL COLORDUINO_X COLORDUINO_Y"
sys.exit()
im=ImageReader(sys.argv[2])
im.resize(colorduinos_horizontal*8,colorduinos_vertical*8)
im.load()
ser = SerialComm(serial_interface)
time.sleep(2)
for x in range(0,8):
for y in range(0,8):
rgb = im.getColorAt(colorduino_x*8 + x, colorduino_y*8 + y)
r = rgb[0]
g = rgb[1]
b = rgb[2]
print "x: %i, y: %i" % (int(colorduino_x)*8+x, int(colorduino_y*8)+y)
line = "%02X,%02X,%02X,%02X,%02X" % (y,x,r,g,b)
ser.writeln(line)
def run(self):
tf.set_random_seed(self.random_seed)
coord = tf.train.Coordinator()
# 读取数据
with tf.name_scope("create_inputs"):
reader = ImageReader(self.data_dir, self.data_train_list,
self.input_size, self.random_scale,
self.random_mirror, self.ignore_label,
self.img_mean, coord)
image_batch, label_batch = reader.dequeue(self.batch_size)
# 网络
net = PSPNet({'data': image_batch},
is_training=True,
num_classes=self.num_classes)
raw_output = net.layers['conv6']
# According from the prototxt in Caffe implement, learning rate must multiply by 10.0 in pyramid module
fc_list = [
'conv5_3_pool1_conv', 'conv5_3_pool2_conv', 'conv5_3_pool3_conv',
'conv5_3_pool6_conv', 'conv6', 'conv5_4'
]
# 所有的变量
restore_var = [v for v in tf.global_variables()]
# 所有可训练变量
all_trainable = [
v for v in tf.trainable_variables()
if ('beta' not in v.name and 'gamma' not in v.name)
or self.train_beta_gamma
]
# fc_list中的全连接层可训练变量和卷积可训练变量
fc_trainable = [
v for v in all_trainable if v.name.split('/')[0] in fc_list
]
conv_trainable = [
v for v in all_trainable if v.name.split('/')[0] not in fc_list
] # lr * 1.0
fc_w_trainable = [v for v in fc_trainable
if 'weights' in v.name] # lr * 10.0
fc_b_trainable = [v for v in fc_trainable
if 'biases' in v.name] # lr * 20.0
# 验证
assert (len(all_trainable) == len(fc_trainable) + len(conv_trainable))
assert (len(fc_trainable) == len(fc_w_trainable) + len(fc_b_trainable))
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, self.num_classes])
label_process = prepare_label(label_batch,
tf.stack(raw_output.get_shape()[1:3]),
num_classes=self.num_classes,
one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_process, [
-1,
])
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, self.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction, labels=gt)
l2_losses = [
self.weight_decay * tf.nn.l2_loss(v)
for v in tf.trainable_variables() if 'weights' in v.name
]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
# Using Poly learning rate policy
base_lr = tf.constant(self.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(
base_lr, tf.pow((1 - step_ph / self.num_steps), self.power))
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
update_ops = None if not self.update_mean_var else tf.get_collection(
tf.GraphKeys.UPDATE_OPS)
# 对变量以不同的学习率优化：分别求梯度、应用梯度
with tf.control_dependencies(update_ops):
opt_conv = tf.train.MomentumOptimizer(learning_rate, self.momentum)
opt_fc_w = tf.train.MomentumOptimizer(learning_rate * 10.0,
self.momentum)
opt_fc_b = tf.train.MomentumOptimizer(learning_rate * 20.0,
self.momentum)
grads = tf.gradients(
reduced_loss, conv_trainable + fc_w_trainable + fc_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_fc_w = grads[len(conv_trainable):(len(conv_trainable) +
len(fc_w_trainable))]
grads_fc_b = grads[(len(conv_trainable) + len(fc_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(
zip(grads_conv, conv_trainable))
train_op_fc_w = opt_fc_w.apply_gradients(
zip(grads_fc_w, fc_w_trainable))
train_op_fc_b = opt_fc_b.apply_gradients(
zip(grads_fc_b, fc_b_trainable))
train_op = tf.group(train_op_conv, train_op_fc_w, train_op_fc_b)
pass
sess = tf.Session(config=self.config)
sess.run(tf.global_variables_initializer())
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=10)
# 加载模型
ckpt = tf.train.get_checkpoint_state(self.log_dir)
if ckpt and ckpt.model_checkpoint_path:
tf.train.Saver(var_list=restore_var).restore(
sess, ckpt.model_checkpoint_path)
Tools.print_info("Restored model parameters from {}".format(
ckpt.model_checkpoint_path))
else:
Tools.print_info('No checkpoint file found.')
pass
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(self.num_steps):
start_time = time.time()
if step % self.save_pred_freq == 0:
loss_value, _ = sess.run([reduced_loss, train_op],
feed_dict={step_ph: step})
saver.save(sess, self.checkpoint_path, global_step=step)
Tools.print_info('The checkpoint has been created.')
else:
loss_value, _ = sess.run([reduced_loss, train_op],
feed_dict={step_ph: step})
duration = time.time() - start_time
Tools.print_info(
'step {:d} \t loss = {:.3f}, ({:.3f} sec/step)'.format(
step, loss_value, duration))
coord.request_stop()
coord.join(threads)
pass
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_up_to(
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 main():
args = get_arguments()
print(args)
coord = tf.train.Coordinator()
tf.reset_default_graph()
with tf.name_scope("create_inputs"):
reader = ImageReader(
DATA_DIRECTORY,
DATA_LIST_PATH,
input_size,
None,
None,
ignore_label,
IMG_MEAN,
coord)
image, label = reader.image, reader.label
image_batch, label_batch = tf.expand_dims(image, dim=0), tf.expand_dims(label, dim=0) # Add one batch dimension.
# Create network.
net = PSPNet({'data': image_batch}, is_training=False, num_classes=num_classes)
with tf.variable_scope('', reuse=True):
flipped_img = tf.image.flip_left_right(image)
flipped_img = tf.expand_dims(flipped_img, dim=0)
net2 = PSPNet({'data': flipped_img}, is_training=False, num_classes=num_classes)
# Which variables to load.
restore_var = tf.global_variables()
# Predictions.
raw_output = net.layers['conv6']
if args.flipped_eval:
flipped_output = tf.image.flip_left_right(tf.squeeze(net2.layers['conv6']))
flipped_output = tf.expand_dims(flipped_output, dim=0)
raw_output = tf.add_n([raw_output, flipped_output])
raw_output_up = tf.image.resize_bilinear(raw_output, size=input_size, align_corners=True)
raw_output_up = tf.argmax(raw_output_up, dimension=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# mIoU
pred_flatten = tf.reshape(pred, [-1,])
raw_gt = tf.reshape(label_batch, [-1,])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
pred = tf.gather(pred_flatten, indices)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(pred, gt, num_classes=num_classes)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
sess.run(tf.local_variables_initializer())
restore_var = tf.global_variables()
ckpt = tf.train.get_checkpoint_state(args.model)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for step in range(num_steps):
preds, _ = sess.run([pred, update_op])
if step > 0 and args.measure_time:
calculate_time(sess, net)
if step % 10 == 0:
print('Finish {0}/{1}'.format(step, num_steps))
print('step {0} mIoU: {1}'.format(step, sess.run(mIoU)))
print('step {0} mIoU: {1}'.format(step, sess.run(mIoU)))
coord.request_stop()
coord.join(threads)
def __init__(self, root, im_file, masks_dir):
self._reader = ImageReader(root, im_file, masks_dir)
self._root = root
self._image = self._reader.image
self._mask = self._reader.mask
def main():
args = get_arguments()
print(args)
coord = tf.train.Coordinator()
tf.reset_default_graph()
with tf.name_scope("create_inputs"):
reader = ImageReader(DATA_DIRECTORY, DATA_LIST_PATH, input_size, None,
None, ignore_label, IMG_MEAN, coord)
image, label = reader.image, reader.label
image_batch, label_batch = tf.expand_dims(image, dim=0), tf.expand_dims(
label, dim=0) # Add one batch dimension.
# Create network.
if args.model[-2:] == 'bn':
net = ICNet_BN({'data': image_batch}, num_classes=num_classes)
else:
net = ICNet({'data': image_batch}, num_classes=num_classes)
# Which variables to load.
restore_var = tf.global_variables()
# Predictions.
raw_output = net.layers['conv6_cls']
raw_output_up = tf.image.resize_bilinear(raw_output,
size=input_size,
align_corners=True)
raw_output_up = tf.argmax(raw_output_up, dimension=3)
raw_pred = tf.expand_dims(raw_output_up, dim=3)
# mIoU
pred_flatten = tf.reshape(raw_pred, [
-1,
])
raw_gt = tf.reshape(label_batch, [
-1,
])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
pred = tf.gather(pred_flatten, indices)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(
pred, gt, num_classes=num_classes)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
sess.run(tf.local_variables_initializer())
restore_var = tf.global_variables()
if args.model == 'train':
print('Restore from train30k model...')
net.load(model_train30k, sess)
elif args.model == 'trainval':
print('Restore from trainval90k model...')
net.load(model_trainval90k, sess)
elif args.model == 'train_bn':
print('Restore from train30k bnnomerge model...')
net.load(model_train30k_bn, sess)
elif args.model == 'trainval_bn':
print('Restore from trainval90k bnnomerge model...')
net.load(model_trainval90k_bn, sess)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for step in range(num_steps):
preds, _ = sess.run([pred, update_op])
if step > 0 and args.measure_time:
calculate_time(sess, net, raw_pred)
if step % 10 == 0:
print('Finish {0}/{1}'.format(step, num_steps))
print('step {0} mIoU: {1}'.format(step, sess.run(mIoU)))
coord.request_stop()
coord.join(threads)
def main(argv=None):
input_size = (cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH)
# Create queue coordinator.
coord = tf.train.Coordinator()
# Load reader.
# Train
print('Train ' + cfg.train_data_list)
with tf.name_scope("create_inputs"):
reader = ImageReader(cfg.train_data_dir, cfg.train_data_list,
input_size, cfg.random_scale, cfg.random_resize,
cfg.random_mirror, cfg.random_color,
cfg.random_crop_pad, cfg.ignore_label,
cfg.IMG_MEAN, coord)
image_batch, label_batch = reader.dequeue(cfg.batch_size)
# Define Network
pred_annotation, logits = inference_deeplabv3_plus_16(
image_batch, is_training=True) # Modified
logits_loss = cross_entropy_loss(logits,
label_batch) # loss1 for ECP dataset
# logits_loss = weighted_cross_entropy_loss(logits, label_batch) # loss2 for RueMonge dataset
# logits_loss = weighted_cross_entropy_loss_4class(logits, label_batch)
# # PSPNet
# pred_annotation, logits, logits_dsn = inference_pspnet(image_batch, is_training=True) # PSPNet
# # logits_loss = cross_entropy_loss(logits, label_batch) + \
# # cross_entropy_loss(logits_dsn, label_batch) # loss1 for ECP dataset
# logits_loss = weighted_cross_entropy_loss(logits, label_batch) + \
# weighted_cross_entropy_loss(logits_dsn, label_batch) # loss2 for RueMonge dataset
ce_loss = logits_loss # cross entropy loss
# Show acc for validation or train dataset
if cfg.is_time_acc or cfg.is_epoch_acc:
with tf.variable_scope('', reuse=True):
val_image_batch = tf.placeholder(
tf.float32,
shape=[1, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name="input_image")
f = open(cfg.val_data_list, 'r')
val_img_list = []
val_label_list = []
for line in f:
try:
image_name, label = line.strip("\n").split(' ')
except ValueError: # Adhoc for test.
