def last_cnn(self, h, resolution, ch_in, ch_out):
with nn.parameter_scope("phase_{}".format(resolution)):
with nn.parameter_scope("conv1"):
h = conv(h,
ch_in,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
with_bias=not self.use_ln,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = LN(h, use_ln=self.use_ln)
h = self.activation(h)
with nn.parameter_scope("conv2"):
h = conv(h,
ch_out,
kernel=(4, 4),
pad=(0, 0),
stride=(1, 1),
with_bias=not self.use_ln,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = LN(h, use_ln=self.use_ln)
h = self.activation(h)
with nn.parameter_scope("linear"):
h = affine(h,
1,
with_bias=True,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
return h
def _inference(self, CC, MLO, keep_prob, is_train):
layers = [3, 16, 32, 64, 64]
cc = CC
mlo = MLO
for i in range(4):
with tf.variable_scope('CC_layers_%s' % i) as scope:
cc = F.conv(cc, layers[i])
cc = F.batch_norm(cc, is_train)
cc = F.activation(cc)
cc = F.max_pool(cc)
with tf.variable_scope('CC_features') as scope:
cc = F.dense(cc, layers[i + 1])
cc = F.batch_norm(cc, is_train)
cc = F.activation(cc)
for j in range(4):
with tf.variable_scope('MLO_layers_%s' % j) as scope:
mlo = F.conv(mlo, layers[j])
mlo = F.batch_norm(mlo, is_train)
mlo = F.activation(mlo)
mlo = F.max_pool(mlo)
with tf.variable_scope('MLO_features') as scope:
mlo = F.dense(mlo, layers[j + 1])
mlo = F.batch_norm(mlo, is_train)
mlo = F.activation(mlo)
with tf.variable_scope('softmax') as scope:
concat = tf.concat(1, [cc, mlo])
h = F.dense(concat, self._num_classes)
return h
def cnn(self, h, resolution, channel, test):
"""CNN block
The following operations are performed two times.
1. Upsampling
2. Conv
3. Pixel-wise normalization
4. Relu
"""
h = F.unpooling(h, kernel=(2, 2))
with nn.parameter_scope("phase_{}".format(resolution)):
with nn.parameter_scope("conv1"):
h = conv(h, channel, kernel=(3, 3), pad=(1, 1), stride=(1, 1),
with_bias=not self.use_bn,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = pixel_wise_feature_vector_normalization(
BN(h, use_bn=self.use_bn, test=test))
h = self.activation(h)
with nn.parameter_scope("conv2"):
h = conv(h, channel, kernel=(3, 3), pad=(1, 1), stride=(1, 1),
with_bias=not self.use_bn,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = pixel_wise_feature_vector_normalization(
BN(h, use_bn=self.use_bn, test=test))
h = self.activation(h)
return h
def cnn(self, h, resolution, ch_in, ch_out):
"""CNN block
The following operations are performed two times.
