def _squeezenet2d(flat_input, keep_prob, n_classes):
w_ini, b_ini, r_ini = initializers()
x_multichannel = tf.reshape(
flat_input, [-1, conf.img_dim[0], conf.img_dim[1], conf.num_ch])
net = conv2d(
x_multichannel,
filters=96,
kernel_size=7,
name='conv1', # keras SqueezeNet uses 3x3 kernel
kernel_initializer=w_ini,
bias_initializer=b_ini,
kernel_regularizer=r_ini)
net = max_pooling2d(net, pool_size=3, strides=2, name='maxpool1')
net = fire_module2d(net, 16, 64, name='fire2')
net = fire_module2d(net, 16, 64, name='fire3')
net = fire_module2d(net, 32, 128, name='fire4')
net = max_pooling2d(net, pool_size=3, strides=2, name='maxpool4')
net = fire_module2d(net, 32, 128, name='fire5')
net = fire_module2d(net, 48, 192, name='fire6')
net = fire_module2d(net, 48, 192, name='fire7')
net = fire_module2d(net, 64, 256, name='fire8')
net = max_pooling2d(net, pool_size=3, strides=2, name='maxpool8')
net = fire_module2d(net, 64, 256, name='fire9')
net = tf.nn.dropout(net, keep_prob=keep_prob, name='dropout9')
net = conv2d(net,
n_classes,
1,
1,
name='conv10',
kernel_initializer=w_ini,
bias_initializer=b_ini,
kernel_regularizer=r_ini)
logits = tf.reduce_mean(net, axis=[1, 2], name='global_avgpool10')
return logits
def vgg_11(x, keep_proba, is_train, num_classes):
# https://arxiv.org/abs/1409.1556 --> configuration A
w_ini, b_ini, r_ini = initializers(conf.bias_init, conf.l2_str)
x_multichannel = tf.reshape(
x, [-1, conf.img_dim[0], conf.img_dim[1], conf.num_ch])
net = _conv2d(x_multichannel, 64, 3, 'conv1')
net = max_pooling2d(net, 2, 2, 'same', name='maxpool1')
net = _conv2d(net, 128, 3, 'conv2')
net = max_pooling2d(net, 2, 2, 'same', name='maxpool2')
net = _conv2d(net, 256, 3, 'conv3_1')
net = _conv2d(net, 256, 3, 'conv3_2')
net = max_pooling2d(net, 2, 2, 'same', name='maxpool3')
net = _conv2d(net, 512, 3, 'conv4_1')
net = _conv2d(net, 512, 3, 'conv4_2')
net = max_pooling2d(net, 2, 2, 'same', name='maxpool4')
net = _conv2d(net, 512, 3, 'conv5_1')
net = _conv2d(net, 512, 3, 'conv5_2')
net = max_pooling2d(net, 2, 2, 'same', name='maxpool5')
net = flatten(net, name='flat_layer')
net = _dense(net, 4096, 'fc6')
net = tf.nn.dropout(net, keep_proba, name='dropout6')
net = _dense(net, 4096, 'fc7')
net = tf.nn.dropout(net, keep_proba, name='dropout7')
# # TODO consider replacing fc layers by conv layers
# # as in https://github.com/tensorflow/models/blob/master/research/slim/nets/vgg.py
# # for "if include_top" see https://github.com/keras-team/keras-applications/blob/master/keras_applications/vgg16.py
# net = _conv2d(net, 4096, 7, 'fc6_replacement', padding='VALID')
# net = tf.nn.dropout(net, keep_proba, name='dropout7')
# net = _conv2d(net, 4096, 1, 'fc7_replacement')
# net = tf.reduce_mean(net, axis=[1, 2], name='global_avg_pool8')
logits = _dense(net, conf.num_cl, 'logits', activation=None)
# softmax performed at loss function
return logits
def add_inference(self, inputs, training, nclass):
df = 'channels_first' if self.data_format == 'NCHW' else 'channels_last'
conv1 = layers.conv2d(inputs=inputs,
filters=32,
kernel_size=[5, 5],
padding="same",
data_format=df,
activation=tf.nn.relu)
pool1 = layers.max_pooling2d(inputs=conv1,
pool_size=[2, 2],
strides=2,
data_format=df)
conv2 = layers.conv2d(inputs=pool1,
filters=64,
kernel_size=[3, 3],
padding="same",
data_format=df,
activation=tf.nn.relu)
pool2 = layers.max_pooling2d(inputs=conv2,
pool_size=[2, 2],
strides=2,
data_format=df)
pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])
dense = layers.dense(inputs=pool2_flat,
units=1024,
activation=tf.nn.relu)
dropout = layers.dropout(inputs=dense, rate=0.4, training=training)
logits = layers.dense(inputs=dropout, units=nclass)
return logits
def vgg_backbone(inputs, **config):
params_conv = {
'padding': 'SAME',
'data_format': config['data_format'],
'activation': tf.nn.relu,
'batch_normalization': False,
'training': config['training'],
'kernel_reg': config.get('kernel_reg', 0.)
