def shallow_net_5x5(x):
net = L.conv(x, name="conv_sn5x5_1", kh=5, kw=5, n_out=24)
net = L.pool(net, name="pool_sn5x5_1", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn5x5_2", kw=3, kh=3, n_out=48)
net = L.pool(net, name="pool_sn5x5_2", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn5x5_3", kw=3, kh=3, n_out=24)
net = L.conv(net, name="conv_sn5x5_4", kw=3, kh=3, n_out=12)
return net
def shallow_net_7x7(x):
net = L.conv(x, name="conv_sn7x7_1", kh=7, kw=7, n_out=20)
net = L.pool(net, name="pool_sn7x7_1", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn7x7_2", kw=5, kh=5, n_out=40)
net = L.pool(net, name="pool_sn7x7_2", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn7x7_3", kw=5, kh=5, n_out=20)
net = L.conv(net, name="conv_sn7x7_4", kw=5, kh=5, n_out=10)
return net
def shallow_net_9x9(x):
net = L.conv(x, name="conv_sn9x9_1", kh=9, kw=9, n_out=16)
net = L.pool(net, name="pool_sn9x9_1", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn9x9_2", kw=7, kh=7, n_out=32)
net = L.pool(net, name="pool_sn9x9_2", kh=2, kw=2, dw=2, dh=2)
net = L.conv(net, name="conv_sn9x9_3", kw=7, kh=7, n_out=16)
net = L.conv(net, name="conv_sn9x9_4", kw=7, kh=7, n_out=8)
return net
def intfVGG_use_layer(input_tensor,
n_classes=1000,
rgb_mean=None,
training=True):
# assuming 224x224x3 input_tensor
# define image mean
if rgb_mean is None:
rgb_mean = np.array([116.779, 123.68, 103.939], dtype=np.float32)
mu = tf.constant(rgb_mean, name="rgb_mean")
keep_prob = 0.5
# subtract image mean
net = tf.subtract(input_tensor, mu, name="input_mean_centered")
# block 1 -- outputs 112x112x64
net = L.conv(net, name="conv1_1", kh=3, kw=3, n_out=64)
net = L.conv(net, name="conv1_2", kh=3, kw=3, n_out=64)
net = L.pool(net, name="pool1", kh=2, kw=2, dw=2, dh=2)
# block 2 -- outputs 56x56x128
net = L.conv(net, name="conv2_1", kh=3, kw=3, n_out=128)
net = L.conv(net, name="conv2_2", kh=3, kw=3, n_out=128)
net = L.pool(net, name="pool2", kh=2, kw=2, dh=2, dw=2)
# # block 3 -- outputs 28x28x256
net = L.conv(net, name="conv3_1", kh=3, kw=3, n_out=256)
net = L.conv(net, name="conv3_2", kh=3, kw=3, n_out=256)
net = L.pool(net, name="pool3", kh=2, kw=2, dh=2, dw=2)
# block 4 -- outputs 14x14x512
net = L.conv(net, name="conv4_1", kh=3, kw=3, n_out=512)
net = L.conv(net, name="conv4_2", kh=3, kw=3, n_out=512)
net = L.conv(net, name="conv4_3", kh=3, kw=3, n_out=512)
net = L.pool(net, name="pool4", kh=2, kw=2, dh=2, dw=2)
# block 5 -- outputs 7x7x512
net = L.conv(net, name="conv5_1", kh=3, kw=3, n_out=512)
net = L.conv(net, name="conv5_2", kh=3, kw=3, n_out=512)
net = L.conv(net, name="conv5_3", kh=3, kw=3, n_out=512)
net = L.pool(net, name="pool5", kh=2, kw=2, dw=2, dh=2)
# flatten
flattened_shape = np.prod([s.value for s in net.get_shape()[1:]])
net = tf.reshape(net, [-1, flattened_shape], name="flatten")
# fully connected
net = L.fully_connected(net, name="fc6", n_out=4096)
net = tf.nn.dropout(net, keep_prob)
net = L.fully_connected(net, name="fc7", n_out=4096)
net = tf.nn.dropout(net, keep_prob)
net = L.fully_connected(net, name="fc8", n_out=n_classes)
return net
def BKStart(x, reuse):
with tf.variable_scope('BKS', reuse=reuse):
n = "BKStart_"
x = conv_layer(x, 1, 32, 5, n + "conv_1", 1, pad='SAME')
x = pool(x, 3, 2, name=n + "max_pool_1", pad='SAME', pool='max')
x = conv_layer(x, 32, 32, 4, n + "conv_2", 1, pad='SAME')
x = pool(x, 3, 2, n + "avg_pool_1", pool='avg')
x = conv_layer(x, 32, 64, 5, n + "conv_3", 1, pad='SAME')
