def _build_model(self, visual_images):
"""
Builds a ResNet-50 network using slim.
"""
# visual_images = tf.placeholder(tf.float32, [None, self.height, self.width, self.channels], name='visual_images')
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
with slim.arg_scope(resnet.resnet_arg_scope(weight_decay=5e-4)):
output, network = resnet50.resnet_v1_50(
visual_images,
num_classes=self.num_classes,
is_training=is_training,
global_pool=False)
# output = tf.squeeze(output, [1, 2])
network.update({
'input': visual_images,
'is_training': is_training,
'keep_prob': keep_prob
})
self.output = output
self.network = network
self.train_vars2 = slim.get_trainable_variables(
self.scope + '/block') + slim.get_trainable_variables(self.scope +
'/conv1')
self.train_vars = slim.get_trainable_variables(
self.scope + '/logits') + slim.get_trainable_variables(self.scope +
'/conv_map')
def _build_model(self, acoustic_images):
"""
Builds the hybrid model using slim and base functions.
"""
# acoustic_images = tf.placeholder(tf.float32, [None, self.height, self.width, self.channels], name='acoustic_images')
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
end_points = OrderedDict({
'input': acoustic_images,
'is_training': is_training,
'keep_prob': keep_prob
})
mean, std, dualcam_net_output, dualcam_net_end_points = self._build_network(acoustic_images, is_training=is_training,
scope=self.scope)
end_points.update(dualcam_net_end_points)
self.mean = mean
self.std = std
self.output = dualcam_net_output # shared_net_output
self.network = end_points
self.train_vars = slim.get_trainable_variables(self.scope)
def _build_model(self, visual_images):
"""
Builds a ResNet-18 network using slim.
"""
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
with slim.arg_scope(resnet18_v1.resnet_arg_scope(weight_decay=5e-4)):
output, network = resnet18_v1.resnet_v1_18(
visual_images,
num_classes=self.num_classes,
map=self.map,
is_training=is_training)
if not self.map:
output = tf.squeeze(output, [1, 2])
network.update({
'input': visual_images,
'is_training': is_training,
'keep_prob': keep_prob
})
self.output = output
self.network = network
# we train all network because we don't have a checkpoint
self.train_vars = slim.get_trainable_variables(self.scope)
def restore_var(model_path, num_class):
x = tf.placeholder(shape=[1, None, None, 3], dtype=tf.float32)
net, end_points = nets.inception.inception_v2(x, num_classes=num_class)
output = end_points['Mixed_5c']
# output=end_points['vgg_16/conv5/conv5_3']
# train_var=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
train_var = slim.get_trainable_variables()
# for var in train_var:
# print(var)
# print(type(var))
restore_var = slim.get_variables_to_restore(exclude=['InceptionV2/Logits'])
for var in restore_var:
print(var)
print(type(var))
init = tf.global_variables_initializer()
variable_restore_op = slim.assign_from_checkpoint_fn(
model_path, restore_var, ignore_missing_vars=True)
data = np.ones(shape=[1, 256, 256, 3], dtype=np.float32)
with tf.Session() as sess:
sess.run(init)
variable_restore_op(sess)
val = sess.run(output, feed_dict={x: data})
print(val.shape)
def _build_model(self, acoustic_images):
"""
Builds the hybrid model using slim and base functions.
"""
# acoustic_images = tf.placeholder(tf.float32, [None, self.height, self.width, self.channels], name='acoustic_images')
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
end_points = OrderedDict({
'input': acoustic_images,
'is_training': is_training,
'keep_prob': keep_prob
})
dualcam_net_output, dualcam_net_end_points = self._build_network(
acoustic_images, scope=self.scope)
end_points.update(dualcam_net_end_points)
# shared_net_output, shared_net_end_points = shared.shared_net(dualcam_net_output,
# num_classes=self.num_classes,
# is_training=is_training,
# keep_prob=keep_prob,
# spatial_squeeze=True,
# scope=self.scope)
# end_points.update(shared_net_end_points)
self.output = dualcam_net_output # shared_net_output
self.network = end_points
self.train_vars = slim.get_trainable_variables(self.scope)
def train_op(self, large_image, alpha, reflectance):
with tf.variable_scope(tf.get_variable_scope(),
reuse=tf.AUTO_REUSE) as scope:
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.train.exponential_decay(
learning_rate=self.learning_rate,
global_step=global_step,
decay_rate=0.96,
decay_steps=10000)
alpha_logits, reflectance_logits = self.interface(large_image)
loss = self.calc_losses(alpha_logits, reflectance_logits, alpha,
reflectance)
optimizer = tf.train.AdamOptimizer(learning_rate)\