image_name = label = line.strip("\n")
val_img_list.append(cfg.val_data_dir + image_name)
val_label_list.append(cfg.val_data_dir + label)
_, val_logits = inference_deeplabv3_plus_16_init(
val_image_batch, is_training=False) # Modified
# _, val_logits, _ = inference_pspnet(val_image_batch, is_training=False) # PSPNet
val_logits_softmax = tf.nn.softmax(val_logits)
tf.group()
l2_loss = [
weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'weights' or 'w' in v.name or 'W' in v.name
] # encode: W, facade: weights
l2_losses = tf.add_n(l2_loss)
# Total loss
loss = ce_loss + l2_losses # + stru_loss
tf.summary.scalar("loss_ce", ce_loss)
tf.summary.scalar("l2_losses", l2_losses)
tf.summary.scalar("total_loss", loss)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
# Using Poly learning rate policy
base_lr = tf.constant(cfg.learning_rate)
learning_rate = tf.scalar_mul(base_lr,
tf.pow((1 - step_ph / global_step), power))
trainable_var = tf.trainable_variables()
# Optimizer
if cfg.optimizer == 'Adam':
optimizer = tf.train.AdamOptimizer(learning_rate)
print('Optimizer: Adam')
elif cfg.optimizer == 'Adam2':
optimizer = tf.train.AdamOptimizer(learning_rate,
beta1=0.9,
beta2=0.99)
elif cfg.optimizer == 'SGD':
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
elif cfg.optimizer == 'Momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
print('Optimizer: Momentum')
elif cfg.optimizer == 'RMSProp':
optimizer = tf.train.RMSPropOptimizer(learning_rate)
# grads = optimizer.compute_gradients(loss, var_list=trainable_var)
# train_op = optimizer.apply_gradients(grads)
## Optimizer definition - nothing different from any classical example
opt = optimizer
## Retrieve all trainable variables you defined in your graph
if cfg.freeze_bn:
tvs = [
v for v in tf.trainable_variables()
if 'beta' not in v.name and 'gamma' not in v.name
]
else:
tvs = [v for v in tf.trainable_variables()]
## Creation of a list of variables with the same shape as the trainable ones
# initialized with 0s
accum_vars = [
tf.Variable(tf.zeros_like(tv.initialized_value()), trainable=False)
for tv in tvs
]
zero_ops = [tv.assign(tf.zeros_like(tv)) for tv in accum_vars]
## Calls the compute_gradients function of the optimizer to obtain... the list of gradients
gvs = opt.compute_gradients(loss, tvs)
## Adds to each element from the list you initialized earlier with zeros its gradient (works because accum_vars and gvs are in the same order)
accum_ops = [accum_vars[i].assign_add(gv[0]) for i, gv in enumerate(gvs)]
## Define the training step (part with variable value update)
train_step = opt.apply_gradients([(accum_vars[i], gv[1])
for i, gv in enumerate(gvs)])
print("Setting up summary op...")
summary_op = tf.summary.merge_all()
# Set gpu usage
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 1.0
sess = tf.Session(config=config)
print("Setting up Saver...")
saver = tf.train.Saver(max_to_keep=cfg.model_save_num)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# create two summary writers to show training loss and validation loss in the same graph
# need to create two folders 'train' and 'validation' inside FLAGS.logs_dir
if not os.path.exists(cfg.logs_dir):
os.makedirs(cfg.logs_dir)
train_writer = tf.summary.FileWriter(cfg.logs_dir + 'train', sess.graph)
test_writer = tf.summary.FileWriter(cfg.logs_dir + 'test')
if not os.path.exists(cfg.save_dir):
os.makedirs(cfg.save_dir)
count = 0
files = os.path.join(cfg.save_dir + 'model.ckpt-*.index')
sfile = glob.glob(files)
if len(sfile) > 0:
sess.run(tf.global_variables_initializer())
sfile = glob.glob(files)
steps = []
for s in sfile:
part = s.split('.')
step = int(part[1].split('-')[1])
steps.append(step)
count = max(steps)
model = cfg.save_dir + 'model.ckpt-' + str(count)
print('\nRestoring weights from: ' + model)
saver.restore(sess, model)
print('End Restore')
else:
# # restore from pre-train on imagenet
variables = tf.global_variables()
sess.run(tf.variables_initializer(variables, name='init'))
# # tensorflow 1
if os.path.exists(cfg.pre_trained_model) or os.path.exists(
cfg.pre_trained_model + '.index'):
var_keep_dic = get_variables_in_checkpoint_file(
cfg.pre_trained_model)
# Get the variables to restore, ignoring the variables to fix
variables_to_restore = get_variables_to_restore(
variables, var_keep_dic)
if len(variables_to_restore) > 0:
restorer = tf.train.Saver(variables_to_restore)
restorer.restore(sess, cfg.pre_trained_model)
print('Model pre-train loaded from ' + cfg.pre_trained_model)
else:
print('Model inited random.')
else:
print('Model inited random.')
# RGB -> BGR
if 'res' in cfg.pre_trained_model:
conv1_rgb = tf.get_variable("conv1_rgb", [7, 7, 3, 64],
trainable=False)
restorer_fc = tf.train.Saver(
{'resnet_v1_50/conv1/weights': conv1_rgb})
restorer_fc.restore(sess, cfg.pre_trained_model)
sess.run(tf.assign(variables[0], tf.reverse(conv1_rgb, [2])))
print('ResNet Conv 1 RGB->BGR')
elif 'vgg' in cfg.pre_trained_model:
conv1_rgb = tf.get_variable("conv1_rgb", [3, 3, 3, 64],
trainable=False)
restorer_fc = tf.train.Saver(
{'vgg_16/conv1/conv1_1/weights': conv1_rgb})
restorer_fc.restore(sess, cfg.pre_trained_model)
sess.run(tf.assign(variables[0], tf.reverse(conv1_rgb, [2])))
print('Vgg Conv 1 RGB->BGR')
_mask = pred_annotation[0]
_img = image_batch[0]
_gt = label_batch[0]
if not os.path.exists(cfg.save_dir + 'temp_img'):
os.mkdir(cfg.save_dir + 'temp_img')
print('Start train ' + cfg.data_dir)
print('---------------Hyper Paras---------------')
print('-- batch_size: ', cfg.batch_size)
print('-- Gradient Accumulation: ', cfg.Gradient_Accumulation)
print('-- image height: ', cfg.IMAGE_HEIGHT)
print('-- image width: ', cfg.IMAGE_WIDTH)
print('-- learning rate: ', cfg.learning_rate)
print('-- GPU: ', cfg.use_gpu)
print('-- optimizer: ', cfg.optimizer)
print('-- class num: ', cfg.NUM_OF_CLASSESS)
print('-- total iter: ', cfg.total_iter)
print('-- Time acc: ', cfg.is_time_acc)
print('-- Acc interval: ', cfg.acc_interval)
print('-- Start Acc iter: ', cfg.start_show_iter)
print('-- Is save step: ', cfg.is_save_step)
print('-- Start save step: ', cfg.start_save_step)
print('-- save ecpoch: ', cfg.save_step_inter)
print('-- model save num: ', cfg.model_save_num)
print('-- summary interval: ', cfg.summary_interval)
print('-- weight decay: ', cfg.weight_decay)
print('-- Freeze BN: ', cfg.freeze_bn)
print('-- Decay rate: ', cfg.decay_rate)
print('-- minScale: ', cfg.minScale)
print('-- maxScale: ', cfg.maxScale)
print('-- random scale: ', cfg.random_scale)
print('-- random mirror: ', cfg.random_mirror)
print('-- random crop: ', cfg.random_crop_pad)
print('-- Validation on :' + str(cfg.val_data_list))
print('-- Pre-trained: ' + cfg.pre_trained_model)
print('----------------End---------------------')
fcfg = open(cfg.save_dir + 'cfg.txt', 'w')
fcfg.write('-- batch_size: ' + str(cfg.batch_size) + '\n')
fcfg.write('-- Gradient Accumulation: ' + str(cfg.Gradient_Accumulation) +
'\n')
fcfg.write('-- image height: ' + str(cfg.IMAGE_HEIGHT) + '\n')
fcfg.write('-- image width: ' + str(cfg.IMAGE_WIDTH) + '\n')
fcfg.write('-- learning rate: ' + str(cfg.learning_rate) + '\n')
fcfg.write('-- GPU: ' + str(cfg.use_gpu) + '\n')
fcfg.write('-- optimizer: ' + str(cfg.optimizer) + '\n')
fcfg.write('-- class num: ' + str(cfg.NUM_OF_CLASSESS) + '\n')
fcfg.write('-- total iter: ' + str(cfg.total_iter) + '\n')
fcfg.write('-- Time acc: ' + str(cfg.is_time_acc) + '\n')
fcfg.write('-- Acc interval: ' + str(cfg.acc_interval) + '\n')
fcfg.write('-- Start Acc iter: ' + str(cfg.start_show_iter) + '\n')
fcfg.write('-- Is save step: ' + str(cfg.is_save_step) + '\n')
fcfg.write('-- Start save step: ' + str(cfg.start_save_step) + '\n')
fcfg.write('-- save ecpoch: ' + str(cfg.save_step_inter) + '\n')
fcfg.write('-- model save num: ' + str(cfg.model_save_num) + '\n')
fcfg.write('-- summary interval: ' + str(cfg.summary_interval) + '\n')
fcfg.write('-- weight decay: ' + str(cfg.weight_decay) + '\n')
fcfg.write('-- Freeze BN: ' + str(cfg.freeze_bn) + '\n')
fcfg.write('-- Decay rate: ' + str(cfg.decay_rate) + '\n')
fcfg.write('-- minScale: ' + str(cfg.minScale) + '\n')
fcfg.write('-- maxScale: ' + str(cfg.maxScale) + '\n')
fcfg.write('-- random scale: ' + str(cfg.random_scale) + '\n')
fcfg.write('-- random mirror: ' + str(cfg.random_mirror) + '\n')
fcfg.write('-- random crop: ' + str(cfg.random_crop_pad) + '\n')
fcfg.write('-- Validation on :' + str(cfg.val_data_list) + '\n')
fcfg.write('-- Pre-trained: ' + cfg.pre_trained_model + '\n')
fcfg.close()
last_summary_time = time.time()
last_acc_time = time.time()
record = train_number / cfg.batch_size # iter number of each epoch
if cfg.is_save_step: # save with step
running_count = count
epo = int(count / record)
if cfg.is_save_epoch: # save with epoch
running_count = int(epo * record)
epo = count
best_acc = 0.5
best_step = 0
train_start_time = time.time()
start_step = running_count
lossTr_list = []
stepes = []
Acc_val_list = []
# Change the graph for read only
sess.graph.finalize()
while running_count < cfg.total_iter:
time_start = time.time()
itr = 0
while itr < int(record):
itr += 1
running_count += 1
# log last 10 model
if running_count > (cfg.total_iter -
10) and cfg.is_save_last10_model:
saver.save(sess, cfg.save_dir + 'model.ckpt',
int(running_count))
print('Model has been saved:' + str(running_count))
# more than total iter, stopping training
if running_count > cfg.total_iter:
break
feed_dict = {step_ph: (running_count)}
# save summary
now = time.time()
if now - last_summary_time > cfg.summary_interval:
summary_str = sess.run(summary_op,
feed_dict={step_ph: running_count})
train_writer.add_summary(summary_str, running_count)
last_summary_time = now
score_map, img, gt = sess.run([_mask, _img, _gt],
feed_dict=feed_dict)
img = np.array(img + cfg.IMG_MEAN, np.uint8)
score_map = score_map * 20
gt = gt * 20
save_temp = np.zeros(
(cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH * 3, 3), np.uint8)
save_temp[0:cfg.IMAGE_HEIGHT, 0:cfg.IMAGE_WIDTH, :] = img
save_temp[0:cfg.IMAGE_HEIGHT,
cfg.IMAGE_WIDTH:cfg.IMAGE_WIDTH * 2, :] = gt
save_temp[0:cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH *
2:cfg.IMAGE_WIDTH * 3, :] = score_map
cv2.imwrite(
cfg.save_dir + 'temp_img/' + str(now) + '_mask.jpg',
save_temp)
time_s = time.time()
# Run the zero_ops to initialize it
sess.run(zero_ops)
# Accumulate the gradients 'n_minibatches' times in accum_vars using accum_ops
for i in range(cfg.Gradient_Accumulation):
sess.run(accum_ops, feed_dict=feed_dict)
train_loss, ls_ce, ls_l2, lr = sess.run(
[loss, ce_loss, l2_losses, learning_rate], feed_dict=feed_dict)
if running_count > 50:
lossTr_list.append(ls_ce)
if start_step == 0:
start_step = 50
# Run the train_step ops to update the weights based on your accumulated gradients
sess.run(train_step, feed_dict=feed_dict)
time_e = time.time()
print(
"Epo: %d, Step: %d, Train_loss:%g, ce: %g, l2:%g, lr:%g, time:%g"
% (epo, running_count, train_loss, ls_ce, ls_l2, lr,
time_e - time_s))
# check accuracy per step of training data
if cfg.is_time_acc and running_count >= cfg.start_show_iter and \
running_count <= cfg.total_iter and (now-last_acc_time) > cfg.acc_interval:
# Test accuracy in val
hist = np.zeros((cfg.NUM_OF_CLASSESS, cfg.NUM_OF_CLASSESS))
for i, img_name in enumerate(val_img_list):
true_val = np.expand_dims(misc.imread(val_label_list[i]),
axis=2)
pred_val = evaluate_accuracy(val_logits_softmax, sess,
val_image_batch, img_name)
hist += fast_hist(true_val.flatten(), pred_val.flatten(),
cfg.NUM_OF_CLASSESS)
hist[0, :] = 0
# overall accuracy
over_acc = np.diag(hist).sum() / hist.sum()
print('>>> Step', running_count, 'overall accuracy', over_acc)
if over_acc > best_acc:
saver.save(sess, cfg.save_dir + 'best.ckpt')
best_acc = over_acc
best_step = running_count
fshow = open(
cfg.save_dir + 'acc: ' + str(best_acc) + ', step: ' +
str(best_step), 'w')
print('>>> best acc: ', best_acc, 'best step: ', best_step)
# per-class accuracy
acc = np.diag(hist) / hist.sum(0)
print('>>> Step', running_count, 'mean accuracy', acc)
last_acc_time = now
stepes.append(running_count)
Acc_val_list.append(over_acc)
# draw plots for visualization ----------------------------
# Plot the figures per 60s
import matplotlib.pyplot as plt
fig1, ax1 = plt.subplots(figsize=(11, 8))
ax1.plot(range(start_step, running_count), lossTr_list)
ax1.set_title("Average training loss vs steps")
ax1.set_xlabel("Steps")
ax1.set_ylabel("Current loss")
plt.savefig(cfg.save_dir + "loss_vs_steps.png")
plt.clf()
fig2, ax2 = plt.subplots(figsize=(11, 8))
ax2.plot(stepes, Acc_val_list, label="Val total acc.")