}
params_pool = {'padding': 'SAME', 'data_format': config['data_format']}
with tf.variable_scope('vgg', reuse=tf.AUTO_REUSE):
x = vgg_block(inputs, 64, 3, 'conv1_1', **params_conv)
x = vgg_block(x, 64, 3, 'conv1_2', **params_conv)
x = tfl.max_pooling2d(x, 2, 2, name='pool1', **params_pool)
x = vgg_block(x, 64, 3, 'conv2_1', **params_conv)
x = vgg_block(x, 64, 3, 'conv2_2', **params_conv)
x = tfl.max_pooling2d(x, 2, 2, name='pool2', **params_pool)
x = vgg_block(x, 128, 3, 'conv3_1', **params_conv)
x = vgg_block(x, 128, 3, 'conv3_2', **params_conv)
x = tfl.max_pooling2d(x, 2, 2, name='pool3', **params_pool)
x = vgg_block(x, 128, 3, 'conv4_1', **params_conv)
x = vgg_block(x, 128, 3, 'conv4_2', **params_conv)
return x
def forward(self, x):
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
# 2. Common Networks Layers
self.conv1 = nn.conv2d(inputs=x,
filters=32,
kernel_size=[3, 3],
activation=tf.nn.relu)
self.conv1 = nn.max_pooling2d(self.conv1, pool_size=3, strides=2)
self.conv2 = nn.conv2d(inputs=self.conv1,
filters=64,
kernel_size=[3, 3],
activation=tf.nn.relu)
self.conv2 = nn.max_pooling2d(self.conv2, pool_size=3, strides=2)
self.conv3 = nn.conv2d(inputs=self.conv2,
filters=64,
kernel_size=[3, 3],
activation=tf.nn.relu)
self.conv3 = nn.max_pooling2d(self.conv3, pool_size=2, strides=2)
self.conv4 = nn.conv2d(inputs=self.conv3,
filters=128,
kernel_size=[2, 2],
activation=tf.nn.relu)
self.conv5 = nn.conv2d(inputs=self.conv4,
filters=64,
kernel_size=[2, 2],
activation=tf.nn.relu)
self.conv6_1 = nn.conv2d(inputs=self.conv5,
filters=1,
kernel_size=[1, 1],
activation=tf.nn.sigmoid)
self.conv6_2 = nn.conv2d(inputs=self.conv5,
filters=4,
kernel_size=[1, 1])
return self.conv6_1, self.conv6_2
def _model(self, inputs, mode, **config):
x = inputs['image']
if config['data_format'] == 'channels_first':
x = tf.transpose(x, [0, 3, 1, 2])
params = {'padding': 'SAME', 'data_format': config['data_format']}
x = tfl.conv2d(x, 32, 5, activation=tf.nn.relu, name='conv1', **params)
x = tfl.max_pooling2d(x, 2, 2, name='pool1', **params)
x = tfl.conv2d(x, 64, 5, activation=tf.nn.relu, name='conv2', **params)
x = tfl.max_pooling2d(x, 2, 2, name='pool2', **params)
x = tfl.flatten(x)
x = tfl.dense(x, 1024, activation=tf.nn.relu, name='fc1')
x = tfl.dense(x, 10, name='fc2')
if mode == Mode.TRAIN:
return {'logits': x}
else:
return {
'logits': x,
'prob': tf.nn.softmax(x),
'pred': tf.argmax(x, axis=-1)
}
def f_net(inputs):
inputs = inputs[0]
inputs = inputs / 128 - 1.0
# (640, 640, 3*n) -> ()
with tf.device('/gpu:0'):
conv1 = layers.conv2d(
inputs=inputs, filters=16, kernel_size=(8, 8), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2),
activation=tf.nn.relu, name='conv1')
print conv1.shape
pool1 = layers.max_pooling2d(
inputs=conv1, pool_size=3, strides=4, name='pool1')
print pool1.shape
conv2 = layers.conv2d(
inputs=pool1, filters=16, kernel_size=(5, 5), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2),
activation=tf.nn.relu, name='conv2')
print conv2.shape
pool2 = layers.max_pooling2d(
inputs=conv2, pool_size=3, strides=3, name='pool2')
print pool2.shape
conv3 = layers.conv2d(
inputs=pool2, filters=64, kernel_size=(3, 3), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2),
activation=tf.nn.relu, name='conv3')
print conv3.shape
pool3 = layers.max_pooling2d(
inputs=conv3, pool_size=3, strides=2, name='pool3', )
print pool3.shape
depth = pool3.get_shape()[1:].num_elements()
inputs = tf.reshape(pool3, shape=[-1, depth])
print inputs.shape
hid1 = layers.dense(
inputs=inputs, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=1e-2), name='hid1')
print hid1.shape
hid2 = layers.dense(
inputs=hid1, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=1e-2), name='hid2_adv')
print hid2.shape
adv = layers.dense(
inputs=hid2, units=len(AGENT_ACTIONS), activation=None,
kernel_initializer=tf.random_uniform_initializer(-3e-3, 3e-3),
kernel_regularizer=l2_regularizer(scale=1e-2), name='adv')
print adv.shape
hid2 = layers.dense(
inputs=hid1, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=1e-2), name='hid2_v')
print hid2.shape
v = layers.dense(
inputs=hid2, units=1, activation=None,
kernel_initializer=tf.random_uniform_initializer(-3e-3, 3e-3),
kernel_regularizer=l2_regularizer(scale=1e-2), name='v')
print v.shape
q = tf.add(adv, v, name='q')
print q.shape
return {"q": q}
def f_net(inputs):
"""
action_num is set 5.
