def denseASPP(inputs,
is_training,
output_stride,
pre_trained_model,
classes,
keep_prob=1.0):
with tf.contrib.slim.arg_scope(
resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
logits, end_points = resnet_v2.resnet_v2_101(
inputs,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=output_stride)
if is_training:
exclude = ['resnet_v2_101' + '/logits', 'global_step']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(
exclude=exclude)
tf.train.init_from_checkpoint(
pre_trained_model,
{v.name.split(':')[0]: v
for v in variables_to_restore})
net = end_points['resnet_v2_101' + '/block4']
with tf.name_scope("denseASPP"):
input = denseASPP_block(net, is_training, keep_prob)
with tf.contrib.slim.arg_scope(
resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
with tf.name_scope("segmentation"):
input_shape = input.get_shape().as_list()
input = tf.nn.dropout(input, keep_prob=keep_prob)
weight_1 = weight_variable([1, 1, input_shape[-1], classes])
bias = bias_variable([classes])
input = tf.nn.conv2d(input, weight_1, [1, 1, 1, 1],
padding='SAME') + bias
with tf.name_scope("upsamling"):
input_shape = input.get_shape().as_list()
input = tf.image.resize_bilinear(input, tf.shape(inputs)[1:3])
output = input
return output
def model(inputs, is_training):
"""Constructs the ResNet model given the inputs."""
if data_format == 'channels_first':
# Convert the inputs from channels_last (NHWC) to channels_first (NCHW).
# This provides a large performance boost on GPU. See
# https://www.tensorflow.org/performance/performance_guide#data_formats
inputs = tf.transpose(inputs, [0, 3, 1, 2])
# tf.logging.info('net shape: {}'.format(inputs.shape))
# encoder
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
logits, end_points = base_model(inputs,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=output_stride)
if is_training:
exclude = [base_architecture + '/logits', 'global_step']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(exclude=exclude)
tf.train.init_from_checkpoint(pre_trained_model,
{v.name.split(':')[0]: v for v in variables_to_restore})
inputs_size = tf.shape(inputs)[1:3]
net = end_points[base_architecture + '/block4']
encoder_output = atrous_spatial_pyramid_pooling(net, output_stride, batch_norm_decay, is_training)
# encoder_output = lstm(encoder_output)
with tf.variable_scope("decoder"):
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
with arg_scope([layers.batch_norm], is_training=is_training):
with tf.variable_scope("low_level_features"):
low_level_features = end_points[base_architecture + '/block1/unit_3/bottleneck_v2/conv1']
low_level_features = layers_lib.conv2d(low_level_features, 48,
[1, 1], stride=1, scope='conv_1x1')
low_level_features_size = tf.shape(low_level_features)[1:3]
with tf.variable_scope("upsampling_logits"):
net = tf.image.resize_bilinear(encoder_output, low_level_features_size, name='upsample_1')
net = tf.concat([net, low_level_features], axis=3, name='concat')
net = layers_lib.conv2d(net, 256, [3, 3], stride=1, scope='conv_3x3_1')
net = layers_lib.conv2d(net, 256, [3, 3], stride=1, scope='conv_3x3_2')
net = layers_lib.conv2d(net, num_classes, [1, 1], activation_fn=None, normalizer_fn=None,
scope='conv_1x1')
logits = tf.image.resize_bilinear(net, inputs_size, name='upsample_2')
return logits
def main(_):
with tf.Graph().as_default():
url = 'https://upload.wikimedia.org/wikipedia/commons/5/5c/Tigershark3.jpg'
image_string = urllib.urlopen(url).read()
image = tf.image.decode_jpeg(image_string, channels=3)
processed_image = inception_preprocessing.preprocess_image(image, image_size, image_size, is_training=False)
processed_images = tf.expand_dims(processed_image, 0)
# Create the model, use the default arg scope to configure the batch norm parameters.
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
logits, _ = resnet_v2.resnet_v2_50(processed_images, num_classes=1001, is_training=False)
probabilities = tf.nn.softmax(logits)
init_fn = slim.assign_from_checkpoint_fn(
os.path.join(checkpoints_dir, 'resnet_v2_50.ckpt'),
slim.get_model_variables('resnet_v2_50'))
with tf.Session() as sess:
init_fn(sess)
np_image, probabilities = sess.run([image, probabilities])
probabilities = np.reshape(probabilities, [1001])
print(probabilities.shape)
sorted_inds = [i[0] for i in sorted(enumerate(-probabilities), key=lambda x: x[1])]
print(sorted_inds)
plt.figure()
plt.imshow(np_image.astype(np.uint8))
plt.axis('off')
plt.show()
names = imagenet.create_readable_names_for_imagenet_labels()
for i in range(5):
index = sorted_inds[i]
# Shift the index of a class name by one.
print('Probability %0.6f%% => [%s]' % (probabilities[index] * 100, names[index + 1]))
def ASPP(inputs, output_stride, batch_norm_decay, is_training, depth=256):
if output_stride not in [8, 16]:
raise ValueError('output_stride must be either 8 or 16.')