.minimize(loss,var_list=slim.get_trainable_variables('cloud_net'),global_step=global_step)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if update_ops:
updates = tf.group(*update_ops)
loss = control_flow_ops.with_dependencies([updates], loss)
tf.summary.scalar('learning_rate', learning_rate)
return optimizer, loss, global_step
def list_trainable_variables(self, listed_parts=None):
model_parts = {
"nothing":
lambda: [],
"fc":
lambda: [self._fc_w, self._fc_b],
"cnn":
lambda: slim.get_trainable_variables("resnet_v2_50"),
"conv1":
lambda: slim.get_trainable_variables("resnet_v2_50/conv1"),
"block1":
lambda: slim.get_trainable_variables("resnet_v2_50/block1"),
"block2":
lambda: slim.get_trainable_variables("resnet_v2_50/block2"),
"block3":
lambda: slim.get_trainable_variables("resnet_v2_50/block3"),
"block4":
lambda: slim.get_trainable_variables("resnet_v2_50/block4"),
"postnorm":
lambda: slim.get_trainable_variables("resnet_v2_50/postnorm"),
"all":
lambda: tf.trainable_variables()
}
if listed_parts:
lst = []
for part in listed_parts:
lst += model_parts[part]()
else:
lst = tf.trainable_variables()
return lst
def test():
# tensor
input_image_tensor = tf.placeholder(dtype=tf.float32,
shape=[None, 40, 40, 1],
name='image-input')
predicted_tensor = ResNet50Model(input_image_tensor, is_training=True)
predicted_tensor = tf.nn.softmax(predicted_tensor)
global_step = tf.Variable(initial_value=0, trainable=False)
# hyper-parameters
test_total = 320000
batch_size = 34
epoch_num = 1
# 输入数据
reader = Reader(
'/home/give/homework/cv/dataset/affiNist/training_and_validation_batches',
'/home/give/homework/cv/dataset/affiNist/test.mat',
batch_size=batch_size)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
ckpt = tf.train.latest_checkpoint(FLAGS.restore_model_path)
print('continue training from previous checkpoint from %s' % ckpt)
start_step = int(os.path.basename(ckpt).split('-')[1])
variable_restore_op = slim.assign_from_checkpoint_fn(
ckpt, slim.get_trainable_variables(), ignore_missing_vars=True)
variable_restore_op(sess)
sess.run(tf.assign(global_step, start_step))
start = 0
predicted = []
while start < test_total:
end = start + batch_size
if end > test_total:
end = test_total
cur_batch_images = reader.test_images[start:end]
predicted_array = sess.run(tf.argmax(predicted_tensor, axis=1),
feed_dict={
input_image_tensor:
np.expand_dims(cur_batch_images,
axis=3)
})
predicted.extend(predicted_array)
print 'Batch Accuracy[%d, %d] : %.4f' % (
start, test_total,
np.mean(
np.asarray(
predicted_array == reader.test_labels[start:end],
np.float32)))
start = end
predicted = np.array(predicted)
print 'Total Accuracy: ', np.mean(
np.asarray(predicted == reader.test_labels, np.float32))
calculate_acc_error(predicted, reader.test_labels)
def get_finetune_trainable_variables(opts):
trainable_variables = []
for scope in opts.trainable_scopes:
trainable_variables.extend(slim.get_trainable_variables(scope))
tf.logging.info("Finetune variables:")
for var in trainable_variables:
tf.logging.info(var)
return trainable_variables
def inception_v2_ssd(img,cfg):
with slim.arg_scope(inception_v2.inception_v2_arg_scope()):
logits, end_point = inception_v2.inception_v2_base(img)
Mixed_3c = end_point['Mixed_3c']
Mixed_4e = end_point['Mixed_4e']
cell_11 = end_point['Mixed_5c']
vbs = slim.get_trainable_variables()
cell_11 = tf.image.resize_bilinear(cell_11,size=[32,32])
cell_11 = tf.concat([cell_11,Mixed_4e],axis=3)
cell_7 = tf.image.resize_bilinear(Mixed_4e,size=[64,64])
cell_7 = tf.concat([cell_7,Mixed_3c],axis=3)
cell_11 = slim.conv2d(cell_11,1024,kernel_size=1,activation_fn=slim.nn.relu)
cell_7 = slim.conv2d(cell_7, 512, kernel_size=3, activation_fn=slim.nn.relu)
cell_7 = slim.conv2d(cell_7, 256, kernel_size=1, activation_fn=slim.nn.relu)
cv6 = slim.conv2d(cell_11, 1024, kernel_size=3, rate=6, activation_fn=slim.nn.relu, scope='conv6')
cv7 = slim.conv2d(cv6, 1024, kernel_size=1, activation_fn=slim.nn.relu, scope='conv7')
s = utils.normalize_to_target(cell_7, target_norm_value=12.0, dim=1)
cv8 = slim.conv2d(cv7, 256, kernel_size=1, stride=1, scope='conv8_0')
cv8 = slim.conv2d(cv8, 512, kernel_size=3, stride=2, scope='conv8_1')
cv9 = slim.conv2d(cv8, 128, kernel_size=1, stride=1, scope='conv9_0')
cv9 = slim.conv2d(cv9, 256, kernel_size=3, stride=2, scope='conv9_1')
cv10 = slim.conv2d(cv9, 128, kernel_size=1, stride=1, scope='conv10_0')
cv10 = slim.conv2d(cv10, 256, kernel_size=3, stride=2, scope='conv10_1')
cv11 = slim.conv2d(cv10, 128, kernel_size=1, stride=1, scope='conv11_0')
cv11 = slim.conv2d(cv11, 256, kernel_size=3, stride=2, scope='conv11_1')
source = [s, cv7, cv8, cv9, cv10, cv11]
conf = []
loc = []
for cv, num in zip(source, cfg.Config['aspect_num']):