ax2.set_title(" Acc vs steps")
ax2.set_xlabel("Steps")
ax2.set_ylabel("Current Acc")
plt.legend(loc='lower right')
plt.savefig(cfg.save_dir + "acc_vs_steps.png")
plt.close('all')
# ----------------------------------------------------------
# Save step model
if cfg.is_save_step and (running_count % cfg.save_step_inter) == 0 \
and running_count >= cfg.start_save_step:
saver.save(sess, cfg.save_dir + 'model.ckpt',
int(running_count))
print('Model has been saved:' + str(running_count))
files = os.path.join(cfg.save_dir +
'model.ckpt-*.data-00000-of-00001')
sfile = glob.glob(files)
if len(sfile) > cfg.model_save_num:
steps = []
for s in sfile:
part = s.split('.')
re = int(part[1].split('-')[1])
steps.append(re)
re = min(steps)
model = cfg.save_dir + 'model.ckpt-' + str(re)
os.remove(model + '.data-00000-of-00001')
os.remove(model + '.index')
os.remove(model + '.meta')
print('Remove Model:' + model)
# Check accuracy per Epoch of training data
if cfg.is_epoch_acc and running_count >= cfg.start_show_iter \
and running_count <= cfg.total_iter:
# Test accuracy in val
hist = np.zeros((cfg.NUM_OF_CLASSESS, cfg.NUM_OF_CLASSESS))
for i, img_name in enumerate(val_img_list):
true_val = np.expand_dims(misc.imread(val_label_list[i]),
axis=2)
pred_val = evaluate_accuracy(val_logits_softmax, sess,
val_image_batch, img_name)
hist += fast_hist(pred_val.flatten(), true_val.flatten(),
cfg.NUM_OF_CLASSESS)
hist[:, 0] = 0
# overall accuracy
over_acc = np.diag(hist).sum() / hist.sum()
print('>>> Step', running_count, 'overall accuracy', over_acc)
if over_acc > best_acc:
saver.save(sess, cfg.save_dir + 'best.ckpt')
best_acc = over_acc
best_step = running_count
fshow = open(
cfg.save_dir + 'acc: ' + str(best_acc) + ', step: ' +
str(best_step), 'w')
print('>>> best acc: ', best_acc, 'best step: ', best_step)
# per-class accuracy
acc = np.diag(hist) / hist.sum(0)
print('>>> Step', running_count, 'mean accuracy', acc)
epo += 1
# Save epoch model
if cfg.is_save_epoch and (epo % cfg.save_epoch_inter
) == 0 and epo >= cfg.start_save_epoch:
saver.save(sess, cfg.save_dir + 'model.ckpt', epo)
print('Model has been saved:' + str(epo))
files = os.path.join(cfg.save_dir +
'model.ckpt-*.data-00000-of-00001')
sfile = glob.glob(files)
if len(sfile) > cfg.model_save_num:
steps = []
for s in sfile:
part = s.split('.')
re = int(part[1].split('-')[1])
steps.append(re)
re = min(steps)
model = cfg.save_dir + 'model.ckpt-' + str(re)
os.remove(model + '.data-00000-of-00001')
os.remove(model + '.index')
os.remove(model + '.meta')
print('Remove Model:' + model)
time_end = time.time()
print('Epo ' + str(epo) + ' use time: ' + str(time_end - time_start))
# saver.save(sess, cfg.save_dir + 'last.ckpt') # save last model
train_end_time = time.time()
print('Train total use: ' +
str((train_end_time - train_start_time) / 3600) + ' h')
coord.request_stop()
coord.join(threads)
def evaluate_checkpoint(model_path, args):
coord = tf.train.Coordinator()
tf.reset_default_graph()
reader = ImageReader(args.data_list,
INPUT_SIZE,
random_scale=False,
random_mirror=False,
ignore_label=IGNORE_LABEL,
img_mean=IMG_MEAN,
coord=coord,
train=False)
image_batch, label_batch = reader.dequeue(args.batch_size)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Create network.
net = ICNet_BN({'data': image_batch},
is_training=False,
num_classes=num_classes)
# Which variables to load.
restore_var = tf.global_variables()
# Predictions.
raw_output = net.layers['conv6_cls']
raw_output_up = tf.image.resize_bilinear(raw_output,
size=INPUT_SIZE,
align_corners=True)
raw_output_up = tf.argmax(raw_output_up, dimension=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# mIoU
pred_flatten = tf.reshape(pred, [
-1,
])
raw_gt = tf.reshape(label_batch, [
-1,
])
if args.ignore_zero:
indices = tf.squeeze(
tf.where(
tf.logical_and(tf.less_equal(raw_gt, num_classes - 1),
tf.greater(raw_gt, 0)), ), 1)
else:
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, num_classes - 1)),
1)
#indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
pred = tf.gather(pred_flatten, indices)
metric, op = tf.contrib.metrics.streaming_mean_iou(pred,
gt,
num_classes=num_classes)
mIoU, update_op = metric, op
# Summaries
miou_op = tf.summary.scalar('mIOU', mIoU)
start = time.time()
logging.info('Starting evaluation at ' +
time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime()))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver = tf.train.Saver(var_list=restore_var)
load(saver, sess, model_path)
for step in range(num_steps):
preds, _ = sess.run([pred, update_op])
if step % 500 == 0:
print('Finish {0}/{1}'.format(step + 1, num_steps))
iou, summ = sess.run([mIoU, miou_op])
sess.close()
coord.request_stop()
#coord.join(threads)
return summ, iou
def main():
tf.set_random_seed(1234)
temp_flags = FLAGS.__flags.items()
temp_flags.sort()
for params, value in FLAGS.__flags.items():
print('{}: {}'.format(params, value))
coord = tf.train.Coordinator()
tf.reset_default_graph()
input_size = [FLAGS.image_height, FLAGS.image_width]
reader = ImageReader(FLAGS.data_dir, FLAGS.data_list, input_size, None,
None, FLAGS.ignore_label, IMG_MEAN, coord)
image, label = reader.image, reader.label
image_batch, label_batch = tf.expand_dims(image, dim=0), tf.expand_dims(
label, dim=0) # Add one batch dimension.
print(image_batch)
net = MobileNet(image_batch, isTraining=False, updateBeta=False)
with tf.variable_scope('', reuse=True):
flipped_img = tf.image.flip_left_right(image)
flipped_img = tf.expand_dims(flipped_img, dim=0)
net2 = MobileNet(flipped_img, isTraining=False, updateBeta=False)
# Which variables to load.
restore_var = tf.global_variables()
raw_output = net
if FLAGS.flipped_eval:
flipped_output = tf.image.flip_left_right(tf.squeeze(net2))
flipped_output = tf.expand_dims(flipped_output, dim=0)
raw_output = tf.add_n([raw_output, flipped_output])
raw_output_up = tf.image.resize_bilinear(raw_output,
size=input_size,
align_corners=True)
raw_output_up = tf.argmax(raw_output_up, dimension=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# mIoU
pred_flatten = tf.reshape(pred, [
-1,
])
raw_gt = tf.reshape(label_batch, [
-1,
])
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, FLAGS.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
pred = tf.gather(pred_flatten, indices)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(
pred, gt, num_classes=FLAGS.num_classes)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
sess.run(tf.local_variables_initializer())
restore_var = tf.global_variables()
load(sess, FLAGS.checkpoint_path, restore_var)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for step in range(1, FLAGS.num_steps + 1):
preds, _ = sess.run([pred, update_op])
# if step > 0 and FLAGS.measure_time:
# calculate_time(sess, net)
if FLAGS.print_each_step and step % 100 == 0:
print('Finish {0}/{1}'.format(step, FLAGS.num_steps))
print('step {0} mIoU: {1}'.format(step, sess.run(mIoU)))
value = sess.run(mIoU)
print('step {0} mIoU: {1}'.format(step, value))
epoch = int(os.path.basename(FLAGS.checkpoint_path).split('-')[1])
writeToLogFile(epoch, value)
coord.request_stop()
coord.join(threads)
def detect_vessels(self, additional_mask=None):
im = self._image
mask = self._mask
# green channel & remove light reflex
im_gray = remove_light_reflex(im)
# CLAHE
clahe = cv2.createCLAHE(3, (7, 7))
im_clahe = clahe.apply(im_gray)
# Haar transform
wp = pywt.WaveletPacket2D(im_clahe,
wavelet='haar',
maxlevel=1,
mode='sym')
wp.decompose()
im_haar = wp['a'].data
im_haar = remove_light_reflex(im_haar)
mask = ImageReader.rescale_mask(im_haar, self.mask)
try:
hasAdditionalMask = additional_mask.size > 0
additional_mask = ImageReader.rescale_mask(im_haar,
additional_mask)
except AttributeError:
hasAdditionalMask = False
# compute the image mean to "invert" the image
# for the gausssian matched filter
mean_thresh, _, _, _ = cv2.mean(im_haar, mask)
if hasAdditionalMask:
im_haar[additional_mask != 0] = mean_thresh
im_haar = mean_thresh - im_haar
# Gaussian Matched and FDOG filter responses
K_MF = gaussian_matched_filter_kernel(31, 5)
K_FDOG = fdog_filter_kernel(31, 5)
kernels_mf = createMatchedFilterBank(K_MF, 12)
kernels_fdog = createMatchedFilterBank(K_FDOG, 12)
im_matched_mf = applyFilters(im_haar, kernels_mf)
im_matched_fdog = applyFilters(im_haar, kernels_fdog)
# normalize local mean of FDOG response
local_mean_fdog = cv2.blur(im_matched_fdog, (11, 11))
local_mean_fdog = cv2.normalize(local_mean_fdog,
alpha=0,
beta=1,
norm_type=cv2.NORM_MINMAX)
# set the threshold matrix
mean_thresh, _, _, _ = cv2.mean(im_matched_mf, mask)
ref_thresh = mean_thresh * self._norm_const
ref_thresh = ref_thresh * (1 + local_mean_fdog)
# show the results
if self._is_debug:
show_images([self._reader.image])
show_images([im_gray, im_haar], titles=["gray", "haar"])
show_images([im_matched_mf, im_matched_fdog],
titles=["mf", "fdog"])
im_vessels = im_matched_mf.copy()
im_vessels[im_vessels < ref_thresh] = 0
im_vessels[mask == 0] = 0
im_vessels[im_vessels != 0] = 1
im_vessels = im_vessels.astype('uint8')
if self._is_debug:
show_images([im_vessels], titles=["vessels"])
self._im_norm = mh.stretch(im_matched_mf)
return ImageReader.rescale_mask(self.image, im_vessels)
def main():
# Create model and start training
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
tf.set_random_seed(args.random_seed)
# Create queue coordinator
coord = tf.train.Coordinator()
# Load Image Reader
with tf.name_scope('create_inputs'):
reader = ImageReader(
args.index_loc,
args.data_dir,
args.mask_dir,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
mode = tf.contrib.learn.ModeKeys.TRAIN
net = DeepLabResNetModel(image_batch, mode, args.num_classes, args.atrous_blocks)
raw_output = net.output
# Trainable Variables
restore_vars = [v for v in tf.global_variables() if 'fc' not in v.name or not args.not_restore_last]
all_trainable = [v for v in tf.trainable_variables() if 'beta' not in v.name and 'gamma' not in v.name]
fc_trainable = [v for v in all_trainable if 'fc' in v.name]
conv_trainable = [v for v in all_trainable if 'fc' not in v.name]
fc_w_trainable = [v for v in fc_trainable if 'weights' in v.name]
fc_b_trainable = [v for v in fc_trainable if 'biases' in v.name]