#
atrous_rates = [6, 12, 18]
if output_stride == 8:
atrous_rates = [2*rate for rate in atrous_rates]
#
#why do we need arg_scope of resnet_v2
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
with arg_scope([layers.batch_norm], is_training=is_training):
inputs_size = tf.shape(inputs)[1:3]
conv_11 = layers_lib.conv2d(inputs, depth, [1, 1], stride = 1, scope = "conv_1x1")
conv_33_1 = layers_lib.conv2d(inputs, depth, [3, 3], stride = 1, rate = atrous_rates[0], scope = 'conv_3x3_1')
conv_33_2 = layers_lib.conv2d(inputs, depth, [3, 3], stride = 1, rate = atrous_rates[1], scope = 'conv_3x3_2')
conv_33_3 = layers_lib.conv2d(inputs, depth, [3, 3], stride = 1, rate = atrous_rates[2], scope = 'conv_3x3_3')
#
with tf.variable_scope("image_level_features"):
image_level_features = tf.reduce_mean(inputs, [1,2], name = 'global_average_pooling', keepdims = True)
image_level_features = layers_lib.conv2d(image_level_features, depth, [1,1], stride = 1, scope = 'conv_1x1')
image_level_features = tf.image.resize_bilinear(image_level_features, inputs_size, name='upsample')
#
net = tf.concat([conv_11, conv_33_1, conv_33_2, conv_33_3, image_level_features], axis = 3, name = 'concat')
net = layers_lib.conv2d(net, depth, [1, 1], stride = 1, scope = 'conv_1x1_concat')
#
return net
def resnet_model(images, is_training, reuse=tf.AUTO_REUSE):
with tf.contrib.framework.arg_scope(resnet_v2.resnet_arg_scope()):
resnet_fn = RESNET_MODELS[model_name]
logits, _ = resnet_fn(images, num_classes, is_training=is_training,
reuse=reuse)
logits = tf.reshape(logits, [-1, num_classes])
return logits
def model(images,
filter_type,
filter_trainable,
weight_decay,
batch_size,
is_training,
num_classes=2):
with slim.arg_scope(resnet_v2.resnet_arg_scope(weight_decay=weight_decay)):
inputs = get_residuals(images, filter_type, filter_trainable)
_, end_points = resnet_small(inputs,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=None,
include_root_block=False)
net = end_points['resnet_small/block4']
net = tf.nn.conv2d_transpose(net, tf.Variable(bilinear_upsample_weights(4,64,1024),dtype=tf.float32,name='bilinear_kernel0'), \
[batch_size, tf.shape(end_points['resnet_small/block2'])[1], tf.shape(end_points['resnet_small/block2'])[2], 64], strides=[1, 4, 4, 1], padding="SAME")
end_points['upsample1'] = net
net = tf.nn.conv2d_transpose(net, tf.Variable(bilinear_upsample_weights(4,4,64),dtype=tf.float32,name='bilinear_kernel1'), \
[batch_size, tf.shape(inputs)[1], tf.shape(inputs)[2], 4], strides=[1, 4, 4, 1], padding="SAME")
end_points['upsample2'] = net
net = layers.batch_norm(net,
activation_fn=tf.nn.relu,
is_training=is_training,
scope='post_norm')
logits = slim.conv2d(net,
num_classes, [5, 5],
activation_fn=None,
normalizer_fn=None,
scope='logits')
preds = tf.cast(tf.argmax(logits, 3), tf.int32)
preds_map = tf.nn.softmax(logits)[:, :, :, 1]
return logits, preds, preds_map, net, end_points, inputs
def load(self, **kwargs):
session = kwargs["session"]
assert isinstance(session, tf.Session)
x_input = tf.placeholder(self.x_dtype, shape=(None, ) + self.x_shape)
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
resnet_v2.resnet_v2_152(x_input,
num_classes=self.n_class,
is_training=False,
reuse=tf.AUTO_REUSE)
model_path = get_model_path('resnet_v2_152')
if not os.path.exists(model_path):
os.makedirs(model_path)
urllib.request.urlretrieve(
'http://download.tensorflow.org/models/resnet_v2_152_2017_04_14.tar.gz',
os.path.join(model_path, 'resnet_v2_152_2017_04_14.tar.gz'),
show_progress)
tar = tarfile.open(
os.path.join(model_path, 'resnet_v2_152_2017_04_14.tar.gz'))
file_names = tar.getnames()
for file_name in file_names:
tar.extract(file_name, model_path)
saver = tf.train.Saver(slim.get_model_variables(scope='resnet_v2'))
saver.restore(session, os.path.join(model_path, 'resnet_v2_152.ckpt'))
def yolonet(images, is_training=True):
with tf.Graph().as_default():
tf.logging.set_verbosity(tf.logging.INFO)
output_depth = NUM_CLASS + 5 * BOX_PER_CELL
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
bottleneck, _ = resnet_v2.resnet_v2_50(images,
global_pool=False,
is_training=is_training)
with arg_scope([layers.batch_norm], is_training=is_training):
net = bottleneck
net = layers_lib.conv2d(net,
512, [1, 1],
normalizer_fn=layers.batch_norm,
scope='yolo_layer1')
net = layers_lib.conv2d(net,
512, [3, 3],
normalizer_fn=layers.batch_norm,
scope='yolo_layer2')
net = layers_lib.conv2d(net,
512, [3, 3],
normalizer_fn=layers.batch_norm,
scope='yolo_layer3')
net = layers_lib.conv2d(net,
output_depth, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='yolo_output')
return net
def after_create_session(self, session, coord=None):
if session.run(tf.train.get_or_create_global_step()) > 0:
self.init_fn(session)
arg_scope = resnet_v2.resnet_arg_scope()
with slim.arg_scope(arg_scope):
gs_init = global_step.assign(0)
session.run(gs_init)
def featureExtractor(input,
feature_norm,
model='101',
reuse=False,
scope='resnet_v2_101'):
with tf.variable_scope(scope, reuse=reuse) as scope:
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
input = scale_RGB(input)
if '50' in model:
net, end_points = resnet_v2.resnet_v2_50(input,
global_pool=True,
is_training=False,
reuse=reuse)
elif '101' in model:
_, end_points = resnet_v2.resnet_v2_101(input,
global_pool=True,
is_training=False,
reuse=reuse)
#f = end_points['stabNet/pathFinder/featureExtractor/resnet_v2_101/block3/unit_23/bottleneck_v2/conv1'] #(18 X 32) X (18 X 32)
f = end_points[
'{}/resnet_v2_101/block4/unit_2/bottleneck_v2/conv1'.