x = pool(x, 3, 2, n + "avg_pool_2", pool='avg')
flattened_shape = np.prod([s.value for s in x.get_shape()[1:]])
x = tf.reshape(x, [-1, flattened_shape], name=n + 'flatten')
x = fc_layer(x, 2048, activation='Relu', name=n + 'FC_1')
#x=dropout_layer(x,keep_prob)
logits = fc_layer(x, 7, activation='None', name=n + 'FC_2')
return logits
def VGG16(x, n_classes, keep_prob):
with tf.name_scope('VGG16'):
# Group 1
x = layers.conv('conv1_1', x, 64, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv1_2', x, 64, [3, 3], [1, 1, 1, 1])
with tf.name_scope('pool1'):
x = layers.pool('pool1', x, [1, 2, 2, 1], [1, 2, 2, 1])
# Group 2
x = layers.conv('conv2_1', x, 128, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv2_2', x, 128, [3, 3], [1, 1, 1, 1])
with tf.name_scope('pool2'):
x = layers.pool('pool2', x, [1, 2, 2, 1], [1, 2, 2, 1])
# Group 3
x = layers.conv('conv3_1', x, 256, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv3_2', x, 256, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv3_3', x, 256, [3, 3], [1, 1, 1, 1])
with tf.name_scope('pool3'):
x = layers.pool('pool3', x, [1, 2, 2, 1], [1, 2, 2, 1])
# Group 4
x = layers.conv('conv4_1', x, 512, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv4_2', x, 512, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv4_3', x, 512, [3, 3], [1, 1, 1, 1])
with tf.name_scope('pool4'):
x = layers.pool('pool4', x, [1, 2, 2, 1], [1, 2, 2, 1])
# Group 5
x = layers.conv('conv5_1', x, 512, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv5_2', x, 512, [3, 3], [1, 1, 1, 1])
x = layers.conv('conv5_3', x, 512, [3, 3], [1, 1, 1, 1])
with tf.name_scope('pool5'):
x = layers.pool('pool5', x, [1, 2, 2, 1], [1, 2, 2, 1])
x = layers.fc_layer('fc6', x, 4096)
x = layers.dropout('drop6', x, keep_prob)
x = layers.fc_layer('fc7', x, 4096)
x = layers.dropout('drop7', x, keep_prob)
x = layers.fc_layer('fc8', x, n_classes)
return x
def BKVGG8(x, keep_prob):
n = "BKVGG8_"
x = conv_layer(x, 1, 32, 3, n + "conv_1", 1, pad='SAME')
x = pool(x, 2, 2, name=n + "max_pool_1", pad='SAME', pool='max')
x = conv_layer(x, 32, 64, 3, n + "conv_2", 1, pad='SAME')
x = pool(x, 2, 2, n + "max_pool_1", pool='max')
x = conv_layer(x, 64, 128, 3, n + "conv_3", 1, pad='SAME')
x = pool(x, 2, 2, n + "max_pool_2", pool='max')
x = conv_layer(x, 128, 256, 3, n + "conv_4", 1, pad='SAME')
x = conv_layer(x, 256, 256, 3, n + "conv_5", 1, pad='SAME')
flattened_shape = np.prod([s.value for s in x.get_shape()[1:]])
x = tf.reshape(x, [-1, flattened_shape], name=n + 'flatten')
x = fc_layer(x, 256, activation='Relu', name=n + 'FC_1')
x = dropout_layer(x, keep_prob)
x = fc_layer(x, 256, activation='Relu', name=n + 'FC_2')
x = dropout_layer(x, keep_prob)
logits = fc_layer(x, 7, activation='None', name=n + 'FC_3')
return logits
def inference(input_tensor,
n_classes,
train=True,
regularizer=None,
evaluate=False):
with tf.name_scope('cnn'):
input_tensor = layers.conv('conv1',
input_tensor,
out_channels=32,
kernel_size=[2, 2],
strides=[1, 2, 2, 1],
train=train)
input_tensor = layers.pool('pool1',
input_tensor,
kernel_size=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
by_max=True)
input_tensor = layers.conv('conv2',
input_tensor,
out_channels=64,
kernel_size=[2, 2],
strides=[1, 1, 1, 1],
train=train)
input_tensor = layers.batch_norm(input_tensor)
input_tensor = layers.pool('pool2',
input_tensor,
kernel_size=[1, 2, 2, 1],
strides=[1, 1, 1, 1],
by_max=True)
input_tensor = layers.full_connect('fc1',