:param inputs:
:return:
"""
inputs = inputs[0]
inputs = inputs/128 - 1.0
action_num = 5
l2 = 1e-4
# (350, 350, 3*n) -> ()
conv1 = layers.conv2d(
inputs=inputs, filters=16, kernel_size=(8, 8), strides=1,
kernel_regularizer=l2_regularizer(scale=l2),
activation=tf.nn.relu, name='conv1')
print conv1.shape
pool1 = layers.max_pooling2d(
inputs=conv1, pool_size=3, strides=4, name='pool1')
print pool1.shape
conv2 = layers.conv2d(
inputs=pool1, filters=16, kernel_size=(5, 5), strides=1,
kernel_regularizer=l2_regularizer(scale=l2),
activation=tf.nn.relu, name='conv2')
print conv2.shape
pool2 = layers.max_pooling2d(
inputs=conv2, pool_size=3, strides=3, name='pool2')
print pool2.shape
conv3 = layers.conv2d(
inputs=pool2, filters=64, kernel_size=(3, 3), strides=1,
kernel_regularizer=l2_regularizer(scale=l2),
activation=tf.nn.relu, name='conv3')
print conv3.shape
pool3 = layers.max_pooling2d(
inputs=conv3, pool_size=3, strides=2, name='pool3',)
print pool3.shape
depth = pool3.get_shape()[1:].num_elements()
inputs = tf.reshape(pool3, shape=[-1, depth])
inputs = tf.stop_gradient(inputs)
print inputs.shape
hid1 = layers.dense(
inputs=inputs, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=l2), name='hid1')
print hid1.shape
hid2 = layers.dense(
inputs=hid1, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=l2), name='hid2')
print hid2.shape
pi = layers.dense(
inputs=hid2, units=action_num, activation=tf.nn.softmax,
kernel_regularizer=l2_regularizer(scale=l2), name='pi')
pi = tf.stop_gradient(pi)
q = layers.dense(inputs=hid2, units=action_num, kernel_regularizer=l2_regularizer(scale=l2), name='q')
return {"pi": pi, "q": q}
def model(inputs, is_training, data_format, num_classes):
net = conv2d(inputs=inputs,
filters=96,
kernel_size=[7, 7],
strides=2,
padding='valid',
data_format=data_format,
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.variance_scaling_initializer(),
bias_initializer=tf.zeros_initializer())
net = max_pooling2d(inputs=net,
pool_size=[3, 3],
strides=2,
data_format=data_format)
net = fire_module(net, 16, 64, data_format)
net = fire_module(net, 16, 64, data_format)
net = fire_module(net, 32, 128, data_format)
net = max_pooling2d(inputs=net,
pool_size=[3, 3],
strides=2,
data_format=data_format)
net = fire_module(net, 32, 128, data_format)
net = fire_module(net, 48, 192, data_format)
net = fire_module(net, 48, 192, data_format)
net = fire_module(net, 64, 256, data_format)
net = max_pooling2d(inputs=net,
pool_size=[3, 3],
strides=2,
data_format=data_format)
net = fire_module(net, 64, 256, data_format)
net = dropout(inputs=net, rate=0.5, training=is_training)
net = conv2d(
inputs=net,
filters=num_classes,
kernel_size=[1, 1],
strides=1,
padding='valid', # no padding eqv. to pad=1 for 1x1 conv?
data_format=data_format,
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.initializers.random_normal(mean=0.0,
stddev=0.01),
bias_initializer=tf.zeros_initializer())
net = average_pooling2d(inputs=net,
pool_size=[13, 13],
strides=1,
data_format=data_format)
# TODO fix for data_format later
logits = tf.squeeze(net, [2, 3])
return logits
def generator_net(inputs, scope, reuse=None, rgb=False): output_channels = 3 if rgb else 1 with tf.variable_scope(scope, reuse=reuse): # branch 1 ( color reconstruction) cv1 = conv2d(inputs, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv9_i') cv1_r = leaky_relu(cv1) res1_c = conv2d(cv1_r, filters=16, kernel_size=5, strides=1, padding='same', activation=None, name='conv3a_1') res1_b = batch_normalization(res1_c) res1_r = leaky_relu(res1_b) res1_d = conv2d(res1_r, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3b_1') res1 = batch_normalization(res1_d) sum1 = cv1 + res1 res2_c = conv2d(sum1, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3a_2') res2_b = batch_normalization(res2_c) res2_r = leaky_relu(res2_b) res2_d = conv2d(res2_r, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3b_2') res2 = batch_normalization(res2_d) br1 = sum1 + res2 # branch 2 (features extraction) br2 = conv2d(inputs, filters=16, kernel_size=5, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf1') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool1') br2 = conv2d(br2, filters=16, kernel_size=3, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf2') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool2a') br2 = conv2d(br2, filters=16, kernel_size=3, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf3') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool2') print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_1") print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_2") print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_3") print(br2.shape) # concatenate branches and reconstruct image sum3 = tf.concat((br1, br2), axis=3); model = conv2d(sum3, filters=output_channels, kernel_size=3, strides=1, padding='same', activation=None, name='conv9_f') return model