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, args.num_classes])
label_proc = prepare_labels(label_batch, tf.stack(raw_output.get_shape()[1:3]), num_classes=args.num_classes, one_hot=False)
raw_gt = tf.reshape(label_proc, [-1,])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, args.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise Softmax Loss
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction, labels=gt)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
variable_summaries(reduced_loss, name='loss')
# Processed predictions: for visualization
raw_output_up = tf.image.resize_bilinear(raw_output, tf.shape(image_batch)[1:3,])
raw_output_up = tf.argmax(raw_output_up, dimension=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# Define loss and optimization parameters
base_lr = tf.constant(args.learning_rate, tf.float64)
global_step = tf.Variable(0, trainable=False, name='global_step')
increment_step = tf.assign(global_step, global_step + 1)
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - global_step / args.num_steps), args.power))
learning_rate = tf.maximum(learning_rate, 8e-7)
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
opt_fc_w = tf.train.MomentumOptimizer(learning_rate * 5.0, args.momentum)
opt_fc_b = tf.train.MomentumOptimizer(learning_rate * 10.0, args.momentum)
grads = tf.gradients(reduced_loss, conv_trainable + fc_w_trainable + fc_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_fc_w = grads[len(conv_trainable) : (len(conv_trainable) + len(fc_w_trainable))]
grads_fc_b = grads[(len(conv_trainable) + len(fc_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(zip(grads_conv, conv_trainable))
train_op_fc_w = opt_fc_w.apply_gradients(zip(grads_fc_w, fc_w_trainable))
train_op_fc_b = opt_fc_b.apply_gradients(zip(grads_fc_b, fc_b_trainable))
train_op = tf.group(increment_step, train_op_conv, train_op_fc_w, train_op_fc_b)
# initial_learning_rate = 1e-2
# learning_rate = tf.train.exponential_decay(initial_learning_rate, global_step, 300, 0.96)
# adam = tf.train.AdamOptimizer(learning_rate).minimize(reduced_loss, global_step=global_step)
# Image Summary
model_dir = args.snapshot_dir + args.model_name
images_summary = tf.py_func(inv_preprocess, [image_batch, args.save_num_images, IMG_MEAN], tf.uint8)
preds_summary = tf.py_func(decode_labels, [pred, args.save_num_images, args.num_classes], tf.uint8)
labels_summary = tf.py_func(decode_labels, [label_batch, args.save_num_images, args.num_classes], tf.uint8)
image_summaries = [images_summary, preds_summary, labels_summary]
image_summary = tf.summary.image('images',
tf.concat(axis=2, values=image_summaries),
max_outputs=args.save_num_images)
# Variable Summary
variable_summaries(fc_w_trainable, 'fc_w')
variable_summaries(fc_b_trainable, 'fc_b')
variable_summaries(learning_rate, 'learning_rate')
# variable_summaries(net.weights, 'aconv_w')
# variable_summaries(net.biases, 'aconv_b')
total_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(model_dir, graph=tf.get_default_graph())
# Set up session
with tf.Session() as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(max_to_keep=3)
if args.snapshot_dir is not None and args.model_name is not None and os.path.exists(model_dir):
loader = tf.train.Saver()
load_model(loader, sess, model_dir)
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# train_op = adam
for step in range(args.num_steps):
start_time = time.time()
if step % args.save_pred_every == 0:
feed = [reduced_loss, image_batch, label_batch, pred, total_summary, global_step, train_op]
loss_value, images, labels, preds, summary, total_steps, _ = sess.run(feed)
summary_writer.add_summary(summary, total_steps)
save_model(saver, sess, model_dir, global_step)
else:
feed = [reduced_loss, global_step, train_op]
loss_value, total_steps, _ = sess.run(feed)
duration = time.time() - start_time
results = 'global step: {:d}, step: {:d} \t loss = {:.3f}, ({:.3f} secs)'\
.format(total_steps, step, loss_value, duration)
if step % WRITE_EVERY == 0:
with open(WRITE_FILE, 'a') as f:
f.write(results + '\n')
print(results)
coord.request_stop()
coord.join(threads)
def evaluate(config,
evaluation_set='val',
determine_confusionMatrix=True,
plot_confusionMatrix=False):
# --------------------------------------------------------------------
# init network
# --------------------------------------------------------------------
tf.compat.v1.reset_default_graph()
# define input placeholders
input_placeholder = {}
input_placeholder.update(
{'is_training': tf.compat.v1.placeholder(dtype=tf.bool, shape=())})
if config.ARCHITECTURE == 'semantic_segmentation':
batch_size = config.BATCH_SIZE * config.TIMESEQUENCE_LENGTH
print('batch_size: {}'.format(config.BATCH_SIZE))
# Search for available GPUs: the result is a list of device ids like `['/gpu:0', '/gpu:1']`
devices = get_available_gpus()
print("found devices: ", devices)
num_GPU = len(devices)
if (num_GPU) == 0:
num_GPU = 1 # CPU support!
# min 1 sample should be applied on a GPU
if (config.BATCH_SIZE < num_GPU):
num_GPU = config.BATCH_SIZE
image_placeholder = []
label_placeholder = []
for iter in range(num_GPU):
if (iter == (num_GPU - 1)):
batch_size_local = batch_size - (num_GPU - 1) * (batch_size //
num_GPU)
else:
batch_size_local = batch_size // num_GPU
print('batch_size /gpu:{} : {}'.format(iter, num_GPU))
image_placeholder.append(
tf.compat.v1.placeholder(
dtype=tf.float32,
shape=(batch_size_local,
config.DATASET_TRAIN.INPUT_SIZE[0],
config.DATASET_TRAIN.INPUT_SIZE[1],
config.DATASET_TRAIN.NUM_CHANNELS)))
label_placeholder.append(
tf.compat.v1.placeholder(
dtype=tf.float32,
shape=(batch_size_local,
config.DATASET_TRAIN.INPUT_SIZE[0],
config.DATASET_TRAIN.INPUT_SIZE[1], 1)))
input_placeholder.update({'image_batch': image_placeholder})
input_placeholder.update({'label_batch': label_placeholder})
else:
print(
'[ERROR] network architecture does not exist!!! Please check your spelling!'
)
raise NotImplementedError
# load network architecture
if config.ARCHITECTURE == 'semantic_segmentation':
model = get_model(config.MODEL)
net = model(
{
'data': input_placeholder['image_batch'],
'is_training': input_placeholder['is_training']
},
is_training=input_placeholder['is_training'],
evaluation=tf.logical_not(input_placeholder['is_training']),
num_classes=config.DATASET_TRAIN.NUM_CLASSES,
filter_scale=config.FILTER_SCALE,
timeSequence=config.TIMESEQUENCE_LENGTH,
variant=config.MODEL_VARIANT)
else:
print(
'[ERROR] network architecture does not exist!!! Please check your spelling!'
)
raise NotImplementedError
# --------------------------------------------------------------------
# determine evaluation metric
# --------------------------------------------------------------------
if config.ARCHITECTURE == 'semantic_segmentation':
list_raw_gt = []
list_pred_flattern_mIoU = []
for iter_gpu in range(len(input_placeholder['image_batch'])):
with tf.device('/gpu:%d' % iter_gpu):
if config.MODEL == 'SegNet_BN' or config.MODEL == 'SegNet_BN_encoder' or config.MODEL == 'SegNet_BN_decoder' or config.MODEL == 'SegNet_BN_encoderDecoder':
raw_output = net.layers['output'][iter_gpu]
raw_output_up = tf.argmax(raw_output,
axis=3,
output_type=tf.int32)
raw_pred_mIoU = tf.expand_dims(raw_output_up, dim=3)
else: # ICNet
ori_shape = config.DATASET_TRAIN.INPUT_SIZE #??
raw_output = net.layers['output'][iter_gpu]
raw_output_up = tf.compat.v1.image.resize_bilinear(
raw_output, size=ori_shape[:2], align_corners=True)
raw_output_up = tf.argmax(raw_output_up,
axis=3,
output_type=tf.int32)
raw_pred_mIoU = tf.expand_dims(raw_output_up, dim=3)
# determine mIoU
if config.USAGE_TIMESEQUENCES: # evaluate only last image of time sequence
pred_of_interest = np.array(
range(config.BATCH_SIZE), dtype=np.int32
) * config.TIMESEQUENCE_LENGTH + config.TIMESEQUENCE_LENGTH - 1
pred_flatten_mIoU = tf.reshape(
tf.gather(raw_pred_mIoU, pred_of_interest), [
-1,
])
raw_gt = tf.reshape(
tf.gather(input_placeholder['label_batch'][iter_gpu],
pred_of_interest), [
-1,
])
else: # evaluate all images of batch size
pred_flatten_mIoU = tf.reshape(raw_pred_mIoU, [
-1,
])
raw_gt = tf.reshape(
input_placeholder['label_batch'][iter_gpu], [
-1,
])
list_raw_gt.append(raw_gt)
list_pred_flattern_mIoU.append(pred_flatten_mIoU)
# combine output of different GPUs
with tf.device('/gpu:%d' % 0):
all_raw_gt = tf.reshape(tf.concat(list_raw_gt, -1), [
-1,
])
all_pred_flatten_mIoU = tf.reshape(
tf.concat(list_pred_flattern_mIoU, -1), [
-1,
])
indices_mIoU = tf.squeeze(
tf.where(
tf.less_equal(raw_gt,
config.DATASET_TRAIN.NUM_CLASSES - 1)), 1)
gt_mIoU = tf.cast(tf.gather(raw_gt, indices_mIoU), tf.int32)
pred_mIoU = tf.gather(pred_flatten_mIoU, indices_mIoU)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(
pred_mIoU, gt_mIoU, num_classes=config.DATASET_VAL.NUM_CLASSES)
# deterimine confusing matrix
if determine_confusionMatrix:
# Create an accumulator variable to hold the counts
confusion = tf.Variable(tf.zeros([
config.DATASET_VAL.NUM_CLASSES, config.DATASET_VAL.NUM_CLASSES
],
dtype=tf.int64),
name='confusion',
collections=[
tf.compat.v1.GraphKeys.LOCAL_VARIABLES
])
# Compute a per-batch confusion
batch_confusion = tf.math.confusion_matrix(
tf.reshape(gt_mIoU, [-1]),
tf.reshape(pred_mIoU, [-1]),
num_classes=config.DATASET_VAL.NUM_CLASSES,
name='batch_confusion')
# Create the update op for doing a "+=" accumulation on the batch
confusion_update = confusion.assign(
confusion + tf.cast(batch_confusion, dtype=tf.int64))
# -----------------------------------------
# init session
# -----------------------------------------
# Set up tf session and initialize variables.
sessConfig = tf.compat.v1.ConfigProto()
sessConfig.gpu_options.allow_growth = True
sessConfig.gpu_options.per_process_gpu_memory_fraction = 0.5
sess = tf.compat.v1.Session(config=sessConfig)
init = tf.compat.v1.global_variables_initializer()
local_init = tf.compat.v1.local_variables_initializer()
sess.run(init)
sess.run(local_init)
# load checkpoint file
print(config.EVALUATION.MODELPATH)
ckpt = tf.compat.v1.train.get_checkpoint_state(config.EVALUATION.MODELPATH)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.compat.v1.train.Saver(
var_list=tf.compat.v1.global_variables())