print(num)
loc.append(slim.conv2d(cv, num * 4, kernel_size=3, stride=1, activation_fn=None))
conf.append(
slim.conv2d(cv, num * cfg.Config['num_classes'], kernel_size=3, stride=1, activation_fn=None))
print(loc)
loc = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, 4)) for o in loc], axis=1)
conf = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, cfg.Config['num_classes'])) for o in conf],
axis=1)
return loc, conf, vbs
def train(img_dir, gt_dir, train_list, pretrained_path):
img_input = tf.placeholder(tf.float32,
shape=[BATCH_SIZE, IMG_SIZE, IMG_SIZE, 3])
gt_input = tf.placeholder(
tf.float32, shape=[BATCH_SIZE, IMG_SIZE // 4, IMG_SIZE // 4, 6])
network = east_model.EAST(training=True, max_len=IMG_SIZE)
pred_score, pred_gmt = network.build(img_input)
loss = network.loss(gt_input[:, :, :, 0:1], pred_score,
gt_input[:, :, :, 1:6], pred_gmt)
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
learning_rate = tf.train.exponential_decay(0.0001,
global_step,
decay_steps=10000,
decay_rate=0.94,
staircase=True)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(loss=loss,
global_step=global_step)
restore_op = slim.assign_from_checkpoint_fn(pretrained_path,
slim.get_trainable_variables(),
ignore_missing_vars=True)
saver = tf.train.Saver()
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
with tf.Session(config=cfg) as sess:
if RESTORE:
saver.restore(sess, tf.train.latest_checkpoint('./checkpoint'))
else:
sess.run(tf.global_variables_initializer())
restore_op(sess)
data_list = data_processor.read_lines(train_list)
for step in range(100001):
img_batch, gt_batch = data_processor.next_batch(
img_dir, gt_dir, data_list, BATCH_SIZE)
s, g, l, lr, _ = sess.run(
[pred_score, pred_gmt, loss, learning_rate, optimizer],
feed_dict={
img_input: img_batch,
gt_input: gt_batch
})
if step % 1000 == 0 and step > 0:
saver.save(sess=sess,
save_path='./checkpoint/east.ckpt',
global_step=step)
if step % 100 == 0:
print(step, lr, l)
def __init__(self, input_shape=None, num_classes=14, nr_frames=5):
self.scope = 'resnet_v1_50'
self.num_classes = num_classes
self.height = input_shape[0]
self.width = input_shape[1]
self.channels = input_shape[2]
self.nr_frames = nr_frames
self.output, self.network = self._build_model()
self.train_vars = slim.get_trainable_variables(self.scope + '/logits')
def __init__(self, input_shape=None, num_classes=14):
self.scope = 'SoundNet'
self.num_classes = num_classes
self.height = input_shape[0] # 22050 * 5
self.width = input_shape[1] # 1
self.channels = input_shape[2] # 1
self.output, self.network = self._build_model()
self.train_vars = slim.get_trainable_variables(self.scope)
def main(_):
# download and conver flower_photos dataset to tfrecord
download_dataset.maybe_download_and_extract(FLOWERS_DATA_DIR, _DATA_URL)
convert_flowers_to_tfrecord.run(FLOWERS_DATA_DIR)
with tf.Graph().as_default() as g:
tf.logging.set_verbosity(tf.logging.INFO)
dataset = flowers.get_split('train', FLOWERS_DATA_DIR)
images, _, labels = load_batch(dataset,
batch_size=BATCH_SIZE,
is_training=True)
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
logits, _ = resnet_v2.resnet_v2_50(images,
num_classes=dataset.num_classes,
is_training=True)
logits = tf.squeeze(tf.convert_to_tensor(logits, tf.float32))
one_hot_labels = slim.one_hot_encoding(labels, dataset.num_classes)
loss = tf.losses.softmax_cross_entropy(onehot_labels=one_hot_labels,
logits=logits)
slim.losses.add_loss(loss)
total_loss = slim.losses.get_total_loss()
tf.summary.scalar('losses/total_loss', total_loss)
# 梯度根性只更新最后一层
var2training = slim.get_trainable_variables(
scope="resnet_v2_50/logits")
optimizer = tf.train.AdamOptimizer()
train_op = slim.learning.create_train_op(
total_loss, optimizer, variables_to_train=var2training)
init_fn = get_init_fn(
checkpoint_exclude_scopes=CHECKPOINT_EXCLUDE_SCOPES,
checkpoint_dir=PRETRAIN_DIR)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
print("start training")
final_loss = slim.learning.train(train_op,
logdir=TRAIN_DIR,
init_fn=init_fn,
number_of_steps=NUMBER_OF_STEPS,
trace_every_n_steps=500,
log_every_n_steps=50,
session_config=config,
save_interval_secs=60)
print('Finished training. Last batch loss %f' % final_loss)
def __init__(self, input_shape=None, num_classes=14, num_frames=12):
self.scope = 'DualCamNet'
self.num_classes = num_classes
self.num_frames = num_frames
self.height = input_shape[0]
self.width = input_shape[1]
self.channels = input_shape[2]
self.output, self.network = self._build_model()
self.train_vars = slim.get_trainable_variables(self.scope)
def _build_model(self, f):
"""
Builds the hybrid model using slim and base functions.