format(scope.name)] #(9 X 16) X (9 X 16)
if feature_norm:
f = featureL2Norm(f)
return f
def CNN(inputs):
with tf.variable_scope("CNN"):
# layer = slim.conv2d(inputs, 64, [8,8], [2,4], normalizer_fn=slim.batch_norm, activation_fn=None)
# layer [B H//2 W//4 64]
# tf.summary.image('zoom', tf.transpose (layer, [3, 1, 2, 0]), max_outputs=6)
# layer = utils_nn.resNet50(layer, True, [2,1])
# [N H//32 W 2048]
# tf.summary.image('2_res50', tf.transpose (layer, [3, 1, 2, 0]), max_outputs=6)
# with slim.arg_scope(inception.inception_v3_arg_scope()):
# with slim.arg_scope([slim.batch_norm, slim.dropout], is_training=True):
# layer, _ = inception.inception_v3_base(inputs, final_endpoint="Mixed_5d")
# layer = utils_nn.resNet101(inputs, True)
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
layer, _ = resnet_v2.resnet_v2_152(inputs,
None,
is_training=True,
global_pool=False,
output_stride=16)
# 直接将网络拉到256 [N 1 256 256]
with tf.variable_scope("Normalize"):
layer = slim.conv2d(layer,
1024, [2, 2], [2, 1],
normalizer_fn=slim.batch_norm,
activation_fn=None)
layer = slim.conv2d(layer,
512, [1, 1],
normalizer_fn=slim.batch_norm,
activation_fn=None)
layer = slim.conv2d(layer,
256, [1, 1],
normalizer_fn=slim.batch_norm,
activation_fn=None)
return layer
def load(self, **kwargs):
session = kwargs["session"]
assert isinstance(session, tf.Session)
x_input = tf.placeholder(tf.float32, shape=(None, ) + self.x_shape)
with tf.contrib.framework.arg_scope(resnet_v2.resnet_arg_scope()):
logits, _ = resnet_v2.resnet_v2_50(x_input,
self.n_class,
is_training=False,
reuse=tf.AUTO_REUSE)
model_path = get_model_path('alp')
url = 'http://download.tensorflow.org/models/adversarial_logit_pairing/imagenet64_alp025_2018_06_26.ckpt.tar.gz'
fname = os.path.join(model_path, url.split('/')[-1])
if not os.path.exists(fname):
if not os.path.exists(model_path):
os.makedirs(model_path)
from six.moves import urllib
urllib.request.urlretrieve(url, fname, show_progress)
import tarfile
t = tarfile.open(fname)
t.extractall(model_path)
print('Extracted model')
saver = tf.train.Saver()
saver.restore(session, fname.split('.tar.gz')[0])
def get_resnet(x_tensor, reuse, is_training, x_batch_size):
with tf.variable_scope('resnet', reuse=reuse):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
resnet, end_points = resnet_v2.resnet_v2_50(
x_tensor,
global_pool=False,
is_training=is_training,
reuse=reuse,
output_stride=32)
global_pool = tf.reduce_mean(resnet, [1, 2])
with tf.variable_scope('fc'):
global_pool = slim.fully_connected(global_pool,
2048,
scope='fc/fc_1')
global_pool = slim.fully_connected(global_pool,
1024,
scope='fc/fc_2')
global_pool = slim.fully_connected(global_pool,
512,
scope='fc/fc_3')
theta = output_layer(global_pool, (grid_h + 1) * (grid_w + 1) * 2)
with tf.name_scope('gen_theta'):
id2_loss = tf.reduce_mean(tf.abs(theta)) * id_mul
return theta, id2_loss, id2_loss
def model(inputs, is_training):
"""Constructs the ResNet model given the inputs."""