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
# --------------------------------------------------------------------
# Evaluate - Iterate over training steps.
# --------------------------------------------------------------------
# evaluate training or validation set
if evaluation_set == "val":
imagereader_val = ImageReader(config.IMAGEREADER.VAL,
config.DATASET_VAL, config.BATCH_SIZE,
config.TIMESEQUENCE_LENGTH)
elif evaluation_set == "train":
imagereader_val = ImageReader(config.IMAGEREADER.VAL,
config.DATASET_TRAIN, config.BATCH_SIZE,
config.TIMESEQUENCE_LENGTH)
elif evaluation_set == "test":
imagereader_val = ImageReader(config.IMAGEREADER.VAL,
config.DATASET_TEST, config.BATCH_SIZE,
config.TIMESEQUENCE_LENGTH)
elif evaluation_set == "all":
imagereader_val = ImageReader(config.IMAGEREADER.VAL,
config.DATASET_ALL, config.BATCH_SIZE,
config.TIMESEQUENCE_LENGTH)
else:
print("Dataset {} does not exist!".format(evaluation_set))
acc_value = 0.0
# --------------------------------------
# perform evaluation - semantic segmentation
# --------------------------------------
if config.ARCHITECTURE == 'semantic_segmentation':
if config.TIMESEQUENCES_SLIDINGWINDOW: # use time sequences
for step in trange(
int(imagereader_val._dataset_amount -
config.BATCH_SIZE * config.TIMESEQUENCE_LENGTH + 1),
desc='evaluation',
leave=True):
start_time = time.time()
training_batch = imagereader_val.getNextMinibatch()
feed_dict = {input_placeholder['is_training']: False}
for iter_GPU in range(len(input_placeholder['image_batch'])):
num_GPU = len(input_placeholder['image_batch'])
batch_size = training_batch['blob_data'].shape[0]
batch_size_local = batch_size // num_GPU
if (iter_GPU == (num_GPU - 1)):
batch_size_act = batch_size - (num_GPU - 1) * (
batch_size // num_GPU)
else:
batch_size_act = batch_size // num_GPU
feed_dict.update({
input_placeholder['image_batch'][iter_GPU]:
training_batch['blob_data'][iter_GPU *
batch_size_local:iter_GPU *
batch_size_local +
batch_size_act, :, :, :],
input_placeholder['label_batch'][iter_GPU]:
training_batch['blob_label']
[iter_GPU *
batch_size_local:iter_GPU * batch_size_local +
batch_size_act, :, :, :]
})
duration = time.time() - start_time
else: # do not use time sequences (normal evaluation)
for step in trange(
int(imagereader_val._dataset_amount /
(config.BATCH_SIZE * config.TIMESEQUENCE_LENGTH)),
desc='evaluation',
leave=True):
start_time = time.time()
training_batch = imagereader_val.getNextMinibatch()
feed_dict = {input_placeholder['is_training']: False}
for iter_GPU in range(len(input_placeholder['image_batch'])):
num_GPU = len(input_placeholder['image_batch'])
batch_size = training_batch['blob_data'].shape[0]
batch_size_local = batch_size // num_GPU
if (iter_GPU == (num_GPU - 1)):
batch_size_act = batch_size - (num_GPU - 1) * (
batch_size // num_GPU)
else:
batch_size_act = batch_size // num_GPU
feed_dict.update({
input_placeholder['image_batch'][iter_GPU]:
training_batch['blob_data'][iter_GPU *
batch_size_local:iter_GPU *
batch_size_local +
batch_size_act, :, :, :],
input_placeholder['label_batch'][iter_GPU]:
training_batch['blob_label']
[iter_GPU *
batch_size_local:iter_GPU * batch_size_local +
batch_size_act, :, :, :]
})
if determine_confusionMatrix:
sess.run([update_op, confusion_update],
feed_dict=feed_dict)
else:
sess.run([update_op], feed_dict=feed_dict)
duration = time.time() - start_time
mIoU_value = sess.run(mIoU)
if determine_confusionMatrix:
confusion_matrix = sess.run(confusion)
# print Accuracy:
np.set_printoptions(linewidth=np.inf) #150)
acc_value = float(np.sum(np.diag(confusion_matrix))) / float(
np.sum(confusion_matrix))
else:
print(
'[ERROR] network architecture does not exist!!! Please check your spelling!'
)
raise NotImplementedError
# --------------------------------------------
# close session
# --------------------------------------------
sess.close()
tf.compat.v1.reset_default_graph()
# --------------------------------------------------------------------
# Show results
# --------------------------------------------------------------------
if determine_confusionMatrix and config.ARCHITECTURE == 'semantic_segmentation':
# determine class-wise IoU
buff = 0.0
print('-------------------------------------------------------------')
for iter in xrange(config.DATASET_VAL.NUM_CLASSES):
if np.sum(
confusion_matrix[iter, :]) == 0: # avoid division by zero
IoU = 0.0
else:
IoU = 100.0 * confusion_matrix[iter, iter] / (
np.sum(confusion_matrix[iter, :]) +
np.sum(confusion_matrix[:, iter]) -
confusion_matrix[iter, iter])
buff = buff + IoU
print('{}: {}'.format(config.DATASET_VAL.CLASSES[iter], IoU))
print('-------------------------------------------------------------')
print('dataset: {} - {}'.format(config.DATASET_NAME,
config.DATASET_WEATHER))
print('Accuracy: {}'.format(acc_value))
print('mIoU: {}'.format(mIoU_value))
print('-------------------------------------------------------------')
if plot_confusionMatrix and config.ARCHITECTURE == 'semantic_segmentation':
print('-------------------------------------------------------------')
print(confusion_matrix)
print('-------------------------------------------------------------')
plot_confusion_matrix(confusion_matrix, config.DATASET_VAL.CLASSES)
return mIoU_value, acc_value
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.estimator import regression
from tflearn.data_preprocessing import ImagePreprocessing
from tflearn.data_augmentation import ImageAugmentation
# Data loading and preprocessing
from tflearn.datasets import cifar10
# (X, Y), (X_test, Y_test) = cifar10.load_data()
label_num = 5
n_epoch = 500
# n_epoch = 5
batch_size = 96
image_size = (32, 32)
(X, Y), (X_test, Y_test) = ImageReader.read_train_test_images_labels(
'../imgs/faces02/trimmed', max_label=label_num, resize=image_size)
# import ipdb
# ipdb.set_trace()
X, Y = shuffle(X, Y)
Y = to_categorical(Y, label_num)
Y_test = to_categorical(Y_test, label_num)
# Real-time data preprocessing
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()
# Real-time data augmentation
img_aug = ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_rotation(max_angle=25.)
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
reader = ImageReader(
args.data_list,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
net = ICNet_BN({'data': image_batch}, is_training = True, num_classes = args.num_classes)
sub4_out = net.layers['sub4_out']
sub24_out = net.layers['sub24_out']
sub124_out = net.layers['conv6']
fc_list = ['conv6']
restore_var = tf.global_variables()
all_trainable = [v for v in tf.trainable_variables() if ('beta' not in v.name and 'gamma' not in v.name) or args.train_beta_gamma]
restore_var = [v for v in tf.global_variables() if not (len([f for f in fc_list if f in v.name])) or not args.not_restore_last]
for v in restore_var:
print(v.name)
loss_sub4 = create_loss(sub4_out, label_batch, args.num_classes, args.ignore_label, args.use_class_weights)
loss_sub24 = create_loss(sub24_out, label_batch, args.num_classes, args.ignore_label, args.use_class_weights)
loss_sub124 = create_loss(sub124_out, label_batch, args.num_classes, args.ignore_label, args.use_class_weights)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
loss = LAMBDA1 * loss_sub4 + LAMBDA2 * loss_sub24 + LAMBDA3 * loss_sub124
reduced_loss = loss + tf.add_n(l2_losses)
##############################
# visualization and summary
##############################
# Processed predictions: for visualisation.
# Sub 4
raw_output_up4 = tf.image.resize_bilinear(sub4_out, tf.shape(image_batch)[1:3,])
raw_output_up4 = tf.argmax(raw_output_up4, dimension = 3)
pred4 = tf.expand_dims(raw_output_up4, dim = 3)
# Sub 24
raw_output_up24 = tf.image.resize_bilinear(sub24_out, tf.shape(image_batch)[1:3,])
raw_output_up24 = tf.argmax(raw_output_up24, dimension=3)
pred24 = tf.expand_dims(raw_output_up24, dim=3)
# Sub 124
raw_output_up124 = tf.image.resize_bilinear(sub124_out, tf.shape(image_batch)[1:3,])
raw_output_up124 = tf.argmax(raw_output_up124, dimension=3)
pred124 = tf.expand_dims(raw_output_up124, dim=3)
images_summary = tf.py_func(inv_preprocess, [image_batch, SAVE_NUM_IMAGES, IMG_MEAN], tf.uint8)
labels_summary = tf.py_func(decode_labels, [label_batch,SAVE_NUM_IMAGES, args.num_classes], tf.uint8)
preds_summary4 = tf.py_func(decode_labels, [pred4, SAVE_NUM_IMAGES, args.num_classes], tf.uint8)
preds_summary24 = tf.py_func(decode_labels, [pred24, SAVE_NUM_IMAGES, args.num_classes], tf.uint8)
preds_summary124 = tf.py_func(decode_labels, [pred124, SAVE_NUM_IMAGES, args.num_classes], tf.uint8)
total_images_summary = tf.summary.image('images',
tf.concat(axis=2, values=[images_summary, labels_summary, preds_summary124]),
max_outputs=SAVE_NUM_IMAGES) # Concatenate row-wise.
total_summary = total_images_summary
loss_summary = tf.summary.scalar('Total_loss', reduced_loss)
#total_summary.append(loss_summary)
summary_writer = tf.summary.FileWriter(args.snapshot_dir,
graph=tf.get_default_graph())
##############################
##############################
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.train.exponential_decay(base_lr, step_ph, args.num_steps, args.power)
lr_summary = tf.summary.scalar('Learning_rate', learning_rate)
#total_summary.append(lr_summary)
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
if args.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
opt_conv = tf.train.AdamOptimizer(learning_rate)
#opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
grads = tf.gradients(reduced_loss, all_trainable)
train_op = opt_conv.apply_gradients(zip(grads, all_trainable))
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list = tf.global_variables(), max_to_keep = 10)
ckpt = tf.train.get_checkpoint_state(args.snapshot_dir)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('Restore from pre-trained model...')
#net.load(args.restore_from, sess, ignore_layers = fc_list)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
#summ_op = tf.summary.merge(total_summary)
# Iterate over training steps.
save_summary_every = 20
for step in range(args.num_steps):
start_time = time.time()
if LR_SHEDULE != {}:
if step >= LR_SHEDULE.keys()[0]:
tf.assign(learning_rate, LR_SHEDULE.popitem()[0])
feed_dict = {step_ph: step}
if not (step % args.save_pred_every == 0):
loss_value, loss1, loss2, loss3, _, lr_sum, l_sum = \
sess.run([reduced_loss, loss_sub4,
loss_sub24, loss_sub124, train_op, lr_summary, loss_summary], feed_dict=feed_dict)
else:
save(saver, sess, args.snapshot_dir, step)
loss_value, loss1, loss2, loss3, _, lr_sum, l_sum, t_sum = \
sess.run([reduced_loss, loss_sub4,
loss_sub24, loss_sub124, train_op, lr_summary, loss_summary, total_summary], feed_dict=feed_dict)
summary_writer.add_summary(t_sum, step)
if step % save_summary_every == 0:
summary_writer.add_summary(lr_sum, step)
summary_writer.add_summary(l_sum, step)
duration = time.time() - start_time
print('step {:d} \t total loss = {:.3f}, sub4 = {:.3f}, sub24 = {:.3f}, sub124 = {:.3f} ({:.3f} sec/step)'.format(step, loss_value, loss1, loss2, loss3, duration))
coord.request_stop()
coord.join(threads)
"""
author: A Skrzypek
"""
from image_reader import ImageReader
from make_plots import ImagePlot
import numpy as np
if __name__ == "__main__":
# read images from path
path_to_test_sample = r"C:\Users\user\Documents\IMAGE_PROJECT\sample_database\jpg_samples"
sample_image_obj = ImageReader(path_to_test_sample)
sample_image_obj.main()
sample_plot_obj = ImagePlot(sample_image_obj.images_storage)
sample_plot_obj.main()
pass
def _tf_common_init(self):
gpu_count = len(self.device_ids)
src_size = self.config['input_size']
input_size_wh = (src_size['width'], src_size['height'])
init_lr = self.config['lr']
power = self.config['lr_decreasing']['power']
momentum = self.config['momentum']
filter_scale = self.config['filter_scale']
weight_decay = self.config['weight_decay']
num_classes = len(self.out_classes)
train_beta_gamma = self.config['train_beta_gamma']
update_mean_var = self.config['update_mean_var']
with tf.device('/cpu:0'):
self.coord = tf.train.Coordinator()
splitted_images = {}
splitted_labels = {}
with tf.name_scope("create_inputs"):
for name, need_shuffle in [
('train', True),
('val', False),
]:
reader = ImageReader(
ia_descrs=self.samples_dct[name],
input_size_wh=input_size_wh,
random_scale=False,
random_mirror=False,
img_mean=IMG_MEAN,
coord=self.coord,
in_pr_meta=self.helper.in_project_meta,
class_to_idx=self.class_title_to_idx,
shuffle=need_shuffle)
batch_sz = self.config['batch_size'][name]
img_batch, lbl_batch = reader.dequeue(batch_sz * gpu_count)
split_images = tf.split(img_batch, gpu_count, 0)
split_labels = tf.split(lbl_batch, gpu_count, 0)
splitted_images[name] = split_images
splitted_labels[name] = split_labels
global_step = tf.get_variable(
'global_step', [],
initializer=tf.constant_initializer(0),
trainable=False,
dtype=tf.int32)
self.tf_label = tf.placeholder(dtype=tf.int32) # , shape=[None])
self.tf_prediction = tf.placeholder(
dtype=tf.int32) # , shape=[None])
self.tf_metric, self.tf_metric_update = tf.metrics.accuracy(
self.tf_label, self.tf_prediction, name="use_metric_acc")
running_vars = tf.get_collection(tf.GraphKeys.LOCAL_VARIABLES,
scope="use_metric_acc")
self.running_vars_initializer = tf.variables_initializer(
var_list=running_vars)
base_lr = tf.constant(init_lr)
self.step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(
base_lr,
tf.pow((1 - self.step_ph / self.total_train_iters), power))
opt = tf.train.MomentumOptimizer(learning_rate, momentum)
all_grads = []
losses = []
with tf.variable_scope(tf.get_variable_scope()):
for curr_dev_id in self.device_ids:
with tf.device('/gpu:{}'.format(curr_dev_id)):
with tf.name_scope(
'clone_{}'.format(curr_dev_id)) as scope:
spl_img = splitted_images['train'][curr_dev_id]
spl_lbl = splitted_labels['train'][curr_dev_id]
net = get_model(spl_img, num_classes, filter_scale)
reduced_loss, internal_losses, prediction, gt, self.v1, self.v2 = \
forward_and_all_losses(
net, spl_lbl, num_classes, self.neutral_input_idx, weight_decay
)
# loss_sub4, loss_sub24, loss_sub124 = internal_losses
losses.append(reduced_loss)
tf.get_variable_scope().reuse_variables()
grads = get_grads(reduced_loss, train_beta_gamma,
update_mean_var)
all_grads.append(grads)
self.total_loss = tf.stack(values=losses)
self.total_loss = tf.reduce_mean(self.total_loss)
mean_grads = average_gradients(all_grads)
all_trainable = get_trainable_vars(train_beta_gamma)
# Apply the gradients to adjust the shared variables.
self.train_op = opt.apply_gradients(zip(mean_grads, all_trainable),
global_step=global_step)