"""
mean, std, dualcam_net_output, dualcam_net_end_points = self._build_network2(
f, is_training=self.is_training, scope=self.scope)
self.end_points.update(dualcam_net_end_points)
self.mean = mean
self.std = std
self.output = dualcam_net_output # shared_net_output
self.network = self.end_points
self.train_vars = slim.get_trainable_variables(self.scope + '/')
def print_model():
def get_nb_params_shape(shape):
nb_params = 1
for dim in shape:
nb_params = nb_params * int(dim)
return nb_params
tot_nb_params = 0
for trainable_variable in slim.get_trainable_variables():
print(trainable_variable.name, trainable_variable.shape)
vshape = trainable_variable.get_shape() # e.g [D,F] or [W,H,C]
current_nb_params = get_nb_params_shape(vshape)
tot_nb_params = tot_nb_params + current_nb_params
print('Total number of trainable params', tot_nb_params)
def __init__(self, s_size, a_size, scope, trainer):
with tf.variable_scope(scope):
print("Scope", scope)
with tf.variable_scope("regular"):
self.network = self.make_network(s_size, a_size)
with tf.variable_scope("target"):
self.target_network = self.make_network(s_size, a_size)
# Only the worker network need ops for loss functions and gradient updating.
if scope != 'global':
self.actions = tf.placeholder(
shape=[None], dtype=tf.int32) # Index of actions taken
self.actions_onehot = tf.one_hot(
self.actions, a_size,
dtype=tf.float32) # 1-hot tensor of actions taken
# losses!
self.target_q_t = tf.placeholder('float32', [None],
name='target_q_t')
self.q_acted = tf.reduce_sum(self.network.value *
self.actions_onehot,
reduction_indices=1,
name='q_acted')
self.loss = tf.reduce_mean(tf.square(self.target_q_t -
self.q_acted),
name='loss')
# Get gradients from local network using local losses
self.local_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope)
self.gradients = tf.gradients(self.loss, self.local_vars)
self.var_norms = tf.global_norm(self.local_vars)
grads, self.grad_norms = tf.clip_by_global_norm(
self.gradients, 40.0)
# Apply local gradients to global network
global_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, 'global/regular')
target_weights = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, "global/target")
# set global target network to be the same as local network
weights = slim.get_trainable_variables(scope=scope +
"/regular")
self.assign_op = {}
for w, t_w in zip(weights, target_weights):
self.assign_op[w.name] = t_w.assign(w)
self.apply_grads = trainer.apply_gradients(
zip(grads, global_vars))
def model(inputs,cfg):
source = []
with tf.variable_scope('vgg_16',default_name=None, values=[inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope([slim.conv2d, slim.max_pool2d]):
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
net = slim.max_pool2d(net, [2, 2], scope='pool1')
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
net = slim.max_pool2d(net, [2, 2], scope='pool2')
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net = slim.max_pool2d(net, [2, 2], scope='pool3')
cov_43 = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
net = slim.max_pool2d(cov_43, [2, 2], scope='pool4')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
net = slim.max_pool2d(net, kernel_size=3,stride=1, scope='pool5')
vbs = slim.get_trainable_variables()
utils.count_parm()
#vbs =None
cv6 = slim.conv2d(net,1024,kernel_size=3,rate=6,activation_fn=slim.nn.relu,scope='conv6')
cv7 = slim.conv2d(cv6,1024,kernel_size=1,activation_fn=slim.nn.relu,scope='conv7')
s = utils.normalize_to_target(cov_43, target_norm_value=2.0, dim=1)
cv8 = slim.conv2d(cv7,256,kernel_size=1,stride=1,scope='conv8_0')
cv8 = slim.conv2d(cv8, 512, kernel_size=3, stride=2, scope='conv8_1')
cv9 = slim.conv2d(cv8, 128, kernel_size=1, stride=1, scope='conv9_0')
cv9 = slim.conv2d(cv9, 256, kernel_size=3, stride=2, scope='conv9_1')
cv10 = slim.conv2d(cv9, 128, kernel_size=1, stride=1, scope='conv10_0')
cv10 = slim.conv2d(cv10, 256, kernel_size=3, stride=2, scope='conv10_1')
cv11 = slim.conv2d(cv10, 128, kernel_size=1, stride=1, scope='conv11_0')
cv11 = slim.conv2d(cv11, 256, kernel_size=3, stride=2, scope='conv11_1')
source = [s,cv7,cv8,cv9,cv10,cv11]
conf = []
loc = []
for cv,num in zip(source,cfg.Config['aspect_num']):