if data_format == 'channels_first':
# Convert the inputs from channels_last (NHWC) to channels_first (NCHW).
# This provides a large performance boost on GPU. See
# https://www.tensorflow.org/performance/performance_guide#data_formats
inputs = tf.transpose(inputs, [0, 3, 1, 2])
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
logits, end_points = base_model(inputs,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=output_stride)
variables_to_restore = None
if is_training:
exclude = [base_architecture + '/logits', 'global_step']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(exclude=exclude)
inputs_size = tf.shape(inputs)[1:3]
net = end_points[base_architecture + '/block4']
net = atrous_spatial_pyramid_pooling(net, output_stride, batch_norm_decay, is_training)
with tf.variable_scope("upsampling_logits"):
net = layers_lib.conv2d(net, num_classes, [1, 1], activation_fn=None, normalizer_fn=None, scope='conv_1x1')
logits = tf.image.resize_bilinear(net, inputs_size, name='upsample')
return logits, variables_to_restore
def _ConvNet2D(self, x, is_training, reuse=False):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
f, _ = resnet_v2.resnet_v2_50(x,
num_classes=None,
is_training=is_training,
global_pool=False,
reuse=reuse)
print("resnet.out.shape: %s" % f.get_shape())
with tf.variable_scope("ConvNet2D", reuse=reuse):
f = tf.reduce_mean(f, [1, 2],
name='global_avg_pooling',
keep_dims=True)
z = slim.conv2d(f,
4096, [1, 1],
padding='VALID',
normalizer_fn=None,
scope='f2zfeture')
z = slim.conv2d(z,
self.z_dim, [1, 1],
padding='VALID',
normalizer_fn=None,
scope='z_2d')
#g_feature = tf.squeeze(g_feature, [1, 2], name='global_spatial_squeeze')
return tf.expand_dims(z, 1)
def resnet_v2_50(init_weights):
tf_compat.reset_default_graph()
image_size = 224
inputs = tf_compat.random_normal([1, image_size, image_size, 3])
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
logits, _ = resnet_v2.resnet_v2_50(inputs, 1000, is_training=False)
return tf_compat.get_default_graph()
def pyramid_pooling(input, is_training):
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
with tf.name_scope("pyramid_pooling"):
input_shape = input.get_shape().as_list()
num_output_features = input_shape[-1] // LEVEL_SIZE
with tf.name_scope("pool_1"):
pool_bin_1 = tf.nn.avg_pool(input, [1, input_shape[1], input_shape[2], 1],
[1, input_shape[1], input_shape[2], 1], padding='VALID')
weight_1 = weight_variable([1, 1, input_shape[-1], num_output_features])
output_1 = tf.nn.conv2d(pool_bin_1, weight_1, [1, 1, 1, 1], padding='SAME')
output_1 = batch_norm(output_1, is_training)
output_1 = tf.nn.relu(output_1)
output_1 = tf.image.resize_bilinear(output_1, [input_shape[1], input_shape[2]])
with tf.name_scope("pool_2"):
pool_bin_2 = tf.nn.avg_pool(input, [1, input_shape[1] // 2, input_shape[2] // 2, 1],
[1, input_shape[1] // 2, input_shape[2] // 2, 1], padding='VALID')
weight_2 = weight_variable([1, 1, input_shape[-1], num_output_features])
output_2 = tf.nn.conv2d(pool_bin_2, weight_2, [1, 1, 1, 1], padding='SAME')
output_2 = batch_norm(output_2, is_training)
output_2 = tf.nn.relu(output_2)
output_2 = tf.image.resize_bilinear(output_2, [input_shape[1], input_shape[2]])
with tf.name_scope("pool_3"):
pool_bin_3 = tf.nn.avg_pool(input, [1, input_shape[1] // 3, input_shape[2] // 3, 1],
[1, input_shape[1] // 3, input_shape[2] // 3, 1], padding='VALID')
weight_3 = weight_variable([1, 1, input_shape[-1], num_output_features])
output_3 = tf.nn.conv2d(pool_bin_3, weight_3, [1, 1, 1, 1], padding='SAME')
output_3 = batch_norm(output_3, is_training)
output_3 = tf.nn.relu(output_3)
output_3 = tf.image.resize_bilinear(output_3, [input_shape[1], input_shape[2]])
with tf.name_scope("pool_6"):
pool_bin_6 = tf.nn.avg_pool(input, [1, input_shape[1] // 6, input_shape[2] // 6, 1],
[1, input_shape[1] // 6, input_shape[2] // 6, 1], padding='VALID')
weight_6 = weight_variable([1, 1, input_shape[-1], num_output_features])
output_6 = tf.nn.conv2d(pool_bin_6, weight_6, [1, 1, 1, 1], padding='SAME')
output_6 = batch_norm(output_6, is_training)
output_6 = tf.nn.relu(output_6)
output_6 = tf.image.resize_bilinear(output_6, [input_shape[1], input_shape[2]])
input = tf.concat([input, output_1], axis=-1)
input = tf.concat([input, output_2], axis=-1)
input = tf.concat([input, output_3], axis=-1)
input = tf.concat([input, output_6], axis=-1)
return input
def create_context_path(input_im):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
last_layer, end_points = resnet_v2.resnet_v2_101(input_im,
is_training=True,