input_tensor,
out_nodes=512,
regularizer=regularizer)
output_tensor = layers.full_connect_not_relu('fc2',
input_tensor,
out_nodes=n_classes,
regularizer=regularizer)
return output_tensor
def generator(inp_z, inp_y, reuse=False):
with tf.variable_scope('Generator', reuse=reuse):
inp = tf.concat([inp_z, inp_y], 1)
sz = 4
g1 = linear(inp, 512 * sz * sz, name=gname + 'deconv1')
g1 = batchnorm(g1, is_training=tf.constant(True), name=gname + 'bn1g')
g1 = lrelu(g1, 0.2)
g1_reshaped = tf.reshape(g1, [-1, 512, sz, sz])
print 'genreshape: ' + str(g1_reshaped.get_shape().as_list())
g2 = nnupsampling(g1_reshaped, [8, 8])
g2 = conv2d(g2, nout=512, kernel=3, name=gname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=gname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=256, kernel=3, name=gname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=gname + 'bn3g')
g3 = lrelu(g3, 0.2)
g4 = nnupsampling(g3, [32, 32])
g4 = conv2d(g4, nout=128, kernel=3, name=gname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=gname + 'bn4g')
g4 = lrelu(g4, 0.2)
g4b = conv2d(g4, nout=128, kernel=3, name=gname + 'deconv4b')
g4b = batchnorm(g4b,
is_training=tf.constant(True),
name=gname + 'bn4bg')
g4b = lrelu(g4b, 0.2)
g5 = nnupsampling(g4b, [64, 64])
g5 = conv2d(g5, nout=64, kernel=3, name=gname + 'deconv5')
g5 = batchnorm(g5, is_training=tf.constant(True), name=gname + 'bn5g')
g5 = lrelu(g5, 0.2)
g5b = conv2d(g5, nout=64, kernel=3, name=gname + 'deconv5b')
g5b = batchnorm(g5b,
is_training=tf.constant(True),
name=gname + 'bn5bg')
g5b = lrelu(g5b, 0.2)
g6 = nnupsampling(g5b, [128, 128])
g6 = conv2d(g6, nout=32, kernel=3, name=gname + 'deconv6')
g6 = batchnorm(g6, is_training=tf.constant(True), name=gname + 'bn6g')
g6 = lrelu(g6, 0.2)
g6b = conv2d(g6, nout=3, kernel=3, name=gname + 'deconv6b')
g6b = tf.nn.tanh(g6b)
g6b_64 = pool(g6b, fsize=3, strides=2, op='avg')
return g6b_64, g6b
def stage_1(x,training):
"""The first stage of the network
Args:
x: the input batch
training: whether we are in the process of training or inference
Returns:
The output of the first stage
"""
x = layers.conv(x, "conv1", 2, 3, 24, 3, training)
x = layers.pool(x, "pool1", 2, 3)
x = layers.down_sample_unit(x, "ds1", 24, training)
x = layers.shuffle_net_unit(x, "sn1_1", 48, training)
x = layers.shuffle_net_unit(x, "sn1_2", 48, training)
stage_1_output = layers.shuffle_net_unit(x, "sn1_3", 48, training)
return stage_1_output
def generator(inp_z, inp_y):
with tf.variable_scope('Generator'):
inp = tf.concat([inp_z, inp_y], 1)
g1 = linear(inp, 512 * 4 * 4, name=gname + 'deconv1')
g1 = batchnorm(g1, is_training=tf.constant(True), name=gname + 'bn1g')
g1 = lrelu(g1, 0.2)
g1_reshaped = tf.reshape(g1, [-1, 512, 4, 4])
print 'genreshape: ' + str(g1_reshaped.get_shape().as_list())
g2 = nnupsampling(g1_reshaped, [8, 8])
g2 = conv2d(g2, nout=256, kernel=3, name=gname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=gname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=128, kernel=3, name=gname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=gname + 'bn3g')
g3 = lrelu(g3, 0.2)
g3b = conv2d(g3, nout=128, kernel=3, name=gname + 'deconv3b')
g3b = batchnorm(g3b,
is_training=tf.constant(True),
name=gname + 'bn3bg')
g3b = lrelu(g3b, 0.2)
g4 = nnupsampling(g3b, [32, 32])
g4 = conv2d(g4, nout=64, kernel=3, name=gname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=gname + 'bn4g')
g4 = lrelu(g4, 0.2)