# # Group all updates to into a single train op.
# train_op = tf.group(apply_gradient_conv_op, apply_gradient_fc_w_op,
# apply_gradient_fc_b_op)
self.total_val_loss, self.v1_val, self.v2_val = get_val_loss(
splitted_images['val'], splitted_labels['val'], num_classes,
weight_decay, self.device_ids, filter_scale,
self.neutral_input_idx)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
config.log_device_placement = False
self.sess = tf.Session(config=config)
init = tf.global_variables_initializer()
self.sess.run(init)
# Saver for storing checkpoints of the model.
self.saver = tf.train.Saver(var_list=tf.global_variables(),
save_relative_paths=True)
def run(self):
coord = tf.train.Coordinator()
tf.reset_default_graph()
with tf.name_scope("create_inputs"):
reader = ImageReader(self.data_dir, self.data_eval_list,
self.input_size, None, None,
self.ignore_label, self.img_mean, coord)
image, label = reader.image, reader.label
image_batch, label_batch = tf.expand_dims(image,
dim=0), tf.expand_dims(label,
dim=0)
# Create network.
net = PSPNet({'data': image_batch},
is_training=False,
num_classes=self.num_classes)
with tf.variable_scope('', reuse=True):
flipped_img = tf.image.flip_left_right(image)
flipped_img = tf.expand_dims(flipped_img, dim=0)
net2 = PSPNet({'data': flipped_img},
is_training=False,
num_classes=self.num_classes)
# Predictions.
raw_output = net.layers['conv6']
if self.is_flip:
flipped_output = tf.image.flip_left_right(
tf.squeeze(net2.layers['conv6']))
flipped_output = tf.expand_dims(flipped_output, dim=0)
raw_output = tf.add_n([raw_output, flipped_output])
raw_output_up = tf.image.resize_bilinear(raw_output,
size=self.input_size,
align_corners=True)
raw_output_up = tf.argmax(raw_output_up, axis=3)
predictions_op = tf.expand_dims(raw_output_up, dim=3)
# mIoU
predictions_flatten = tf.reshape(predictions_op, [
-1,
])
raw_gt = tf.reshape(label_batch, [
-1,
])
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, self.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
predictions = tf.gather(predictions_flatten, indices)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(
predictions, gt, num_classes=self.num_classes)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
sess.run(tf.local_variables_initializer())
ckpt = tf.train.get_checkpoint_state(self.log_dir)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=tf.global_variables())
loader.restore(sess, ckpt.model_checkpoint_path)
Tools.print_info("Restored model parameters from {}".format(
ckpt.model_checkpoint_path))
else:
Tools.print_info('No checkpoint file found.')
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for step in range(self.num_steps):
predictions_result, _ = sess.run([predictions, update_op])
if step > 0 and self.is_measure_time:
self.calculate_time(sess, net)
if step % 1 == 0:
Tools.print_info('Finish {0}/{1} mIoU: {2}'.format(
step, self.num_steps, sess.run(mIoU)))
Tools.print_info('step {0} mIoU: {1}'.format(self.num_steps,
sess.run(mIoU)))
coord.request_stop()
coord.join(threads)
pass
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
reader = ImageReader(
' ',
args.data_list,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
net = ICNet_BN({'data': image_batch}, is_training=True, num_classes=args.num_classes)
sub4_out = net.layers['sub4_out']
sub24_out = net.layers['sub24_out']
sub124_out = net.layers['conv6_cls']
restore_var = tf.global_variables()
all_trainable = [v for v in tf.trainable_variables() if ('beta' not in v.name and 'gamma' not in v.name) or args.train_beta_gamma]
loss_sub4 = create_loss(sub4_out, label_batch, args.num_classes, args.ignore_label)
loss_sub24 = create_loss(sub24_out, label_batch, args.num_classes, args.ignore_label)
loss_sub124 = create_loss(sub124_out, label_batch, args.num_classes, args.ignore_label)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
reduced_loss = LAMBDA1 * loss_sub4 + LAMBDA2 * loss_sub24 + LAMBDA3 * loss_sub124 + tf.add_n(l2_losses)
# Using Poly learning rate policy
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - step_ph / args.num_steps), args.power))
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
if args.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
grads = tf.gradients(reduced_loss, all_trainable)
train_op = opt_conv.apply_gradients(zip(grads, all_trainable))
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=5)
ckpt = tf.train.get_checkpoint_state(args.snapshot_dir)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('Restore from pre-trained model...')
net.load(args.restore_from, sess)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, loss1, loss2, loss3, _ = sess.run([reduced_loss, loss_sub4, loss_sub24, loss_sub124, train_op], feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
else:
loss_value, loss1, loss2, loss3, _ = sess.run([reduced_loss, loss_sub4, loss_sub24, loss_sub124, train_op], feed_dict=feed_dict)
duration = time.time() - start_time
print('step {:d} \t total loss = {:.3f}, sub4 = {:.3f}, sub24 = {:.3f}, sub124 = {:.3f} ({:.3f} sec/step)'.format(step, loss_value, loss1, loss2, loss3, duration))
coord.request_stop()
coord.join(threads)
def evaluate_checkpoint(model_path, args):
coord = tf.train.Coordinator()
tf.reset_default_graph()
reader = ImageReader(
args.data_list,
INPUT_SIZE,
random_scale = False,
random_mirror = False,
ignore_label = IGNORE_LABEL,
img_mean = IMG_MEAN,
coord = coord,
train = False)
image_batch, label_batch = reader.dequeue(args.batch_size)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# Start queue threads.
threads = tf.train.start_queue_runners(coord = coord, sess = sess)
# Create network.
#net = ICNet_BN({'data': image_batch}, is_training = False, num_classes = num_classes)
net = unext(image_batch, is_train = False, n_out = NUM_CLASSES)
# Predictions.
#raw_output = net.layers['conv6']
raw_output = net.outputs
raw_output_up = tf.image.resize_bilinear(raw_output, size = INPUT_SIZE, align_corners = True)
raw_output_up = tf.argmax(raw_output_up, dimension = 3)
pred = tf.expand_dims(raw_output_up, dim = 3)
# mIoU
pred_flatten = tf.reshape(pred, [-1,])
raw_gt = tf.reshape(label_batch, [-1,])
indices = tf.squeeze(tf.where(tf.not_equal(raw_gt, IGNORE_LABEL)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
pred = tf.gather(pred_flatten, indices)
iou_metric, iou_op = tf.metrics.mean_iou(pred, gt, num_classes = num_classes)
acc_metric, acc_op = tf.metrics.accuracy(pred, gt)
# Summaries
iou_summ_op = tf.summary.scalar('mIOU', iou_metric)
acc_summ_op = tf.summary.scalar('Accuracy', acc_metric)
start = time.time()
logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S',
time.gmtime()))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver = tf.train.Saver(var_list = tf.global_variables())
load(saver, sess, model_path)
for step in range(int(num_steps / batch_size)):
preds, _, _ = sess.run([raw_output_up, iou_op, acc_op])
if step % int(100 / batch_size) == 0:
print('Finish {0}/{1}'.format(step + 1, int(num_steps / batch_size)))
iou, iou_summ, acc, acc_summ = sess.run([iou_metric, iou_summ_op, acc_metric, acc_summ_op])
sess.close()
coord.request_stop()
#coord.join(threads)
return iou, iou_summ, acc, acc_summ
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
reader = ImageReader(
DATA_DIR,
DATA_LIST_PATH,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
net = ICNet_BN({'data': image_batch}, is_training=True, num_classes=args.num_classes, filter_scale=args.filter_scale)
sub4_out = net.layers['sub4_out']
sub24_out = net.layers['sub24_out']
sub124_out = net.layers['conv6_cls']
restore_var = tf.global_variables()
all_trainable = [v for v in tf.trainable_variables() if ('beta' not in v.name and 'gamma' not in v.name) or args.train_beta_gamma]
loss_sub4 = create_loss(sub4_out, label_batch, args.num_classes, args.ignore_label)
loss_sub24 = create_loss(sub24_out, label_batch, args.num_classes, args.ignore_label)
loss_sub124 = create_loss(sub124_out, label_batch, args.num_classes, args.ignore_label)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
reduced_loss = LAMBDA1 * loss_sub4 + LAMBDA2 * loss_sub24 + LAMBDA3 * loss_sub124 + tf.add_n(l2_losses)
# Using Poly learning rate policy
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - step_ph / args.num_steps), args.power))
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
if args.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
grads = tf.gradients(reduced_loss, all_trainable)
train_op = opt_conv.apply_gradients(zip(grads, all_trainable))
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=5)
ckpt = tf.train.get_checkpoint_state(args.snapshot_dir)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('Restore from pre-trained model...')
net.load(args.restore_from, sess)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, loss1, loss2, loss3, _ = sess.run([reduced_loss, loss_sub4, loss_sub24, loss_sub124, train_op], feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
else:
loss_value, loss1, loss2, loss3, _ = sess.run([reduced_loss, loss_sub4, loss_sub24, loss_sub124, train_op], feed_dict=feed_dict)
duration = time.time() - start_time
print('step {:d} \t total loss = {:.3f}, sub4 = {:.3f}, sub24 = {:.3f}, sub124 = {:.3f} ({:.3f} sec/step)'.format(step, loss_value, loss1, loss2, loss3, duration))
coord.request_stop()
coord.join(threads)
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
reader = ImageReader(args.data_list, input_size, args.random_scale,
args.random_mirror, args.ignore_label, IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
net = PSPNet50({'data': image_batch},
is_training=True,
num_classes=args.num_classes)
raw_output = net.layers['conv6']
# According from the prototxt in Caffe implement, learning rate must multiply by 10.0 in pyramid module
fc_list = [
'conv5_3_pool1_conv', 'conv5_3_pool2_conv', 'conv5_3_pool3_conv',
'conv5_3_pool6_conv', 'conv6', 'conv5_4'
]
restore_var = [
v for v in tf.global_variables()
if not (len([f for f in fc_list
if f in v.name])) or not args.not_restore_last
]
all_trainable = [
v for v in tf.trainable_variables()
if 'gamma' not in v.name and 'beta' not in v.name
]
fc_trainable = [
v for v in all_trainable if v.name.split('/')[0] in fc_list
]
conv_trainable = [
v for v in all_trainable if v.name.split('/')[0] not in fc_list
] # lr * 1.0
fc_w_trainable = [v for v in fc_trainable
if 'weights' in v.name] # lr * 10.0
fc_b_trainable = [v for v in fc_trainable
if 'biases' in v.name] # lr * 20.0
assert (len(all_trainable) == len(fc_trainable) + len(conv_trainable))
assert (len(fc_trainable) == len(fc_w_trainable) + len(fc_b_trainable))
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, args.num_classes])
label_proc = prepare_label(label_batch,
tf.stack(raw_output.get_shape()[1:3]),
num_classes=args.num_classes,
one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_proc, [
-1,
])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, args.num_classes - 1)),
1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction,
labels=gt)
# Make mistakes for class N more important for network
if USE_CLASS_WEIGHTS:
if len(CLASS_WEIGHTS) != NUM_CLASSES:
print('Incorrect class weights, it will be not used')
else:
mask = tf.zeros_like(loss)
for i, w in enumerate(CLASS_WEIGHTS):
mask = mask + tf.cast(tf.equal(gt, i),
tf.float32) * tf.constant(w)
loss = loss * mask
l2_losses = [
args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'weights' in v.name
]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
# Processed predictions: for visualisation.
raw_output_up = tf.image.resize_bilinear(raw_output,
tf.shape(image_batch)[1:3, ])
raw_output_up = tf.argmax(raw_output_up, dimension=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# Image summary.