scope='resnet_v2_101',
global_pool=False)
frontend_scope = 'resnet_v2_101'
init_fn = slim.assign_from_checkpoint_fn(
model_path=os.path.join('models', 'resnet_v2_101.ckpt'),
var_list=slim.get_model_variables('resnet_v2_101'),
ignore_missing_vars=True)
layer_reduced16 = end_points[frontend_scope + '/block2']
layer_reduced32 = last_layer
layer_arm16 = arm_module(layer_reduced16, n_filter_maps=512)
layer_arm32 = arm_module(layer_reduced32, n_filter_maps=2048)
layer_global_context = tf.reduce_mean(last_layer,
axis=[1, 2],
keepdims=True,
name='global_context')
## Combining Context Features
layer_context1 = tf.math.multiply(layer_arm32, layer_global_context)
layer_context1 = layers.UpSampling2D(
size=4, interpolation='bilinear')(layer_context1)
layer_context2 = layers.UpSampling2D(
size=2, interpolation='bilinear')(layer_arm16)
context_output = tf.concat([layer_context1, layer_context2], axis=-1)
return context_output, init_fn
def PSPNet(inputs, is_training, output_stride, pre_trained_model, classes):
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
logits, end_points = resnet_v2.resnet_v2_101(inputs, num_classes=None, is_training=is_training,
global_pool=False, output_stride=output_stride)
if is_training:
exclude = ['resnet_v2_101' + '/logits', 'global_step']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(exclude=exclude)
tf.train.init_from_checkpoint(pre_trained_model, {v.name.split(':')[0]: v for v in variables_to_restore})
net = end_points['resnet_v2_101' + '/block4']
encoder_output = pyramid_pooling(net, is_training)
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
with tf.name_scope("auli_logits"):
key = 'resnet_v2_101/block3'
auxi_logits = end_points[key]
auli_shape = auxi_logits.get_shape().as_list()
weight_3 = weight_variable([3, 3, auli_shape[-1], auli_shape[-1] // 4])
auxi_logits = tf.nn.conv2d(auxi_logits, weight_3, [1, 1, 1, 1], padding='SAME')
auxi_logits = batch_norm(auxi_logits, is_training)
auxi_logits = tf.nn.relu(auxi_logits)
weight_1 = weight_variable([1, 1, auli_shape[-1] // 4, classes])
bias = bias_variable([classes])
auxi_logits = tf.nn.conv2d(auxi_logits, weight_1, [1, 1, 1, 1], padding='SAME') + bias
auxi_logits = tf.image.resize_bilinear(auxi_logits, tf.shape(inputs)[1:3])
with tf.name_scope("segmentation"):
encoder_output_shape = encoder_output.get_shape().as_list()
weight_3 = weight_variable([3, 3, encoder_output_shape[-1], encoder_output_shape[-1] // 4])
net = tf.nn.conv2d(encoder_output, weight_3, [1, 1, 1, 1], padding='SAME')
net = batch_norm(net, is_training)
net = tf.nn.relu(net)
weight_1 = weight_variable([1, 1, encoder_output_shape[-1] // 4, classes])
bias = bias_variable([classes])
net = tf.nn.conv2d(net, weight_1, [1, 1, 1, 1], padding='SAME') + bias
logits = tf.image.resize_bilinear(net, tf.shape(inputs)[1:3])
return auxi_logits, logits
def main(_):
batch_shape = [FLAGS.batch_size, FLAGS.image_height, FLAGS.image_width, 3]
num_classes = 1001
itr = 30
a = FLAGS.input_dir
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
# Prepare graph
x_input = tf.placeholder(tf.float32, shape=batch_shape)
img_resize_tensor = tf.placeholder(tf.int32, [2])
x_input_resize = tf.image.resize_images(
x_input,
img_resize_tensor,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
shape_tensor = tf.placeholder(tf.int32, [3])
padded_input = padding_layer_iyswim(x_input_resize, shape_tensor)
# 330 is the last value to keep 8*8 output, 362 is the last value to keep 9*9 output, stride = 32
padded_input.set_shape(
(FLAGS.batch_size, FLAGS.image_resize, FLAGS.image_resize, 3))
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
net, end_points = resnet_v2.resnet_v2_50(x_input,
num_classes=num_classes,
is_training=False)
predicted_labels = tf.argmax(end_points['predictions'], 3)
# Run computation
saver = tf.train.Saver(slim.get_model_variables())
session_creator = tf.train.ChiefSessionCreator(
scaffold=tf.train.Scaffold(saver=saver),
checkpoint_filename_with_path=FLAGS.checkpoint_path,
master=FLAGS.master)
with tf.train.MonitoredSession(
session_creator=session_creator) as sess:
with tf.gfile.Open(FLAGS.output_file, 'w') as out_file:
for filenames, images in load_images(FLAGS.input_dir,
batch_shape):
if np.random.randint(0, 2, size=1) == 1:
images = images[:, :, ::-1, :]
resize_shape_ = np.random.randint(310, 331)
labels = sess.run(
predicted_labels,
feed_dict={
x_input:
images,
img_resize_tensor: [resize_shape_] * 2,
shape_tensor:
np.array([
random.randint(
0, FLAGS.image_resize - resize_shape_),
random.randint(