g4b = conv2d(g4, nout=64, kernel=3, name=gname + 'deconv4b')
g4b = batchnorm(g4b,
is_training=tf.constant(True),
name=gname + 'bn4bg')
g4b = lrelu(g4b, 0.2)
g5 = nnupsampling(g4b, [64, 64])
g5 = conv2d(g5, nout=32, kernel=3, name=gname + 'deconv5')
g5 = batchnorm(g5, is_training=tf.constant(True), name=gname + 'bn5g')
g5 = lrelu(g5, 0.2)
g5b = conv2d(g5, nout=3, kernel=3, name=gname + 'deconv5b')
g5b = tf.nn.tanh(g5b)
g5b_32 = pool(g5b, fsize=3, strides=2, op='avg', pad='SAME')
return g5b_32, g5b
def _build_network(self, inputs: tf.Tensor, sparse_layers: list,
activation_fn: callable) -> tf.Tensor:
with tf.variable_scope('network'):
net = inputs
self.weight_tensors = []
bias_initializer = tf.constant_initializer(0.1)
# block 1 -- outputs 112x112x64
net, w, b = L.sparse_conv(net, sparse_layers[0], name="conv1_1")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[1], name="conv1_2")
self.weight_tensors.append(w)
net = L.pool(net, name="pool1", kh=2, kw=2, dw=2, dh=2)
# block 2 -- outputs 56x56x128
net, w, b = L.sparse_conv(net, sparse_layers[2], name="conv2_1")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[3], name="conv2_2")
self.weight_tensors.append(w)
net = L.pool(net, name="pool2", kh=2, kw=2, dh=2, dw=2)
# # block 3 -- outputs 28x28x256
net, w, b = L.sparse_conv(net, sparse_layers[4], name="conv3_1")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[5], name="conv3_2")
self.weight_tensors.append(w)
net = L.pool(net, name="pool3", kh=2, kw=2, dh=2, dw=2)
# block 4 -- outputs 14x14x512
net, w, b = L.sparse_conv(net, sparse_layers[6], name="conv4_1")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[7], name="conv4_2")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[8], name="conv4_3")
self.weight_tensors.append(w)
net = L.pool(net, name="pool4", kh=2, kw=2, dh=2, dw=2)
# block 5 -- outputs 7x7x512
net, w, b = L.sparse_conv(net, sparse_layers[9], name="conv5_1")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[10], name="conv5_2")
self.weight_tensors.append(w)
net, w, b = L.sparse_conv(net, sparse_layers[11], name="conv5_3")
self.weight_tensors.append(w)
net = L.pool(net, name="pool5", kh=2, kw=2, dw=2, dh=2)
# flatten
flattened_shape = np.prod([s.value for s in net.get_shape()[1:]])
net = tf.reshape(net, [-1, flattened_shape], name="flatten")
# fully connected
net, w, b = L.sparse_fully_connected(net,
sparse_layers[12],
name="fc6")
self.weight_tensors.append(w)
net, w, b = L.sparse_fully_connected(net,
sparse_layers[13],
name="fc7")
self.weight_tensors.append(w)
net, w, b = L.sparse_fully_connected(net,
sparse_layers[14],
name="fc8_2",
activation_fn=None)
self.weight_tensors.append(w)
return net
def build(input_tensor, n_classes=1000, rgb_mean=None, training=True):
# assuming 224x224x3 input_tensor
# define image mean
if rgb_mean is None:
rgb_mean = np.array([116.779, 123.68, 103.939], dtype=np.float32)
mu = tf.constant(rgb_mean, name="rgb_mean")
keep_prob = 0.5
# subtract image mean
input_mean_centered = tf.subtract(input_tensor,
mu,
name="input_mean_centered")
# block 1 -- outputs 112x112x64
conv1_1 = L.conv(input_mean_centered, name="conv1_1", kh=3, kw=3, n_out=64)
conv1_2 = L.conv(conv1_1, name="conv1_2", kh=3, kw=3, n_out=64)
pool1 = L.pool(conv1_2, name="pool1", kh=2, kw=2, dw=2, dh=2)
# block 2 -- outputs 56x56x128
conv2_1 = L.conv(pool1, name="conv2_1", kh=3, kw=3, n_out=128)
conv2_2 = L.conv(conv2_1, name="conv2_2", kh=3, kw=3, n_out=128)