images_summary = tf.py_func(inv_preprocess,
[image_batch, args.save_num_images, IMG_MEAN],
tf.uint8)
labels_summary = tf.py_func(
decode_labels, [label_batch, args.save_num_images, args.num_classes],
tf.uint8)
preds_summary = tf.py_func(decode_labels,
[pred, args.save_num_images, args.num_classes],
tf.uint8)
total_summary = tf.summary.image(
'images',
tf.concat(axis=2,
values=[images_summary, labels_summary, preds_summary]),
max_outputs=args.save_num_images) # Concatenate row-wise.
summary_writer = tf.summary.FileWriter(args.snapshot_dir,
graph=tf.get_default_graph())
# Using Poly learning rate policy
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(
base_lr, tf.pow((1 - step_ph / args.num_steps), args.power))
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
if args.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
opt_fc_w = tf.train.MomentumOptimizer(learning_rate * 10.0,
args.momentum)
opt_fc_b = tf.train.MomentumOptimizer(learning_rate * 20.0,
args.momentum)
grads = tf.gradients(reduced_loss,
conv_trainable + fc_w_trainable + fc_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_fc_w = grads[len(conv_trainable):(len(conv_trainable) +
len(fc_w_trainable))]
grads_fc_b = grads[(len(conv_trainable) + len(fc_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(
zip(grads_conv, conv_trainable))
train_op_fc_w = opt_fc_w.apply_gradients(
zip(grads_fc_w, fc_w_trainable))
train_op_fc_b = opt_fc_b.apply_gradients(
zip(grads_fc_b, fc_b_trainable))
train_op = tf.group(train_op_conv, train_op_fc_w, train_op_fc_b)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
# config.gpu_options.allow_growth = True
# config.allow_soft_placement = True
# config.intra_op_parallelism_threads = 1
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=10)
ckpt = tf.train.get_checkpoint_state(SNAPSHOT_DIR)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(
os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
load_step = 0
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, _, summary = sess.run(
[reduced_loss, train_op, total_summary], feed_dict=feed_dict)
summary_writer.add_summary(summary, step)
save(saver, sess, args.snapshot_dir, step)
else:
z, t, o, p, loss_value, _ = sess.run(
[raw_gt, raw_output, gt, prediction, reduced_loss, train_op],
feed_dict=feed_dict)
print(z.shape, t.shape, o.shape, p.shape)
duration = time.time() - start_time
print('step {:d} \t loss = {:.3f}, ({:.3f} sec/step)'.format(
step, loss_value, duration))
coord.request_stop()
coord.join(threads)
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
#tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
with tf.Graph().as_default(), tf.device('/cpu:0'):
# Using Poly learning rate policy
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.train.exponential_decay(base_lr,
step_ph,
20000,
0.5,
staircase=True)
tf.summary.scalar('lr', learning_rate)
opt = tf.train.MomentumOptimizer(learning_rate, 0.9)
#opt = tf.train.RMSPropOptimizer(learning_rate, 0.9, momentum=0.9, epsilon=1e-10)
#opt = tf.train.AdamOptimizer(learning_rate)
losses = []
train_op = []
total_batch_size = args.batch_size * args.gpu_nums
with tf.name_scope('PSPnet') as scope:
with tf.name_scope("create_inputs"):
reader = ImageReader(args.data_dir, args.data_list, input_size,
args.random_blur, args.random_scale,
args.random_mirror, args.random_rotate,
args.ignore_label, IMG_MEAN, coord)
image_batch, label_batch = reader.dequeue(total_batch_size)
images_splits = tf.split(axis=0,
num_or_size_splits=args.gpu_nums,
value=image_batch)
labels_splits = tf.split(axis=0,
num_or_size_splits=args.gpu_nums,
value=label_batch)
net = PSPNet({'data': images_splits},
is_training=True,
num_classes=args.num_classes)
raw_output_list = net.layers['conv6']
add_list = [
'conv5_3_pool1_conv', 'conv5_3_pool1_conv_bn',
'conv5_3_pool2_conv', 'conv5_3_pool2_conv_bn',
'conv5_3_pool3_conv', 'conv5_3_pool3_conv_bn',
'conv5_3_pool6_conv', 'conv5_3_pool6_conv_bn', 'conv5_4',
'conv5_4_bn', 'conv6'
]
num_valide_pixel = 0
for i in range(len(raw_output_list)):
with tf.device('/gpu:%d' % i):
raw_output_up = tf.image.resize_bilinear(
raw_output_list[i],
size=input_size,
align_corners=True)
tf.summary.image('images_{}'.format(i),
images_splits[i] + IMG_MEAN,
max_outputs=4)
tf.summary.image('labels_{}'.format(i),
labels_splits[i],
max_outputs=4)
tf.summary.image('predict_{}'.format(i),
tf.cast(
tf.expand_dims(
tf.argmax(raw_output_up, -1), 3),
tf.float32),
max_outputs=4)
all_trainable = [v for v in tf.trainable_variables()]
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output_up,
[-1, args.num_classes])
label_proc = prepare_label(
labels_splits[i],
tf.stack(raw_output_up.get_shape()[1:3]),
num_classes=args.num_classes,
one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_proc, [
-1,
])
#indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, args.num_classes - 1)), 1)
indices = tf.where(
tf.logical_and(tf.less(raw_gt, args.num_classes),
tf.greater_equal(raw_gt, 0)))
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
mIoU, update_op = tf.contrib.metrics.streaming_mean_iou(
tf.argmax(tf.nn.softmax(prediction), axis=-1),
gt,
num_classes=args.num_classes)
tf.summary.scalar('mean IoU_{}'.format(i), mIoU)
train_op.append(update_op)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=prediction, labels=gt)
num_valide_pixel += tf.shape(gt)[0]
losses.append(tf.reduce_sum(loss))
l2_losses = [
args.weight_decay * tf.nn.l2_loss(v)
for v in tf.trainable_variables() if 'weights' in v.name
]
reduced_loss = tf.truediv(
tf.reduce_sum(losses), tf.cast(
num_valide_pixel, tf.float32)) + tf.add_n(l2_losses)
tf.summary.scalar('average_loss', reduced_loss)
grads = tf.gradients(reduced_loss,
all_trainable,
colocate_gradients_with_ops=True)
variable_averages = tf.train.ExponentialMovingAverage(0.99, step_ph)
variables_to_average = (tf.trainable_variables() +
tf.moving_average_variables())
variables_averages_op = variable_averages.apply(variables_to_average)
train_op = tf.group(opt.apply_gradients(zip(grads, all_trainable)),
*train_op)
train_op = tf.group(train_op, variables_averages_op)
summary_op = tf.summary.merge_all()
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.allow_soft_placement = True
sess = tf.Session(config=config)
init = [
tf.global_variables_initializer(),
tf.local_variables_initializer()
]
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=2)
#restore from resnet imagenet, bised and local_step is in moving_average
restore_var = [
v for v in tf.trainable_variables() if 'conv5_3_pool' not in v.name
and 'conv5_4' not in v.name and 'conv6' not in v.name
] + [
v for v in tf.global_variables()
if ('moving_mean' in v.name or 'moving_variance' in v.name) and
('conv5_3_pool' not in v.name and 'conv5_4_bn' not in v.name
and 'biased' not in v.name and 'local_step' not in v.name)
]
ckpt = tf.train.get_checkpoint_state(args.restore_from)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
"""
#restore from snapshot
restore_var = tf.global_variables()
ckpt = tf.train.get_checkpoint_state(args.snapshot_dir)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var, allow_empty=True)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
load_step = 0
"""
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
summary_writer = tf.summary.FileWriter(args.snapshot_dir,
graph=sess.graph)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0 and step != 0:
loss_value, _ = sess.run([reduced_loss, train_op],
feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
elif step % 100 == 0:
summary_str, loss_value, _, IOU = sess.run(
[summary_op, reduced_loss, train_op, mIoU],
feed_dict=feed_dict)
duration = time.time() - start_time
summary_writer.add_summary(summary_str, step)
print(
'step {:d} \t loss = {:.3f}, mean_IoU = {:.3f}, ({:.3f} sec/step)'
.format(step, loss_value, IOU, duration))
else:
loss_value, _ = sess.run([reduced_loss, train_op],
feed_dict=feed_dict)
coord.request_stop()
coord.join(threads)
def main():
train_list = list(np.load(config.data_file)['arr_0'])
valid_list = list(np.load(config.data_file)['arr_1'])
#print( type( train_list) ) #, type( train_aug_list ) )
#print( len( train_list) ) # len( train_aug_list ) )
#train_list = train_list[0:150000]
#valid_list = valid_list[0:1000]
print('Data detail')
print( '{0}/{1} - images for training/validation'.format( \
len( train_list) , len( valid_list) ) )
# Prepare Theano variables for inputs and targets
net, train_fn, val_fn = experiment.train_setup()
# Finally, launch the training loop.
print("Starting training...")
# We iterate over epochs:
reader_threads = []
for th in range(num_readers):
t = ImageReader( dest_queue, source_queue, \
config.image_height, config.image_width, \
config.image_channel )
t.setDaemon(True)
t.start()
#reader_threads.append ( t )
#for th in reader_threads:
# th,start()
start_loss = config.start_loss
for epoch in range(config.num_epochs):
print('Running epoch: {0}'.format(epoch + 1))
# In each epoch, we do a full pass over the training data:
train_loss = 0.
train_batches = 0
start_time = time.time()
for batch in ml_utility.iterate_minibatches( train_list, \
config.batch_size, \
shuffle=True):
source_queue.put(batch)
train_batches += 1
#print( 'Loading {0} batches'.format( train_batches ) )
train_batches -= 1
source_counter = source_queue.qsize()
dest_counter = 0
while dest_counter < train_batches:
if dest_queue.qsize() > 0:
#print( 'Started batch training' )
inputs, targets = dest_queue.get()
dest_queue.task_done()
#assume color for now
loss = train_fn(inputs, targets)
train_loss += loss
dest_counter += 1
print('trained with {0} batches'.format(dest_counter))
#print( train_batches )
# And a full pass over the validation data:
val_loss = 0.
val_acc = 0.
val_batches = 0
val_preds = None
val_truths = None
print('validation starts ')
for batch in ml_utility.iterate_minibatches2( valid_list, \
config.batch_size, \
config.image_channel, \
config.image_height, \
config.image_width ):
#inputs, targets = ml_utility.imo2py( batch)
inputs, targets = batch
#assume color for now
loss, pred_scores, acc = val_fn(inputs, targets)
pred_scores = pred_scores[:, 1]
if val_preds is None:
val_preds = pred_scores
val_truths = targets
else:
val_preds = np.concatenate( ( val_preds, pred_scores), \
axis = 0 )
val_truths = np.concatenate( (val_truths, targets ), \
axis = 0 )
val_loss += loss
val_acc += np.sum(acc)
val_batches += 1
# Then we print the results for this epoch:
epoch_time = time.time() - start_time
train_loss = train_loss / train_batches
val_loss = val_loss / val_batches
val_acc = val_acc / len(valid_list)
val_preds = np.array(val_preds)
val_truths = np.array(val_truths)
val_preds = np.reshape(val_preds, (-1, 1))
val_truths = np.reshape(val_truths, (-1, 1))
val_class = np.argmax(val_preds, axis=1)
#auc_score = roc_auc_score( val_truths, val_preds )
print("Epoch {} of {} took {:.3f}s".format(epoch + 1,
config.num_epochs,
time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_loss))
print(" validation loss:\t\t{:.6f}".format(val_loss))
print(" Performance metrics ")
print(" Validation Accuracy : {0}".format(val_acc * 100.))