loc.append(slim.conv2d(cv,num*4,kernel_size=3,stride=1,activation_fn=None))
conf.append(slim.conv2d(cv, num*cfg.Config['num_classes'], kernel_size=3, stride=1, activation_fn=None))
print(loc)
loc = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, 4)) for o in loc], axis=1)
conf = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, cfg.Config['num_classes'])) for o in conf], axis=1)
return loc,conf,vbs
def resnetv2_ssd(img):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
net, end_points = resnet_v2.resnet_v2_50(img, is_training=True)
c1 = end_points['resnet_v2_50/block1']
c2 = end_points['resnet_v2_50/block2']
base_16_0 = end_points['resnet_v2_50/block3']
base_16_1 = end_points['resnet_v2_50/block4']
vbs = slim.get_trainable_variables()
# vbs = None
base_16_0 = slim.conv2d(base_16_0, 512, 1)
base_16_1 = slim.conv2d(base_16_1, 512, 1)
c3 = tf.concat([base_16_0, base_16_1], axis=3)
return c1, c2, c3, vbs
def __init__(self, num_classes=None):
self.scope = 'AVNet'
self.num_classes = num_classes
self.video_height = 224
self.video_width = 224
self.video_channels = 3
self.audio_height = 36
self.audio_width = 48
self.audio_channels = 12
self.num_frames = 12
self.output, self.network = self._build_model()
self.audio_train_vars = slim.get_trainable_variables('DualCamNet')
self.video_train_vars = slim.get_trainable_variables(
'resnet_v1_50/logits')
self.audio_video_train_vars = slim.get_trainable_variables(self.scope)
self.train_vars = self.audio_train_vars + self.video_train_vars + self.audio_video_train_vars
def inception_v2_ssd(img,cfg):
with slim.arg_scope(inception_v2.inception_v2_arg_scope()):
logits, end_point = inception_v2.inception_v2_base(img)
Mixed_3c = end_point['Mixed_3c']
Mixed_4e = end_point['Mixed_4e']
Mixed_5c = end_point['Mixed_5c']
vbs = slim.get_trainable_variables()
#vbs = None
cell_11 = tf.image.resize_bilinear(Mixed_5c,size=[int(32*(cfg.image_size/512)),int(32*(cfg.image_size/512))])
cell_11 = tf.concat([cell_11,Mixed_4e],axis=3)
cell_7 = tf.image.resize_bilinear(Mixed_4e,size=[int(64*(cfg.image_size/512)),int(64*(cfg.image_size/512))])
cell_7 = tf.concat([cell_7,Mixed_3c],axis=3)
mask_fp = get_mask_fp(Mixed_3c ,Mixed_4e,Mixed_5c)
cell_11 = slim.conv2d(cell_11,1024,kernel_size=1,activation_fn=slim.nn.relu)
cell_7 = slim.conv2d(cell_7, 512, kernel_size=3, activation_fn=slim.nn.relu)
cell_7 = slim.conv2d(cell_7, 256, kernel_size=1, activation_fn=slim.nn.relu)
cv6 = slim.conv2d(cell_11, 1024, kernel_size=3, rate=6, activation_fn=slim.nn.relu, scope='conv6')
cv7 = slim.conv2d(cv6, 1024, kernel_size=1, activation_fn=slim.nn.relu, scope='conv7')
s = utils.normalize_to_target(cell_7, target_norm_value=cfg.norm_value, dim=1)
cv8 = inception(cv7, out_put=512, name='cv8', stride=2)
cv9 = inception(cv8, out_put=256, name='cv9', stride=2)
cv10 = inception(cv9, out_put=256, name='cv10', stride=2)
cv11 = inception(cv10, out_put=256,name= 'cv11', stride=2)
source = [s, cv7, cv8, cv9, cv10, cv11]
conf = []
loc = []
for cv, num in zip(source, cfg.Config['aspect_num']):
loc.append(slim.conv2d(cv, num * 4, kernel_size=3, stride=1, activation_fn=None))
conf.append(
slim.conv2d(cv, num * cfg.Config['num_classes'], kernel_size=3, stride=1, activation_fn=None))
loc = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, 4)) for o in loc], axis=1)
conf = tf.concat([tf.reshape(o, shape=(cfg.batch_size, -1, cfg.Config['num_classes'])) for o in conf],
axis=1)
return loc, conf,mask_fp, vbs
def variables_to_restore_and_train(pre_trained_model):
if 'vgg_16' in pre_trained_model:
exclude = ['fully_connected']
train_sc = ['fully_connected']
elif 'inception_v3' in pre_trained_model:
exclude = [
'InceptionV3/Logits', 'InceptionV3/AuxLogits', 'fully_connected'
]
train_sc = ['fully_connected']
elif 'resnet_v1_50' in pre_trained_model:
exclude = ['resnet_v1_50/logits']
train_sc = ['resnet_v1_50/logits']
else:
exclude = []
train_sc = []
variables_to_restore = slim.get_variables_to_restore(exclude=exclude)
variables_to_train = []
for sc in train_sc:
variables_to_train += slim.get_trainable_variables(sc)
return variables_to_train, variables_to_restore
def _build_model(self, acoustic_images):
"""
Builds the hybrid model using slim and base functions.