def create_resnet(x, num_blocks, num_outputs, training, activation=tf.nn.elu, output_activation=None, bottleneck=False, name=None, reuse=False):
create_block = create_bottleneck_block if bottleneck else create_resnet_block
unit = 256 if bottleneck else 64
name = "ResNET" if name is None else name
with tf.variable_scope(name, reuse=reuse):
b = layers.conv2d(x, 64, kernel_size=7, strides=2, activation=activation, padding="SAME", kernel_initializer=xavier_initializer())
b = layers.max_pooling2d(b, pool_size=3, strides=2, padding="SAME")
for i, num_repeats in enumerate(num_blocks):
b = create_block(b, unit*2**i, training, activation, skip_connection=True)
for _ in range(num_repeats-1):
b = create_block(b, unit*2**i, training, activation)
b = downsample(b, unit*2**(i+1), training)
# use global average pooling
b = layers.conv2d(b, num_outputs, kernel_size=1, activation=None, padding="SAME")
fts = tf.reduce_mean(b, [1, 2])
if output_activation is not None:
fts = output_activation(fts)
return fts
def __call__(self):
""" Builds the network. """
x = conv2d(self.img_input, self.nb_filter, self.initial_kernel, kernel_initializer='he_normal', padding='same',
strides=self.initial_strides, use_bias=False, **self.conv_kwargs)
if self.subsample_initial_block:
x = batch_normalization(x, **self.bn_kwargs)
x = tf.nn.relu(x)
x = max_pooling2d(x, (3, 3), data_format=self.data_format, strides=(2, 2), padding='same')
# Add dense blocks
nb_filter = self.nb_filter
for block_idx in range(self.nb_dense_block - 1):
with tf.variable_scope('denseblock_{}'.format(block_idx)):
x, nb_filter = self._dense_block(x, self.nb_layers[block_idx], nb_filter)
# add transition_block
x = self._transition_block(x, nb_filter)
nb_filter = int(nb_filter * self.compression)
# The last dense_block does not have a transition_block
x, nb_filter = self._dense_block(x, self.final_nb_layer, self.nb_filter)
x = batch_normalization(x, **self.bn_kwargs)
x = tf.nn.relu(x)
x = GlobalAveragePooling2D(data_format=self.data_format)(x)
if self.include_top:
x = dense(x, self.nb_classes)
return x
def inceptionBlock(X_in, c1, c3_r, c3, c5_r, c5, p3_r, ns=None):
""" inception building block
Args:
X_in the input tensor
c1...p3_r conv layer filter size of inception,
'_r' means reduction 1x1 conv layer
ns name scope
"""
global inceptionCount
inceptionCount += 1
if ns is None:
ns = "inception{}".format(inceptionCount)
with tf.name_scope(ns):
conv1 = conv2d(X_in, c1, 1, padding='SAME')
conv3 = conv2d(X_in, c3_r, 1, padding='SAME')
conv3 = tf.nn.relu(conv3)
conv3 = conv2d(conv3, c3, 3, padding='SAME')
conv5 = conv2d(X_in, c5_r, 1, padding='SAME')
conv5 = tf.nn.relu(conv5)
conv5 = conv2d(conv5, c5, 5, padding='SAME')
maxpool = max_pooling2d(X_in, 3, 1, padding='SAME')
maxpool = conv2d(maxpool, p3_r, 1, padding='SAME')
logits = tf.concat([conv1, conv3, conv5, maxpool], 3)
logits = tf.nn.relu(logits)
return logits
def __init__(self, state_size, action_size):
self.states = tf.placeholder(tf.float32,
shape=[None, *state_size],
name="states")
self.labels = tf.placeholder(
tf.int32, shape=[None, 1],
name="labels") # size 1 because sparse loss is used.
# conv1 = layers.conv2d(self.states, filters=16, kernel_size=(8, 8), strides=(4, 4), activation='relu',
# name="conv1"),
# conv2 = layers.conv2d(conv1, filters=32, kernel_size=(4, 4), strides=(2, 2), activation='relu', name="conv2"),
# flatten = layers.flatten(conv2),
# dense = layers.dense(flatten, 256, activation='relu', name="features"),
#
# self.logits = layers.dense(dense, action_size, name="logits")
# self.value = layers.dense(dense, 1, name="values")
# conv1 = conv2d(self.states, filters=32, kernel_size=(3, 3), name='conv1')
with tf.variable_scope('layers'):
conv1 = layers.conv2d(self.states,
filters=32,
kernel_size=(3, 3),
activation='relu',
name='conv1')
conv2 = layers.conv2d(conv1,
filters=64,
kernel_size=(3, 3),
activation='relu',
name='conv2')