0, FLAGS.image_resize - resize_shape_),
FLAGS.image_resize
])
})
labels = labels.flatten()
for filename, label in zip(filenames, labels):
out_file.write('{0},{1}\n'.format(filename, label))
def RES(inputs, seq_len, reuse = False):
with tf.variable_scope("OCR", reuse=reuse):
print("inputs shape:",inputs.shape)
# layer = utils_nn.resNet101V2(inputs, True) # N H W/16 2048
# layer = utils_nn.resNet50(inputs, True, [2,1]) # (N H/16 W 2048)
# with slim.arg_scope(inception.inception_v3_arg_scope()):
# with slim.arg_scope([slim.batch_norm, slim.dropout], is_training=True):
# layer, _ = inception.inception_v3_base(inputs, final_endpoint="Mixed_5d")
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
layer, _ = resnet_v2.resnet_v2_152(inputs,
None,
is_training=True,
global_pool=False,
output_stride=16)
print("ResNet shape:",layer.shape)
# 直接将网络拉到256 [N 1 512 256]
with tf.variable_scope("Normalize"):
layer = slim.conv2d(layer, 1024, [2,2], [2,1], normalizer_fn=slim.batch_norm, activation_fn=None)
layer = slim.conv2d(layer, 512, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
layer = slim.conv2d(layer, 256, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
# layer = utils_nn.resNet101(inputs, True)
# with tf.variable_scope("ResNext"):
# layer = slim.conv2d(inputs, 64, [2,4], [2,4], normalizer_fn=slim.batch_norm, activation_fn=None)
# tf.summary.image('1_2_4_zoom', tf.transpose (layer, [3, 1, 2, 0]), max_outputs=6)
# layer = utils_nn.resNext50(layer, True, [2,1]) # (N H/16 W 2048)
# tf.summary.image('2_res50', tf.transpose (layer, [3, 1, 2, 0]), max_outputs=6)
temp_layer = layer
# with tf.variable_scope("Normalize"):
# layer = slim.conv2d(layer, 1024, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
# layer = slim.conv2d(layer, 512, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
# layer = slim.conv2d(layer, 256, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
# layer = slim.conv2d(layer, 128, [1,1], normalizer_fn=slim.batch_norm, activation_fn=None)
# 将图像高度和宽度 // [2, 4]
# layer = slim.avg_pool2d(layer, [2, 4], [2, 4])
print("ResNet shape:",layer.shape)
# 增加坐标信息，增加的个数为 embedd_size
# max_width_height, embedd_size
# max_width_height 为缩放后的 w 的最大宽度，实际上的最大图片宽度为 max_width_height * 4
with tf.variable_scope("Coordinates"):
max_width_height = MAX_IMAGE_WIDTH//8
embedd_size = 64
layer = Coordinates(layer, max_width_height, embedd_size)
print("Coordinates shape:",layer.shape)
with tf.variable_scope("LSTM"):
layer = tf.squeeze(layer, squeeze_dims=1)
print("SEQ shape:",layer.shape)
layer = LSTM(layer, 256+embedd_size, seq_len) # N, W*H, 256
print("lstm shape:",layer.shape)
return layer, temp_layer
def built_network(self, inputs, is_training, dropout_rate):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
net, end_points = resnet_v2.resnet_v2_50(inputs,
self.opt.num_classes,
is_training=is_training)
net = tf.squeeze(net, axis=[1, 2])
return net
def resnet_model(images, is_training, reuse=tf.AUTO_REUSE):
with tf.contrib.framework.arg_scope(resnet_v2.resnet_arg_scope()):
resnet_fn = resnet_v2.resnet_v2_50
logits, _ = resnet_fn(images,
num_classes,
is_training=is_training,
reuse=reuse)
logits = tf.reshape(logits, [-1, num_classes])
return logits
def deeplab_v3_plus(inputs, is_training, output_stride, pre_trained_model):
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
logits, end_points = resnet_v2.resnet_v2_101(inputs, num_classes=None, is_training=is_training, global_pool=False, output_stride=output_stride)
if is_training:
exclude = ['resnet_v2_101' + '/logits', 'global_step']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(exclude=exclude)
tf.train.init_from_checkpoint(pre_trained_model, {v.name.split(':')[0]: v for v in variables_to_restore})
net = end_points['resnet_v2_101' + '/block4']
encoder_output = aspp(net, output_stride, is_training)
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=_BATCH_NORM_DECAY)):
with tf.name_scope('low_level_features'):
low_level_features = end_points['resnet_v2_101' + '/block1/unit_3/bottleneck_v2/conv1']
in_channels = low_level_features.get_shape().as_list()[-1]
low_level_shape = tf.shape(low_level_features)
weight_1x1_low_level = weight_variable([1, 1, in_channels, 48], name='weight_1x1_low_level')
conv_1x1_low_level = tf.nn.conv2d(low_level_features, weight_1x1_low_level, [1, 1, 1, 1], padding='SAME', name='conv_1x1_low_level')
conv_1x1_low_level = tf.nn.relu(batch_norm(conv_1x1_low_level, is_training), name='relu_1x1_low_level')
low_level_features_size = low_level_shape[1:3]