"""
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
end_points = OrderedDict({
'input': acoustic_images,
'is_training': is_training,
'keep_prob': keep_prob
})
dualcam_net_output, dualcam_net_end_points = self._build_network(
acoustic_images, is_training=is_training, scope=self.scope)
end_points.update(dualcam_net_end_points)
self.output = dualcam_net_output
self.network = end_points
self.train_vars = slim.get_trainable_variables(self.scope)
def create_restore_hook(opts):
variables_to_restore = slim.get_variables_to_restore(
exclude=opts.exclude_restore_scopes)
tf.logging.info('Restore variables: ')
for var in variables_to_restore:
tf.logging.info(var)
checkpoint_path = opts.pretrained_weights_path
if tf.gfile.IsDirectory(checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(checkpoint_path)
tf.logging.info('checkpoint_path = {}'.format(checkpoint_path))
init_fn = slim.assign_from_checkpoint_fn(checkpoint_path,
variables_to_restore,
ignore_missing_vars=False)
tf.logging.info('Global trainable variables: ')
for var in slim.get_trainable_variables():
tf.logging.info(var)
return _RestoreHook(init_fn)
def _build_opt(self, c,
freeze_cnn=cfg.FREEZE_CNN,
freeze_rnn=cfg.FREEZE_RNN,
log=no_op):
log('- build-opt -')
opt = tf.train.AdamOptimizer(learning_rate=self.learning_rate_)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if freeze_cnn:
log('freezing cnn')
train_vars = [v for v in slim.get_trainable_variables() if ('vo/sconv' not in v.name) and ('vo/conv' not in v.name)]
if freeze_rnn:
train_vars = [v for v in train_vars if ('vo/reduction' not in v.name) and ('vo/rnn' not in v.name)]
log('train variables:')
for v in train_vars:
log('\t {} : {}'.format(v.name, v.shape))
log('---------')
else:
# train everything
train_vars = None
with tf.control_dependencies(update_ops):
#grads, vars = zip(*opt.compute_gradients(c))
#grads_c, _ = tf.clip_by_global_norm(grads, 1.0)
#train_op = opt.apply_gradients(zip(grads_c, vars), global_step=self.step_)
##train_op = opt.minimize(c, global_step=self.step_)
train_op = tf.contrib.layers.optimize_loss(c, self.step_,
learning_rate=self.learning_rate_,
optimizer='Adam',
clip_gradients=3.0,
summaries=['loss', 'learning_rate', 'global_gradient_norm', 'gradients'],
variables=train_vars
)
log('-------------')
return train_op
def _build_model(self, melspectrum):
"""
Builds the hybrid model using slim and base functions.
"""
# acoustic_images = tf.placeholder(tf.float32, [None, self.height, self.width, self.channels], name='acoustic_images')
# is_training = tf.placeholder(tf.bool, name='is_training')
# keep_prob = tf.placeholder(tf.float32, name='keep_prob')
# end_points = OrderedDict({
# 'input': melspectrum,
# 'is_training': is_training,
# 'keep_prob': keep_prob
# })
output = self._build_network(melspectrum, scope=self.scope)
# end_points.update(dualcam_net_end_points)
self.output = output # shared_net_output
# self.network = end_points
self.train_vars = slim.get_trainable_variables(self.scope)
def validation_dice(val_dir):
input_image_tensor = tf.placeholder(dtype=tf.float32,
shape=[
None, Config.vox_size[0],
Config.vox_size[1],
Config.vox_size[2], 1
],
name='image_tensor')
input_gt_tensor = tf.placeholder(dtype=tf.int32,
shape=[
None, Config.vox_size[0],
Config.vox_size[1], Config.vox_size[2]
],
name='gt_tensor')
pred_last, pred_6, pred_3 = model(input_image_tensor)
global_step = tf.train.get_or_create_global_step()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver = tf.train.Saver(tf.global_variables(), max_to_keep=5)
ckpt = tf.train.latest_checkpoint(FLAGS.model_restore_path)
print('continue training from previous checkpoint from %s' % ckpt)
start_step = int(os.path.basename(ckpt).split('-')[1])
variable_restore_op = slim.assign_from_checkpoint_fn(
ckpt, slim.get_trainable_variables(), ignore_missing_vars=True)
variable_restore_op(sess)
sess.run(tf.assign(global_step, start_step))
# obtain the data of validation
image_paths = glob(os.path.join(val_dir, 'volume-*.nii'))
for image_path in image_paths:
basename = os.path.basename(image_path)
file_id = basename.split('-')[1].split('.')[0]
gt_path = os.path.join(val_dir, 'segmentation-' + file_id + '.nii')
pred_result = compute_onefile()
def _build_model(self, spectrogram):
"""
Builds the model using slim.