max_pool = layers.max_pooling2d(conv2, 2, 1, name='max_pool')
drop_1 = layers.dropout(max_pool, 0.25, name='drop1')
flatten = layers.flatten(drop_1, name='flatten')
dense = layers.dense(flatten, 128, activation='relu', name='dense')
drop2 = layers.dropout(dense, 0.5, name='drop2')
logits = layers.dense(drop2,
action_size,
activation='softmax',
name='logits')
self.output = tf.nn.softmax(logits, name='output')
# tf.one_hot(tf.arg_max(self.output, 1), depth=10)
print(tf.arg_max(self.output, 1))
self.test = tf.one_hot(tf.arg_max(self.output, 1), depth=10)
# input()
self.cost = tf.losses.sparse_softmax_cross_entropy(self.labels, logits)
self.acc, self.acc_op = tf.metrics.accuracy(self.labels,
tf.arg_max(self.output, 1))
# self.grad = tf.gradients(self.cost, self.states, stop_gradients=[self.states])
self.grad = tf.gradients(self.cost, tf.trainable_variables())
self.optimizer = tf.train.AdamOptimizer()
# print(tf.trainable_variables())
# print(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='layers'))
# print(self.grad)
self.apply_grad = self.optimizer.apply_gradients(
zip(
self.grad,
tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='layers')))
def mixed_7a(self, x):
with tf.variable_scope("mixed_7a"):
with tf.variable_scope("branch0"):
x0 = self.basic_conv2d(x,
256,
kernel_size=1,
stride=1,
padding="same",
namescope="conv1",
use_bias=False)
x0 = self.basic_conv2d(x0,
384,
kernel_size=3,
stride=2,
padding="same",
namescope="conv2",
use_bias=False)
with tf.variable_scope("branch1"):
x1 = self.basic_conv2d(x,
256,
kernel_size=1,
stride=1,
padding="same",
namescope="conv1",
use_bias=False)
x1 = self.basic_conv2d(x1,
288,
kernel_size=3,
stride=2,
padding="same",
namescope="conv2",
use_bias=False)
with tf.variable_scope("branch2"):
x2 = self.basic_conv2d(x,
256,
kernel_size=1,
stride=1,
padding="same",
namescope="conv1",
use_bias=False)
x2 = self.basic_conv2d(x2,
288,
kernel_size=3,
stride=1,
padding="same",
namescope="conv2",
use_bias=False)
x2 = self.basic_conv2d(x2,
320,
kernel_size=3,
stride=2,
padding="same",
namescope="conv3",
use_bias=False)
with tf.variable_scope("branch3"):
x3 = layers.max_pooling2d(x, 3, strides=2, padding="same")
x = tf.concat([x0, x1, x2, x3], axis=-1)
x = layers.dropout(x, noise_shape=[None, 1, 1, None])
return x
def f_net(inputs, num_outputs, is_training):
inputs = inputs/128 - 1.0
conv1 = layers.conv2d(
inputs=inputs, filters=16, kernel_size=(8, 8), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2), name='conv1'
)
pool1 = layers.max_pooling2d(
inputs=conv1, pool_size=3, strides=4, name='pool1'
)
conv2 = layers.conv2d(
inputs=pool1, filters=16, kernel_size=(5, 5), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2), name='conv2'
)
pool2 = layers.max_pooling2d(
inputs=conv2, pool_size=3, strides=3, name='pool2'
)
conv3 = layers.conv2d(
inputs=pool2, filters=64, kernel_size=(3, 3), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2), name='conv3'
)
pool3 = layers.max_pooling2d(
inputs=conv3, pool_size=3, strides=8, name='pool3',
)
conv4 = layers.conv2d(
inputs=pool3, filters=64, kernel_size=(3, 3), strides=1,
kernel_regularizer=l2_regularizer(scale=1e-2), name='conv4'
)
pool4 = layers.max_pooling2d(
inputs=conv4, pool_size=3, strides=8, name='pool4'
)
depth = pool4.get_shape()[1:].num_elements()
inputs = tf.reshape(pool4, shape=[-1, depth])
hid1 = layers.dense(
inputs=inputs, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=1e-2), name='hid1'
)
hid2 = layers.dense(
inputs=hid1, units=256, activation=tf.nn.relu,
kernel_regularizer=l2_regularizer(scale=1e-2), name='hid2'
)
q = layers.dense(
inputs=hid2, units=num_outputs, activation=None,
kernel_regularizer=l2_regularizer(scale=1e-2), name='q'
)
q = tf.squeeze(q, name='out_sqz')
return q
def net(inputs, is_training):
x = tf.reshape(inputs, [-1, 28, 28, 1])
dropout_params = {
'init_min': 0.1,
'init_max': 0.1,
'weight_regularizer': 1e-6,
'dropout_regularizer': 1e-5,
'training': is_training
}
x, reg = concrete_dropout(x, name='conv1_dropout', **dropout_params)
x = tfl.conv2d(x,
32,
5,
activation=tf.nn.relu,
padding='SAME',
kernel_regularizer=reg,
bias_regularizer=reg,
name='conv1')
x = tfl.max_pooling2d(x, 2, 2, padding='SAME', name='pool1')
x, reg = concrete_dropout(x, name='conv2_dropout', **dropout_params)
x = tfl.conv2d(x,
64,
5,
activation=tf.nn.relu,
padding='SAME',
kernel_regularizer=reg,
bias_regularizer=reg,
name='conv2')