with tf.name_scope("upsamling_logits"):
net = tf.image.resize_bilinear(encoder_output, low_level_features_size, name='upsample_1')
net = tf.concat([net, low_level_features], axis=-1, name='concat')
weight_3x3_upsamle_1 = weight_variable([3, 3, net.get_shape().as_list()[-1], 256], name='weight_3x3_upsamle_1')
weight_3x3_upsamle_2 = weight_variable([3, 3, 256, 256], name='weight_3x3_upsamle_2')
weight_3x3_upsamle_3 = weight_variable([1, 1, 256, CLASSES], name='weight_3x3_upsamle_3') # => weight_1x1_upsamle_3
bias = bias_variable([CLASSES], name='softmax_bias')
net = tf.nn.conv2d(net, weight_3x3_upsamle_1, [1, 1, 1, 1], padding='SAME', name='conv_3x3_upsamle_1')
net = tf.nn.relu(batch_norm(net, is_training), name='conv_3x3_relu_1')
net = tf.nn.conv2d(net, weight_3x3_upsamle_2, [1, 1, 1, 1], padding='SAME', name='conv_3x3_upsamle_2')
net = tf.nn.relu(batch_norm(net, is_training), name='conv_3x3_relu_2')
net = tf.nn.conv2d(net, weight_3x3_upsamle_3, [1, 1, 1, 1], padding='SAME', name='conv_1x1_upsamle_3') + bias
logits = tf.image.resize_bilinear(net, tf.shape(inputs)[1:3], name='upsample_2')
return logits
def build(self):
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
logits, self.endpoints = resnet_v2.resnet_v2_50(
self.inputs['images'],
num_classes=self.num_classes,
is_training=self.is_training)
self.outputs['logits'] = tf.reshape(logits, [-1, self.num_classes])
self.outputs['argmax'] = tf.argmax(self.outputs['logits'],
axis=1,
name='output/predict')
def built_network(self, inputs1, inputs2, is_training=False):
inputs = tf.concat([inputs1, inputs2], axis=0)
with tf.variable_scope("Seg"):
with tf.contrib.slim.arg_scope(
resnet_v2.resnet_arg_scope(
batch_norm_decay=_BATCH_NORM_DECAY)):
logits, end_points = resnet_v2.resnet_v2_50(
inputs, is_training=is_training)
return end_points
def atrous_spatial_pyramid_pooling(self, inputs, output_stride, batch_norm_decay, is_training, depth=256):
"""Atrous Spatial Pyramid Pooling.
Args:
inputs: A tensor of size [batch, height, width, channels].
output_stride: The ResNet unit's stride. Determines the rates for atrous convolution.
the rates are (6, 12, 18) when the stride is 16, and doubled when 8.
batch_norm_decay: The moving average decay when estimating layer activation
statistics in batch normalization.
is_training: A boolean denoting whether the input is for training.
depth: The depth of the ResNet unit output.
Returns:
The atrous spatial pyramid pooling output.
"""
with tf.variable_scope("aspp"):
if output_stride not in [8, 16]:
raise ValueError('output_stride must be either 8 or 16.')
atrous_rates = [6, 12, 18]
if output_stride == 8:
atrous_rates = [2*rate for rate in atrous_rates]
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
with arg_scope([layers.batch_norm], is_training=is_training):
inputs_size = tf.shape(inputs)[1:3]
# (a) one 1x1 convolution and three 3x3 convolutions with rates = (6, 12, 18) when output stride = 16.
# the rates are doubled when output stride = 8.
conv_1x1 = layers_lib.conv2d(
inputs, depth, [1, 1], stride=1, scope="conv_1x1")
conv_3x3_1 = layers_lib.conv2d(
inputs, depth, [3, 3], stride=1, rate=atrous_rates[0], scope='conv_3x3_1')
conv_3x3_2 = layers_lib.conv2d(
inputs, depth, [3, 3], stride=1, rate=atrous_rates[1], scope='conv_3x3_2')
conv_3x3_3 = layers_lib.conv2d(
inputs, depth, [3, 3], stride=1, rate=atrous_rates[2], scope='conv_3x3_3')
# (b) the image-level features
with tf.variable_scope("image_level_features"):
# global average pooling
image_level_features = tf.reduce_mean(
inputs, [1, 2], name='global_average_pooling', keepdims=True)
# 1x1 convolution with 256 filters( and batch normalization)
image_level_features = layers_lib.conv2d(image_level_features, depth, [
1, 1], stride=1, scope='conv_1x1')
# bilinearly upsample features
image_level_features = tf.image.resize_bilinear(
image_level_features, inputs_size, name='upsample')
net = tf.concat([conv_1x1, conv_3x3_1, conv_3x3_2, conv_3x3_3,
image_level_features], axis=3, name='concat')
net = layers_lib.conv2d(
net, depth, [1, 1], stride=1, scope='conv_1x1_concat')
return net
def get_backbone(self,features):
if self.flags.model_variant.startswith('xception'):
assert False,'not implement'
elif self.flags.model_variant=='resnet_v2_50':
# inputs has shape [batch, 513, 513, 3]
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
net, end_points = resnet_v2.resnet_v2_50(features,
self.num_classes,
is_training=False,
global_pool=False,
output_stride=self.output_stride)
elif self.flags.model_variant=='resnet_v1_50':