"""
is_training = tf.placeholder(tf.bool, name='is_training')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
end_points = OrderedDict({
'input': spectrogram,
'is_training': is_training,
'keep_prob': keep_prob
})
with slim.arg_scope(self._build_arg_scope(weight_decay=5e-4)):
hearnet_output, hearnet_end_points = self._build_network(
spectrogram,
num_classes=self.num_classes,
is_training=is_training,
keep_prob=keep_prob,
scope=self.scope)
end_points.update(hearnet_end_points)
shared_net_output, shared_net_end_points = shared.shared_net(
hearnet_output,
num_classes=self.num_classes,
is_training=is_training,
keep_prob=keep_prob,
spatial_squeeze=True,
scope=self.scope,
embedding=self.embedding)
end_points.update(shared_net_end_points)
self.output = shared_net_output
self.network = end_points
self.train_vars = slim.get_trainable_variables(self.scope)
def main(argv=None):
import os
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu_list
if not tf.gfile.Exists(FLAGS.checkpoint_path):
# tf.gfile.MkDir(FLAGS.checkpoint_path)
os.makedirs(FLAGS.checkpoint_path)
else:
if not FLAGS.restore:
tf.gfile.DeleteRecursively(FLAGS.checkpoint_path)
tf.gfile.MkDir(FLAGS.checkpoint_path)
input_images = tf.placeholder(tf.float32, shape=[None, None, None, 3], name='input_images')
input_score_maps = tf.placeholder(tf.float32, shape=[None, None, None, 1], name='input_score_maps')
if FLAGS.geometry == 'RBOX':
input_geo_maps = tf.placeholder(tf.float32, shape=[None, None, None, 5], name='input_geo_maps')
else:
input_geo_maps = tf.placeholder(tf.float32, shape=[None, None, None, 8], name='input_geo_maps')
input_training_masks = tf.placeholder(tf.float32, shape=[None, None, None, 1], name='input_training_masks')
global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False)
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate, global_step, decay_steps=10000,
decay_rate=0.94, staircase=True)
# add summary
tf.summary.scalar('learning_rate', learning_rate)
opt = tf.train.AdamOptimizer(learning_rate)
# opt = tf.train.MomentumOptimizer(learning_rate, 0.9)
# split
input_images_split = tf.split(input_images, len(gpus))
input_score_maps_split = tf.split(input_score_maps, len(gpus))
input_geo_maps_split = tf.split(input_geo_maps, len(gpus))
input_training_masks_split = tf.split(input_training_masks, len(gpus))
tower_grads = []
reuse_variables = None
for i, gpu_id in enumerate(gpus):
with tf.device('/gpu:%d' % gpu_id):
with tf.name_scope('model_%d' % gpu_id) as scope:
iis = input_images_split[i]
isms = input_score_maps_split[i]
igms = input_geo_maps_split[i]
itms = input_training_masks_split[i]
total_loss, model_loss = tower_loss(iis, isms, igms, itms, reuse_variables)
batch_norm_updates_op = tf.group(*tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope))
reuse_variables = True
grads = opt.compute_gradients(total_loss)
tower_grads.append(grads)
grads = average_gradients(tower_grads)
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
summary_op = tf.summary.merge_all()
# save moving average
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
# batch norm updates
with tf.control_dependencies([variables_averages_op, apply_gradient_op, batch_norm_updates_op]):
train_op = tf.no_op(name='train_op')
saver = tf.train.Saver(tf.global_variables())
summary_writer = tf.summary.FileWriter(FLAGS.checkpoint_path, tf.get_default_graph())
init = tf.global_variables_initializer()
if FLAGS.pretrained_model_path is not None:
variable_restore_op = slim.assign_from_checkpoint_fn(FLAGS.pretrained_model_path, slim.get_trainable_variables(),
ignore_missing_vars=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
if FLAGS.restore:
print('continue training from previous checkpoint')
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
saver.restore(sess, ckpt)
else:
sess.run(init)
if FLAGS.pretrained_model_path is not None:
variable_restore_op(sess)
data_generator = read_icdar.get_batch(num_workers=FLAGS.num_readers,
input_size=FLAGS.input_size,
batch_size=FLAGS.batch_size_per_gpu * len(gpus))
start = time.time()
for step in range(FLAGS.start_steps, FLAGS.max_steps):
data = next(data_generator)
feed_dict = {input_images: data[0],
input_score_maps: data[2],
input_geo_maps: data[3],
input_training_masks: data[4]}
ml, tl, _ = sess.run([model_loss, total_loss, train_op], feed_dict=feed_dict)
if np.isnan(tl):
print('Loss diverged, stop training')
break
if step % 10 == 0:
avg_time_per_step = (time.time() - start)/10
avg_examples_per_second = (10 * FLAGS.batch_size_per_gpu * len(gpus))/(time.time() - start)
start = time.time()
print('Step {:06d}, model loss {:.4f}, total loss {:.4f}, {:.2f} seconds/step, '
'{:.2f} examples/second'.format(step, ml, tl, avg_time_per_step, avg_examples_per_second))