x = tfl.max_pooling2d(x, 2, 2, padding='SAME', name='pool2')
x = tf.reshape(x, [-1, 7 * 7 * 64], name='flatten')
x, reg = concrete_dropout(x, name='fc1_dropout', **dropout_params)
x = tfl.dense(x,
1024,
activation=tf.nn.relu,
name='fc1',
kernel_regularizer=reg,
bias_regularizer=reg)
outputs = tfl.dense(x, 10, name='fc2')
return outputs
def model_old(self): # not use roi pooling
h_00 = tl.batch_normalization(tf.cast(self.m_image_data,
tf.float32)) # 256*256
h_01 = tl.conv2d(h_00, 8, 3) # 256*256
h_02 = tl.conv2d(h_01, 8, 3) # 256*256
h_03 = tl.max_pooling2d(h_02) # 128*128
h_04 = tl.conv2d(h_03, 16, 3) # 128*128
h_05 = tl.conv2d(h_04, 16, 3) # 128*128
h_06 = tl.max_pooling2d(h_05) # 64*64
h_07 = tl.conv2d(h_06, 32, 3) # 64*64
h_08 = tl.conv2d(h_07, 32, 3) # 64*64
h_09 = tl.max_pooling2d(h_08) # 32*32
h_10 = tl.conv2d(h_09, 32, 3) # 32*32
h_11 = tl.conv2d(h_10, 32, 3) # 32*32
h_12 = tl.max_pooling2d(h_11) # 16*16
conv_predict_label = tl.conv2d(h_12,
len(self.config.anchors) * 2,
1,
activation_fn=None)
conv_predict_proposal_regress = tl.conv2d(h_12,
len(self.config.anchors) * 4,
1,
activation_fn=None)
predict_label_logits = tf.reshape(conv_predict_label,
[-1, self.config.sample_size, 2])
self.predict_label = tf.nn.softmax(predict_label_logits)
self.predict_proposal_regress = tf.reshape(
conv_predict_proposal_regress, [-1, self.config.sample_size, 4])
self.loss_label = tf.losses.softmax_cross_entropy(
self.m_label, predict_label_logits, weights=self.m_label_weight)
self.loss_proposal_regress = tf.losses.mean_squared_error(
self.m_proposal_regress,
self.predict_proposal_regress,
weights=self.m_proposal_regress_weight)
self.loss = self.loss_label + self.loss_proposal_regress
return self.predict_label, self.predict_proposal_regress, self.loss
def DownSampler(input, num_output): ''' 下采样单元，用卷积层和池化层拼接即可 num_output: 输出的特征数 ''' net = conv2d(input, num_output, kernel_size = (3, 3), strides = (1, 1), padding = "SAME") net = max_pooling2d(net, pool_size = (2,2), strides = 2) return net
def vgg(inputs, ys, knn_inputs, knn_ys):
# block 1
x = conv2d(inputs, filters=64, kernel_size=(3, 3), padding='same', activation=ReLU)
x = conv2d(x, filters=64, kernel_size=(3, 3), padding='same', activation=ReLU)
x = max_pooling2d(x, pool_size=(2, 2), strides=(2, 2))
# block 2
x = conv2d(x, filters=128, kernel_size=(3, 3), padding='same', activation=ReLU)
x = conv2d(x, filters=128, kernel_size=(3, 3), padding='same', activation=ReLU)
x = max_pooling2d(x, pool_size=(2, 2), strides=(2, 2))
# final
x = flatten(x)
x = dense(x, units=1024, activation=ReLU)
x = dropout(x, rate=0.5, training=)
x = dense(x, units=1024, activation=ReLU)
x = dense(x, units=self.data_loader.num_outputs)
return x
def lenet5_2d(x, keep_prob, is_train, conf):
"""modified LeNet5 (ReLU instead sigmoid, dropout after FC layers)"""
def _dense(inputs, neurons, name, activation=tf.nn.relu):
"""wrapper function for dense layer"""
w_ini, b_ini, r_ini = initializers(conf.bias_init, conf.l2_str)
return dense(inputs,
neurons,
activation=activation,
bias_initializer=b_ini,
kernel_initializer=w_ini,
kernel_regularizer=r_ini,
name=name)
def _conv2d(inputs, n_filters, size, name, padding='VALID'):
"""wrapper function for 2D convolutional layer"""
w_ini, b_ini, r_ini = initializers(conf.bias_init, conf.l2_str)
return conv2d(inputs,
n_filters,
size,
padding=padding,
activation=tf.nn.relu,
bias_initializer=b_ini,
kernel_initializer=w_ini,
kernel_regularizer=r_ini,
name=name)
# split signals according to length and channels
x_multichannel = tf.reshape(
x, [-1, conf.img_dim[0], conf.img_dim[1], conf.num_ch])
# inference
net = _conv2d(x_multichannel, 6, 5, 'conv1', padding='SAME')
net = max_pooling2d(net, [2, 2], 2, name='pool1')
net = _conv2d(net, 16, 5, 'conv2')
net = max_pooling2d(net, [2, 2], 2, name='pool2')
net = flatten(net, 'flat_layer')
net = _dense(net, 120, 'fc1')
net = tf.nn.dropout(net, keep_prob, name='dropout1')
net = _dense(net, 84, 'fc2')
net = tf.nn.dropout(net, keep_prob, name='dropout2')
# output layer (softmax applied at loss function)
logits = _dense(net, conf.num_cl, activation=None, name='logits')
return logits
def __init__(self, images, params, reuse=False):
with tf.variable_scope("RGBNetwork", reuse=reuse) as scope:
x = tf.reshape(images, [-1, 200, 200, 3])
x = conv2d(x, 64, (10, 10), strides=(4, 4), padding='same')
x = leaky_relu(x)
x = max_pooling2d(x, (5, 5), strides=(2, 2), padding='same')