# The key difference of the full preactivation 'v2' variant compared to the
# 'v1' variant in [1] is the use of batch normalization before every weight layer.
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
net, end_points = resnet_v1.resnet_v1_50(features,
self.num_classes,
is_training=False,
global_pool=False,
output_stride=self.output_stride)
elif self.flags.model_variant=='resnet_v2_101':
# inputs has shape [batch, 513, 513, 3]
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
net, end_points = resnet_v2.resnet_v2_101(features,
self.num_classes,
is_training=False,
global_pool=False,
output_stride=self.output_stride)
elif self.flags.model_variant=='resnet_v1_101':
# The key difference of the full preactivation 'v2' variant compared to the
# 'v1' variant in [1] is the use of batch normalization before every weight layer.
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
net, end_points = resnet_v1.resnet_v1_101(features,
self.num_classes,
is_training=False,
global_pool=False,
output_stride=self.output_stride)
else:
assert False,'not implement'
print(end_points.keys())
print(net)
def resnet_model(images, is_training, reuse=tf.AUTO_REUSE):
with tf.contrib.framework.arg_scope(resnet_v2.resnet_arg_scope()):
resnet_fn = RESNET_MODELS[model_name]
logits, _ = resnet_fn(images,
num_classes,
is_training=is_training,
reuse=reuse)
logits = tf.reshape(logits, [
-1, num_classes
]) # from [bs x 1 x 1 x num_classes] to [bs x num_classes]
return logits
def __init__(self, inputs, true_labels, is_train=False, num_classes=None):
self.true_labels = true_labels
self.NUM_CLASSES = num_classes
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
self.output, self.features = resnet_v2.resnet_v2_50(
inputs=inputs, num_classes=None, is_training=False)
self.classifier, _ = cnn.fc(input=self.forward_pass(),
num_outputs=self.NUM_CLASSES,
use_relu=False,
name='classifier')
def DeepLabNet(input_batch, is_training, num_classes, output_stride = 16, batch_norm_decay = 0.9997, backbone = 'resnet_v2_101'):
#Use channels_first to boost on GPU
#Deeplab V3+ with resnet as backbone
inputs_size = tf.shape(input_batch)[1:3]
with tf.variable_scope('deeplab'):
#ResNet as the encoder
with tf.variable_scope('encoder'):
if backbone == 'resnet_v2_50':
base_model = resnet_v2.resnet_v2_50
else:
base_model = resnet_v2.resnet_v2_101
#
#Implement tensorflow resnetV2
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay=batch_norm_decay)):
logits, end_points = base_model(input_batch,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=output_stride)
#ASPP in the middle layers
with tf.variable_scope('aspp'):
net = end_points['deeplab/encoder/' + backbone + '/block4']
encoder_output = ASPP(net, output_stride, batch_norm_decay, is_training)
#
#Decoder
with tf.variable_scope('decoder'):
with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope(batch_norm_decay = batch_norm_decay)):
with arg_scope([layers.batch_norm], is_training=is_training):
with tf.variable_scope("low_level_features"):
low_level_features = end_points['deeplab/encoder/' + backbone + '/block1/unit_3/bottleneck_v2/conv1']
low_level_features = layers_lib.conv2d(low_level_features, 48, [1,1], stride = 1, scope = 'conv_1x1')
low_level_features_size = tf.shape(low_level_features)[1:3]
with tf.variable_scope("upsampling_logits"):
net = tf.image.resize_bilinear(encoder_output, low_level_features_size, name = 'upsample_1')
net = tf.concat([net, low_level_features], axis = 3, name = 'concat')
net = layers_lib.conv2d(net, 256, [3,3], stride = 1, scope = 'conv_3x3_1')
net = layers_lib.conv2d(net, 256, [3,3], stride = 1, scope = 'conv_3x3_2')
net = layers_lib.conv2d(net, num_classes, [1, 1], activation_fn = None, normalizer_fn = None, scope = 'conv_1x1')
logits = tf.image.resize_bilinear(net, inputs_size, name = 'upsample_2')
#
return logits
def build(self, images):
"""Builds a ResNet50 embedder for the input images.
It assumes that the range of the pixel values in the images tensor is
[0,255] and should be castable to tf.uint8.
Args:
images: a tensor that contains the input images which has the shape of
NxTxHxWx3 where N is the batch size, T is the maximum length of the
sequence, H and W are the height and width of the images and C is the
number of channels.
Returns:
The embedding of the input image with the shape of NxTxL where L is the
embedding size of the output.
Raises:
ValueError: if the shape of the input does not agree with the expected
shape explained in the Args section.
"""
shape = images.get_shape().as_list()
if len(shape) != 5:
raise ValueError(
'The tensor shape should have 5 elements, {} is provided'.format(
len(shape)))
if shape[4] != 3:
raise ValueError('Three channels are expected for the input image')
images = tf.cast(images, tf.uint8)
images = tf.reshape(images,
[shape[0] * shape[1], shape[2], shape[3], shape[4]])
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
def preprocess_fn(x):
x = tf.expand_dims(x, 0)
x = tf.image.resize_bilinear(x, [299, 299],
align_corners=False)
return(tf.squeeze(x, [0]))
images = tf.map_fn(preprocess_fn, images, dtype=tf.float32)
net, _ = resnet_v2.resnet_v2_50(
images, is_training=False, global_pool=True)
output = tf.reshape(net, [shape[0], shape[1], -1])
return output