logging.info('Step {:06d}, model loss {:.4f}, total loss {:.4f}, {:.2f} seconds/step, '
'{:.2f} examples/second'.format(step, ml, tl, avg_time_per_step, avg_examples_per_second))
if step % FLAGS.save_checkpoint_steps == 0:
saver.save(sess, FLAGS.checkpoint_path + 'model.ckpt', global_step=global_step)
if step % FLAGS.save_summary_steps == 0:
_, tl, summary_str = sess.run([train_op, total_loss, summary_op], feed_dict=feed_dict)
summary_writer.add_summary(summary_str, global_step=step)
def main(argv=None):
import os
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu_list
if not tf.gfile.Exists(FLAGS.checkpoint_path):
tf.gfile.MkDir(FLAGS.checkpoint_path)
else:
if not FLAGS.restore:
tf.gfile.DeleteRecursively(FLAGS.checkpoint_path)
tf.gfile.MkDir(FLAGS.checkpoint_path)
input_images = tf.placeholder(tf.float32, shape=[None, None, None, 3], name='input_images')
input_score_maps = tf.placeholder(tf.float32, shape=[None, None, None, 1], name='input_score_maps')
if FLAGS.geometry == 'RBOX':
input_geo_maps = tf.placeholder(tf.float32, shape=[None, None, None, 5], name='input_geo_maps')
else:
input_geo_maps = tf.placeholder(tf.float32, shape=[None, None, None, 8], name='input_geo_maps')
input_training_masks = tf.placeholder(tf.float32, shape=[None, None, None, 1], name='input_training_masks')
global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False)
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate, global_step, decay_steps=10000, decay_rate=0.94, staircase=True)
# add summary
tf.summary.scalar('learning_rate', learning_rate)
opt = tf.train.AdamOptimizer(learning_rate)
# opt = tf.train.MomentumOptimizer(learning_rate, 0.9)
# split
input_images_split = tf.split(input_images, len(gpus))
input_score_maps_split = tf.split(input_score_maps, len(gpus))
input_geo_maps_split = tf.split(input_geo_maps, len(gpus))
input_training_masks_split = tf.split(input_training_masks, len(gpus))
tower_grads = []
reuse_variables = None
for i, gpu_id in enumerate(gpus):
with tf.device('/gpu:%d' % gpu_id):
with tf.name_scope('model_%d' % gpu_id) as scope:
iis = input_images_split[i]
isms = input_score_maps_split[i]
igms = input_geo_maps_split[i]
itms = input_training_masks_split[i]
total_loss, model_loss = tower_loss(iis, isms, igms, itms, reuse_variables)
batch_norm_updates_op = tf.group(*tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope))
reuse_variables = True
grads = opt.compute_gradients(total_loss)
tower_grads.append(grads)
grads = average_gradients(tower_grads)
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
summary_op = tf.summary.merge_all()
# save moving average
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
# batch norm updates
with tf.control_dependencies([variables_averages_op, apply_gradient_op, batch_norm_updates_op]):
train_op = tf.no_op(name='train_op')
saver = tf.train.Saver(tf.global_variables())
summary_writer = tf.summary.FileWriter(FLAGS.checkpoint_path, tf.get_default_graph())
init = tf.global_variables_initializer()
if FLAGS.pretrained_model_path is not None:
variable_restore_op = slim.assign_from_checkpoint_fn(FLAGS.pretrained_model_path, slim.get_trainable_variables(),
ignore_missing_vars=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
if FLAGS.restore:
print 'continue training from previous checkpoint'
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
saver.restore(sess, ckpt)
else:
sess.run(init)
if FLAGS.pretrained_model_path is not None:
variable_restore_op(sess)
data_generator = icdar.get_batch(num_workers=FLAGS.num_readers,
input_size=FLAGS.input_size,
batch_size=FLAGS.batch_size * len(gpus))
start = time.time()
for step in xrange(FLAGS.max_steps):
data = data_generator.next()
ml, tl, _ = sess.run([model_loss, total_loss, train_op], feed_dict={input_images: data[0],
input_score_maps: data[2],
input_geo_maps: data[3],
input_training_masks: data[4]})
if np.isnan(tl):
print 'Loss diverged, stop training'
break
if step % 10 == 0:
avg_time_per_step = (time.time() - start)/10
avg_examples_per_second = (10 * FLAGS.batch_size * len(gpus))/(time.time() - start)
start = time.time()
print 'Step {:06d}, model loss {:.4f}, total loss {:.4f}, {:.2f} seconds/step, {:.2f} examples/second'.format(
step, ml, tl, avg_time_per_step, avg_examples_per_second)
if step % FLAGS.save_checkpoint_steps == 0:
saver.save(sess, FLAGS.checkpoint_path + 'model.ckpt', global_step=global_step)
if step % FLAGS.save_summary_steps == 0:
_, tl, summary_str = sess.run([train_op, total_loss, summary_op], feed_dict={input_images: data[0],
input_score_maps: data[2],
input_geo_maps: data[3],
input_training_masks: data[4]})
summary_writer.add_summary(summary_str, global_step=step)