pool2 = L.pool(conv2_2, name="pool2", kh=2, kw=2, dh=2, dw=2)
# # block 3 -- outputs 28x28x256
conv3_1 = L.conv(pool2, name="conv3_1", kh=3, kw=3, n_out=256)
conv3_2 = L.conv(conv3_1, name="conv3_2", kh=3, kw=3, n_out=256)
pool3 = L.pool(conv3_2, name="pool3", kh=2, kw=2, dh=2, dw=2)
# block 4 -- outputs 14x14x512
conv4_1 = L.conv(pool3, name="conv4_1", kh=3, kw=3, n_out=512)
conv4_2 = L.conv(conv4_1, name="conv4_2", kh=3, kw=3, n_out=512)
conv4_3 = L.conv(conv4_2, name="conv4_3", kh=3, kw=3, n_out=512)
pool4 = L.pool(conv4_3, name="pool4", kh=2, kw=2, dh=2, dw=2)
# block 5 -- outputs 7x7x512
conv5_1 = L.conv(pool4, name="conv5_1", kh=3, kw=3, n_out=512)
conv5_2 = L.conv(conv5_1, name="conv5_2", kh=3, kw=3, n_out=512)
conv5_3 = L.conv(conv5_2, name="conv5_3", kh=3, kw=3, n_out=512)
# pool5 = L.pool(conv5_3, name="pool5", kh=2, kw=2, dw=2, dh=2)
pool2_max = L.pool(pool2, name="pool2_max", kh=2, kw=2, dw=2, dh=2)
pool2_max_conv = L.conv(pool2_max,
name="pool2_max_conv",
kh=3,
kw=3,
n_out=512)
pool3_conv = L.conv(pool3, name="pool3_conv", kh=3, kw=3, n_out=512)
conv4_3_conv = L.conv(conv4_3, name="conv4_3_conv", kh=3, kw=3, n_out=512)
conv5_3_conv = L.conv(conv5_3, name="conv5_3_conv", kh=3, kw=3, n_out=512)
conv5_3_conv_upsampling = tf.image.resize_images(
conv5_3_conv, [28, 28],
method=tf.image.ResizeMethod.BILINEAR,
align_corners=False)
concatTensor = tf.concat([pool2_max_conv, pool3_conv], 0)
concatTensor = tf.concat([concatTensor, conv4_3_conv], 0)
concatTensor = tf.concat([concatTensor, conv5_3_conv_upsampling], 0)
concatTensor_conv = L.conv(concatTensor,
name="concatTensor_conv",
kh=3,
kw=3,
n_out=128)
hidden_layer_size = 30
input_size = 8
target_size = 10
rnn = LSTM_cell(input_size, hidden_layer_size, target_size)
# scope = "conv_inputs"
# logger.info("Building %s" % scope)
# Getting all outputs from rnn
outputs = rnn.get_outputs()
# Getting final output through indexing after reversing
last_output = outputs[-1]
# main reccurent layers
# l_hid = concatTensor_conv
# for idx in range(conf.recurrent_length):
# scope = 'LSTM%d' % idx
# l[scope] = l_hid = diagonal_bilstm(l_hid, conf, scope=scope)
# cell = tf.contrib.rnn.Conv2DLSTMCell(input_shape=[28, 28],
# # size of input feeding into network (needed for the zero state)
# kernel_shape=[3, 3], # for a 3 by 3 conv
# output_channels=128) # number of feature maps
##########################################################################
# Now running the dynamic rnn
##########################################################################
# (outputs, state) = tf.nn.dynamic_rnn(cell, concatTensor_conv, time_major=False, dtype=tf.float32)
##########################################################################
# Now treat the hidden state out of the conv lstm as the new image
##########################################################################
x_image = state[0]
# tf.contrib.rnn.ConvLSTMCell(conv_ndims,input_shape,output_channels, kernel_shape,use_bias=True,skip_connection=False,forget_bias=1.0,
# initializers=None, name='conv_lstm_cell')
# flatten
# flattened_shape = np.prod([s.value for s in net.get_shape()[1:]])
# net = tf.reshape(net, [-1, flattened_shape], name="flatten")
# fully connected
# net = L.fully_connected(net, name="fc6", n_out=4096)
# net = tf.nn.dropout(net, keep_prob)
# net = L.fully_connected(net, name="fc7", n_out=4096)
# net = tf.nn.dropout(net, keep_prob)
# net = L.fully_connected(net, name="fc8", n_out=n_classes)
return net