#print(" AUC ROC:{0}".format( auc_score ) )
ml_utility.save_epoch_info( epoch+1,\
epoch_time,\
train_loss, \
val_loss, \
val_acc, \
config.stat_file )
#save the model after every 10 iterations
if (val_loss <= start_loss):
np.savez(config.model_file, *get_all_param_values(net))
start_loss = val_loss
def test_setup(self):
# Create queue coordinator.
self.coord = tf.train.Coordinator()
# Load reader
with tf.name_scope("create_inputs"):
reader = ImageReader(
self.conf.data_dir,
self.conf.valid_data_list,
None, # the images have different sizes
False, # no data-aug
False, # no data-aug
self.conf.ignore_label,
IMG_MEAN,
self.coord)
image, label = reader.image, reader.label # [h, w, 3 or 1]
# Add one batch dimension [1, h, w, 3 or 1]
self.image_batch, self.label_batch = tf.expand_dims(
image, dim=0), tf.expand_dims(label, dim=0)
# 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, False)
else:
net = ResNet_segmentation(self.image_batch, self.conf.num_classes,
False, self.conf.encoder_name)
# predictions
raw_output = net.outputs
raw_output = tf.image.resize_bilinear(
raw_output,
tf.shape(self.image_batch)[1:3, ])
raw_output = tf.argmax(raw_output, axis=3)
pred = tf.expand_dims(raw_output, dim=3)
self.pred = tf.reshape(pred, [
-1,
])
# labels
gt = tf.reshape(self.label_batch, [
-1,
])
# Ignoring all labels greater than or equal to n_classes.
temp = tf.less_equal(gt, self.conf.num_classes - 1)
weights = tf.cast(temp, tf.int32)
# fix for tf 1.3.0
gt = tf.where(temp, gt, tf.cast(temp, tf.uint8))
# Pixel accuracy
self.accu, self.accu_update_op = tf.contrib.metrics.streaming_accuracy(
self.pred, gt, weights=weights)
# mIoU
self.mIoU, self.mIou_update_op = tf.contrib.metrics.streaming_mean_iou(
self.pred, gt, num_classes=self.conf.num_classes, weights=weights)
# confusion matrix
self.confusion_matrix = tf.contrib.metrics.confusion_matrix(
self.pred, gt, num_classes=self.conf.num_classes, weights=weights)
# Loader for loading the checkpoint
self.loader = tf.train.Saver(var_list=tf.global_variables())
def main():
temp_flags = FLAGS.__flags.items()
temp_flags.sort()
for params, value in FLAGS.__flags.items():
print('{}: {}'.format(params, value))
input_size = (FLAGS.train_image_size, FLAGS.train_image_size)
tf.set_random_seed(1234)
coord = tf.train.Coordinator()
reader = ImageReader(FLAGS.data_dir, FLAGS.data_list, input_size,
FLAGS.random_scale, FLAGS.random_mirror,
FLAGS.ignore_label, IMG_MEAN, coord)
image_batch, label_batch = reader.dequeue(FLAGS.batch_size)
raw_output = MobileNet(image_batch,
isTraining=True,
updateBeta=FLAGS.update_beta)
psp_list = [
'conv_ds_15a', 'conv_ds_15b', 'conv_ds_15c', 'conv_ds_15d',
'conv_ds_16', 'conv_ds_17'
]
all_trainable = [v for v in tf.trainable_variables()]
if FLAGS.update_beta == False:
all_trainable = [v for v in all_trainable if 'beta' not in v.name]
psp_trainable = [
v for v in all_trainable if v.name.split('/')[1] in psp_list and (
'weights' in v.name or 'biases' in v.name)
]
conv_trainable = [v for v in all_trainable
if v not in psp_trainable] # lr * 1.0
psp_w_trainable = [v for v in psp_trainable
if 'weights' in v.name] # lr * 10.0
psp_b_trainable = [v for v in psp_trainable
if 'biases' in v.name] # lr * 20.0
assert (len(all_trainable) == len(psp_trainable) + len(conv_trainable))
assert (len(psp_trainable) == len(psp_w_trainable) + len(psp_b_trainable))
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, FLAGS.num_classes])
label_proc = prepare_label(label_batch,
tf.stack(raw_output.get_shape()[1:3]),
num_classes=FLAGS.num_classes,
one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_proc, [
-1,
])
indices = tf.squeeze(
tf.where(tf.less_equal(raw_gt, FLAGS.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction,
labels=gt)
# Regularisation loss
l2_losses = [
FLAGS.weight_decay * tf.nn.l2_loss(v)
for v in tf.trainable_variables() if 'weights' in v.name
]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
#TODO auxilary loss
#Using Poly learning rate policy
current_epoch = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.train.polynomial_decay(
FLAGS.start_learning_rate,
current_epoch,
FLAGS.decay_steps,
end_learning_rate=FLAGS.end_learning_rate,
power=FLAGS.learning_rate_decay_power,
name="poly_learning_rate")
if FLAGS.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
if FLAGS.optimizer == 'momentum':
opt_conv = tf.train.MomentumOptimizer(learning_rate,
FLAGS.momentum)
opt_psp_w = tf.train.MomentumOptimizer(learning_rate * 10.0,
FLAGS.momentum)
opt_psp_b = tf.train.MomentumOptimizer(learning_rate * 20.0,
FLAGS.momentum)
elif FLAGS.optimizer == 'rmsprop':
opt_conv = tf.train.RMSPropOptimizer(
learning_rate,
decay=FLAGS.rmsprop_decay,
momentum=FLAGS.rmsprop_momentum,
epsilon=FLAGS.opt_epsilon)
opt_psp_w = tf.train.RMSPropOptimizer(
learning_rate * 10.0,
decay=FLAGS.rmsprop_decay,
momentum=FLAGS.rmsprop_momentum,
epsilon=FLAGS.opt_epsilon)
opt_psp_b = tf.train.RMSPropOptimizer(
learning_rate * 20.0,
decay=FLAGS.rmsprop_decay,
momentum=FLAGS.rmsprop_momentum,
epsilon=FLAGS.opt_epsilon)
grads = tf.gradients(
reduced_loss, conv_trainable + psp_w_trainable + psp_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_psp_w = grads[len(conv_trainable):(len(conv_trainable) +
len(psp_w_trainable))]
grads_psp_b = grads[(len(conv_trainable) + len(psp_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(
zip(grads_conv, conv_trainable))
train_op_psp_w = opt_psp_w.apply_gradients(
zip(grads_psp_w, psp_w_trainable))
train_op_psp_b = opt_psp_b.apply_gradients(
zip(grads_psp_b, psp_b_trainable))
train_op = tf.group(train_op_conv, train_op_psp_w, train_op_psp_b)
restore_var = tf.global_variables()
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=500)
load(sess, FLAGS.pretrained_checkpoint, restore_var)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for epoch in range(FLAGS.start_epoch,
FLAGS.start_epoch + FLAGS.num_epochs):
total_loss = 0.0
for step in range(1, FLAGS.num_steps + 1):
start_time = time.time()
feed_dict = {current_epoch: epoch}
loss_value, _ = sess.run([reduced_loss, train_op],
feed_dict=feed_dict)
duration = time.time() - start_time
print('step {:d} \t loss = {:.3f}, ({:.3f} sec/step)'.format(
step, loss_value, duration))
#TODO ignore NaN loss
total_loss += loss_value
save(saver, sess, FLAGS.log_dir, epoch)
total_loss /= FLAGS.num_steps
print('Epoch {:d} completed! Total Loss = {:.3f}'.format(
epoch, total_loss))
coord.request_stop()
coord.join(threads)
def main():
"""Create the model and start the training."""
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
tf.set_random_seed(args.random_seed)
coord = tf.train.Coordinator()
with tf.name_scope("create_inputs"):
reader = ImageReader(
args.data_dir,
args.data_list,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
IMG_MEAN,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
net = PSPNet({'data': image_batch}, is_training=True, num_classes=args.num_classes)
raw_output = net.layers['conv6']
# According from the prototxt in Caffe implement, learning rate must multiply by 10.0 in pyramid module
fc_list = ['conv5_3_pool1_conv', 'conv5_3_pool2_conv', 'conv5_3_pool3_conv', 'conv5_3_pool6_conv', 'conv6', 'conv5_4']
restore_var = [v for v in tf.global_variables()]
all_trainable = [v for v in tf.trainable_variables() if 'gamma' not in v.name and 'beta' not in v.name]
fc_trainable = [v for v in all_trainable if v.name.split('/')[0] in fc_list]
conv_trainable = [v for v in all_trainable if v.name.split('/')[0] not in fc_list] # lr * 1.0
fc_w_trainable = [v for v in fc_trainable if 'weights' in v.name] # lr * 10.0
fc_b_trainable = [v for v in fc_trainable if 'biases' in v.name] # lr * 20.0
assert(len(all_trainable) == len(fc_trainable) + len(conv_trainable))
assert(len(fc_trainable) == len(fc_w_trainable) + len(fc_b_trainable))
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, args.num_classes])
label_proc = prepare_label(label_batch, tf.stack(raw_output.get_shape()[1:3]), num_classes=args.num_classes, one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_proc, [-1,])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, args.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction, labels=gt)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
# Using Poly learning rate policy
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - step_ph / args.num_steps), args.power))
# Gets moving_mean and moving_variance update operations from tf.GraphKeys.UPDATE_OPS
if args.update_mean_var == False:
update_ops = None
else:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
opt_fc_w = tf.train.MomentumOptimizer(learning_rate * 10.0, args.momentum)
opt_fc_b = tf.train.MomentumOptimizer(learning_rate * 20.0, args.momentum)
grads = tf.gradients(reduced_loss, conv_trainable + fc_w_trainable + fc_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_fc_w = grads[len(conv_trainable) : (len(conv_trainable) + len(fc_w_trainable))]
grads_fc_b = grads[(len(conv_trainable) + len(fc_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(zip(grads_conv, conv_trainable))
train_op_fc_w = opt_fc_w.apply_gradients(zip(grads_fc_w, fc_w_trainable))
train_op_fc_b = opt_fc_b.apply_gradients(zip(grads_fc_b, fc_b_trainable))
train_op = tf.group(train_op_conv, train_op_fc_w, train_op_fc_b)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=10)
ckpt = tf.train.get_checkpoint_state(SNAPSHOT_DIR)
if ckpt and ckpt.model_checkpoint_path:
loader = tf.train.Saver(var_list=restore_var)
load_step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
load(loader, sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found.')
load_step = 0
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, _ = sess.run([reduced_loss, train_op], feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
else:
loss_value, _ = sess.run([reduced_loss, train_op], feed_dict=feed_dict)
duration = time.time() - start_time
print('step {:d} \t loss = {:.3f}, ({:.3f} sec/step)'.format(step, loss_value, duration))
coord.request_stop()
coord.join(threads)
class Evaluator():
def __init__(self):
self.stop = False
if not tf.gfile.Exists(FLAGS.test_load_queue_path):
self.settings = {
"best_acc": None,
"best_checkpoint": None,
"last_checkpoint": None,
"acc_increasing": None,
"last_accs": deque()
}
else:
self.settings = np.load(FLAGS.test_load_queue_path)[()]
self.setup()
def setup(self):
self.recreate_directory_structure()
self.coord = tf.train.Coordinator()
# Load reader.
with tf.name_scope("create_inputs"):
self.reader = ImageReader("./val.npy", True, self.coord)
self.image_batch, self.label_list_batch = self.reader.dequeue(
FLAGS.batch_size)
global_step = tf.Variable(0,
dtype=tf.int32,
name='global_step',
trainable=False)
self.net = CatznDogs({'data': self.image_batch}, global_step)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
# Start queue threads.
self.threads = tf.train.start_queue_runners(coord=self.coord,
sess=self.sess)
self.ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
def run(self):
while not self.stop:
all_models_paths = self.ckpt.all_model_checkpoint_paths
index_current_model = list(all_models_paths).index(
self.ckpt.model_checkpoint_path)
if self.settings["last_checkpoint"]:
index_last_evaluated_model = list(all_models_paths).index(
self.settings["last_checkpoint"])
else:
index_last_evaluated_model = -1
if index_current_model != index_last_evaluated_model:
index_model_under_evaluation = index_last_evaluated_model + 1
self.settings["last_checkpoint"] = all_models_paths[
index_model_under_evaluation]
print("Evaluator started evaluating")
acc_pred = self.net.test(
self.image_batch,
self.label_list_batch,
self.coord,
self.sess,
self.reader.nb_samples,
checkpoint_iteration=index_model_under_evaluation)
self.settings["last_accs"].append(acc_pred)
if not self.settings[
"best_acc"] or acc_pred < self.settings["best_acc"]:
self.settings["best_acc"] = acc_pred
self.settings["best_checkpoint"] = self.settings[
"last_checkpoint"]
self.settings["acc_increasing"] = 0
else:
self.settings["acc_increasing"] += 1
if self.settings["acc_increasing"] >= 5:
self.stop = 1
np.save(FLAGS.test_load_queue_path, self.settings)
print("Best model is {} with best Acc {}".format(
self.settings["best_checkpoint"],
self.settings["best_acc"]))
else:
time.sleep(10)
self.coord.request_stop()
self.coord.join(self.threads)
print("Best model is {} with best Acc {}".format(
self.settings["best_checkpoint"], self.settings["best_acc"]))
def recreate_directory_structure(self):
if not tf.gfile.Exists(FLAGS.test_summaries_dir):
tf.gfile.MakeDirs(FLAGS.test_summaries_dir)
else:
tf.gfile.DeleteRecursively(FLAGS.test_summaries_dir)
tf.gfile.MakeDirs(FLAGS.test_summaries_dir)