x = conv2d(x, 128, (2, 2), strides=(2, 2), padding='same')
x = leaky_relu(x)
x = max_pooling2d(x, (2, 2), strides=(2, 2), padding='same')
x = conv2d(x, 256, (3, 3), strides=(1, 1), padding='same')
x = leaky_relu(x)
x = conv2d(x, 512, (3, 3), strides=(1, 1), padding='same')
x = leaky_relu(x)
x = flatten(x)
x = dense(x, 256, activation=leaky_relu)
x = dense(x, 200 * 200, activation=None)
x = tf.reshape(x, [-1, 200, 200])
self.output = tf.cast(x, tf.float32)
self.parameters = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="SimpleNetwork")
def forward(self, x):
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
# 2. Common Networks Layers
self.conv1 = nn.conv2d(inputs=x,
filters=28,
kernel_size=[3, 3],
activation=tf.nn.relu)
self.conv1 = nn.max_pooling2d(self.conv1, pool_size=3, strides=2)
self.conv2 = nn.conv2d(inputs=self.conv1,
filters=48,
kernel_size=[3, 3],
activation=tf.nn.relu)
self.conv2 = nn.max_pooling2d(self.conv2, pool_size=3, strides=2)
self.conv3 = nn.conv2d(inputs=self.conv2,
filters=64,
kernel_size=[2, 2],
activation=tf.nn.relu)
# self.conv4 = nn.conv2d(inputs=self.conv3, filters=64, kernel_size=[2, 2], activation=tf.nn.relu)
# self.conv5_1 = nn.conv2d(inputs=self.conv4, filters=1, kernel_size=[1, 1], activation=tf.nn.sigmoid)
# self.conv5_2 = nn.conv2d(inputs=self.conv4, filters=4, kernel_size=[1, 1])
#
# return self.conv5_1, self.conv5_2
# This is a wrong method, because the batch number is unknown
# self.conv_flat = tf.reshape(self.conv3, [self.conv3.get_shape()[0], -1])
self.conv_flat = tf.reshape(self.conv3, (-1, 2 * 2 * 64))
self.fc1 = tf.layers.dense(inputs=self.conv_flat,
units=128,
activation=tf.nn.relu)
self.fc1 = tf.layers.dense(inputs=self.fc1,
units=64,
activation=tf.nn.relu)
self.fc1 = tf.layers.dense(inputs=self.fc1,
units=32,
activation=tf.nn.relu)
self.fc2_1 = tf.layers.dense(inputs=self.fc1,
units=1,
activation=tf.nn.sigmoid)
self.fc2_2 = tf.layers.dense(inputs=self.fc1, units=4)
return self.fc2_1, self.fc2_2
def inference(self, input):
conv = conv2d(input,
self.noutput,
kernel_size=(3, 3),
strides=(2, 2),
padding="SAME")
maxpool = max_pooling2d(input, pool_size=(2, 2), strides=2)
concat = tf.concat([conv, maxpool], -1)
bn = batch_normalization(concat, self.training)
return tf.nn.relu(bn)
def classifier_net(inputs, scope, reuse=None): with tf.variable_scope(scope, reuse=reuse): net = conv2d(inputs, filters=32, kernel_size=5, strides=1, activation=tf.nn.leaky_relu, name='conv1') net = conv2d(net, filters=64, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv2') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool1') net = conv2d(net, filters=64, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv3') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool2a') net = conv2d(net, filters=64, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv4') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool2') net = conv2d(net, filters=64, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv5') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool3a') net = conv2d(net, filters=64, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv6') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool3') net = conv2d(net, filters=32, kernel_size=3, strides=1, activation=tf.nn.leaky_relu, name='conv7') net = max_pooling2d(net, pool_size=2, strides=2, padding='same', name='maxpool4') net = flatten(net, name='flatten') # net = dense(net, 1024, activation=tf.nn.leaky_relu, name='dense1') # net = batch_normalization(net) net = dense(net, 512, activation=tf.nn.leaky_relu, name='dense2') # net = batch_normalization(net) net = dense(net, 256, activation=tf.nn.leaky_relu, name='dense3') # net = batch_normalization(net) net = dense(net, 128, activation=tf.nn.leaky_relu, name='dense4') net = dense(net, 1, activation=tf.nn.sigmoid, name='out') return net
def cnn_model(features, mode):
"""
cnn model structure
:param features: images
:return: predicts
"""
input_layer = tf.reshape(features, shape=[-1, 28, 28, 1], name='input')
# conv1
# trainable can change in middle of training
conv1 = layers.conv2d(inputs=input_layer, filters=32, kernel_size=[3, 3], padding="same", activation=tf.nn.relu,
name='conv1')
pool1 = layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2, name='pool1')
# conv2
conv2 = layers.conv2d(inputs=pool1, filters=64, kernel_size=[5, 5], padding="same", activation=tf.nn.relu,
name='conv2')
pool2 = layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2, name='pool2')
# fully connected layer
pool2_flat = layers.flatten(pool2, name='flatten')
dense = layers.dense(inputs=pool2_flat, units=512, activation=tf.nn.relu, name='dense_layer')
dropout = layers.dropout(inputs=dense, rate=0.4, training=(mode == tf.estimator.ModeKeys.TRAIN), name='dropout')
# output layer
logits = tf.layers.dense(inputs=dropout, units=10, name='logits')
return logits
def _squeezenet(images, num_classes=1000):
#print("image",images)
net = conv2d(images, 96, [7, 7], strides=2, padding='SAME', kernel_regularizer=regularizer, activation=tf.nn.relu)
net = max_pooling2d(net, [3, 3], strides=2)
#print("@here ",net)
net = fire_module(net, 16, 64, scope='fire2')
net = fire_module(net, 16, 64, scope='fire3')
net = fire_module(net, 32, 128, scope='fire4')
net = max_pooling2d(net, [3, 3], strides=2)
net = fire_module(net, 32, 128, scope='fire5')
net = fire_module(net, 48, 192, scope='fire6')
net = fire_module(net, 48, 192, scope='fire7')
net = fire_module(net, 64, 256, scope='fire8')
net = max_pooling2d(net, [3, 3], strides=2)
net = fire_module(net, 64, 256, scope='fire9')
#print("@here fire_module ",net)
net = conv2d(net, num_classes, [1, 1], strides=1, padding='SAME')
#print("@here conv2d ",net)
net = max_pooling2d(net, [13, 13], strides=1)
#print("@here pooling ",net)
logits = tf.squeeze(net, [1,2], name='logits')
#print("@here logits ",logits)
return logits
def build(self, input):
self.conv1 = self.conv_block(input, 64, "conv1")
pool1 = max_pooling2d(self.conv1, pool_size=(2, 2), strides=2)
drop1 = dropout(pool1, .25)
self.conv2 = self.conv_block(drop1, 128, 'conv2')
pool2 = max_pooling2d(self.conv2, pool_size=(2, 2), strides=2)
drop2 = dropout(pool2, .25)
self.conv3 = self.conv_block(drop2, 256, 'conv3')
pool3 = max_pooling2d(self.conv3, pool_size=(2, 2), strides=2)
drop3 = dropout(pool3, .25)
self.conv4 = self.conv_block(drop3, 512, 'conv4')
pool4 = max_pooling2d(self.conv4, pool_size=(2, 2), strides=2)
drop4 = dropout(pool4, .25)
self.conv5 = conv2d(drop4,
1024,
kernel_size=(3, 3),
activation=tf.nn.relu,
padding='SAME')
self.conv5_2 = conv2d(self.conv5,
1024,
kernel_size=(3, 3),
activation=tf.nn.relu,
padding='SAME')
self.deconv4 = self.deconv_block(self.conv5_2, 512, self.conv4, "SAME",
'deconv4')
self.deconv3 = self.deconv_block(self.deconv4, 256, self.conv3,
"VALID", 'deconv3')
self.deconv2 = self.deconv_block(self.deconv3, 128, self.conv2, "SAME",
'deconv2')
self.deconv1 = self.deconv_block(self.deconv2, 64, self.conv1, "VALID",
'deconv1')
self.output = conv2d(self.deconv1,
filters=1,
kernel_size=1,
name='logits')
def forward(self, x, is_training):
# Assumes [Batch, Time, Freq, Chan]
self.inputs = layers.conv2d(
x, filters=128, kernel_size=[3, 3], strides=1,
activation=tf.nn.relu
)
print(x.shape)
x = layers.conv2d(
x, filters=128, kernel_size=[3, 3], strides=1,
activation=tf.nn.relu
)
print(x.shape)
x = layers.max_pooling2d(
x, pool_size=[3, 3], strides=[3, 3]
)
print(x.shape)
x = layers.conv2d(
x, filters=256, kernel_size=[3, 3], strides=1,
activation=tf.nn.relu
)
print(x.shape)
x = layers.conv2d(
x, filters=256, kernel_size=[3, 3], strides=1,
activation=tf.nn.relu
)
print(x.shape)
x = layers.max_pooling2d(
x, pool_size=[3, 3], strides=[3, 3]
)
print(x.shape)
x = tf.contrib.layers.flatten(x)
print(x.shape)
self.raw_scores = layers.dense( # Linear map to output
inputs=x, units=self.FLAGS.num_classes, activation=None
)
return self.raw_scores
def _encode_brick(incoming,
nb_filters,
is_training,
scope,
use_bn=True,
trainable=True):
""" Encoding brick: conv --> conv --> max pool.
"""
with tf.variable_scope(scope):
conv1 = layers.conv2d(incoming,
filters=nb_filters,
kernel_size=3,
strides=1,
padding='same',
kernel_initializer=he_init,
bias_initializer=b_init,
trainable=trainable)
if use_bn:
conv1 = layers.batch_normalization(conv1,
training=is_training,
trainable=trainable)
conv1_act = tf.nn.relu(conv1)
conv2 = layers.conv2d(conv1_act,
filters=nb_filters,
kernel_size=3,
strides=1,
padding='same',
kernel_initializer=he_init,
bias_initializer=b_init,
trainable=trainable)
if use_bn:
conv2 = layers.batch_normalization(conv2,
training=is_training,
trainable=trainable)
conv2_act = tf.nn.relu(conv2)
pool = layers.max_pooling2d(conv2_act,
pool_size=2,
strides=2,
padding='same')
with tf.variable_scope('concat_layer_out'):
concat_layer_out = conv2_act
return pool, concat_layer_out