1. Conv
2. Layer normalization
3. Leaky relu
"""
with nn.parameter_scope("phase_{}".format(resolution)):
with nn.parameter_scope("conv1"):
h = conv(h,
ch_in,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
with_bias=not self.use_ln,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = LN(h, use_ln=self.use_ln)
h = self.activation(h)
with nn.parameter_scope("conv2"):
h = conv(h,
ch_out,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
with_bias=not self.use_ln,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = LN(h, use_ln=self.use_ln)
h = self.activation(h)
h = F.average_pooling(h, kernel=(2, 2))
return h
def _inference(self, X, keep_prob):
h = F.max_pool(F.activation(F.conv(X, 64)))
h = F.max_pool(F.activation(F.conv(h, 128)))
h = F.max_pool(F.activation(F.conv(h, 256)))
h = F.activation(F.dense(F.flatten(h), 1024))
h = F.dense(h, self._num_classes)
return tf.nn.softmax(h)
def _inference(self, X, keep_prob):
h = F.max_pool(F.activation(F.conv(X, 64)))
h = F.max_pool(F.activation(F.conv(h, 128)))
h = F.max_pool(F.activation(F.conv(h, 256)))
h = F.activation(F.dense(F.flatten(h), 1024))
h = F.dense(h, self._num_classes)
return h
def _residual(self, h, channels, strides, keep_prob, is_train):
h0 = h
h1 = F.conv(F.activation(F.batch_norm(self, 'bn1', h0, is_train)), channels, strides)
h1 = F.dropout(h1, keep_prob, is_train)
h2 = F.conv(F.activation(F.batch_norm(self, 'bn2', h1, is_train)), channels)
if F.volume(h0) == F.volume(h2):
h = h0 + h2
else :
h4 = F.conv(h0, channels, strides)
h = h2 + h4
return h
def _residual(self, h, channels, strides, keep_prob):
h0 = h
h1 = F.dropout(
F.conv(F.activation(F.batch_normalization(h0)), channels, strides),
keep_prob)
h2 = F.conv(F.activation(F.batch_normalization(h1)), channels)
# c.f. http://gitxiv.com/comments/7rffyqcPLirEEsmpX
if F.volume(h0) == F.volume(h2):
h = h2 + h0
else:
h4 = F.conv(h0, channels, strides)
h = h2 + h4
return h
def _residual(self, h, channels, strides, keep_prob, is_train):
h0 = h
with tf.variable_scope('residual_first'):
h1 = F.conv(F.activation(F.batch_norm(h0, is_train)), channels, strides)
h1 = F.dropout(h1, keep_prob, is_train)
with tf.variable_scope('residual_second'):
h2 = F.conv(F.activation(F.batch_norm(h1, is_train)), channels)
if F.volume(h0) == F.volume(h2):
h = h0 + h2
else :
h4 = F.conv(h0, channels, strides)
h = h2 + h4
return h
def _residual(self, h, channels, strides):
h0 = h
h1 = F.activation(
F.batch_normalization(
F.conv(h0, channels, strides, bias_term=False)))
h2 = F.batch_normalization(F.conv(h1, channels, bias_term=False))
if F.volume(h0) == F.volume(h2):
h = h2 + h0
else:
h3 = F.avg_pool(h0)
h4 = tf.pad(h3,
[[0, 0], [0, 0], [0, 0], [channels / 4, channels / 4]])
h = h2 + h4
return h
def _residual(self, h, channels, strides):
h0 = h
h1 = F.activation(
F.batch_normalization(
F.conv(h0, channels, strides, bias_term=False)))
h2 = F.batch_normalization(F.conv(h1, channels, bias_term=False))
# c.f. http://gitxiv.com/comments/7rffyqcPLirEEsmpX
if F.volume(h0) == F.volume(h2):
h = h2 + h0
else:
h3 = F.avg_pool(h0)
h4 = tf.pad(h3,
[[0, 0], [0, 0], [0, 0], [channels / 4, channels / 4]])
h = h2 + h4
return F.activation(h)
def first_cnn(self, h, resolution, channel, test=False):
with nn.parameter_scope("phase_{}".format(resolution)):
# affine is 1x1 conv with 4x4 kernel and 3x3 pad.
with nn.parameter_scope("conv1"):
h = affine(h,
channel * 4 * 4,
with_bias=not self.use_bn,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = BN(h, use_bn=self.use_bn, test=test)
h = F.reshape(h, (h.shape[0], channel, 4, 4))
h = pixel_wise_feature_vector_normalization(
BN(h, use_bn=self.use_bn, test=test))
h = self.activation(h)
with nn.parameter_scope("conv2"):
h = conv(h,
channel,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
with_bias=not self.use_bn,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = pixel_wise_feature_vector_normalization(
BN(h, use_bn=self.use_bn, test=test))
h = self.activation(h)
return h
def _inference(self, X, keep_prob, is_train):
dropout_rate = [0.9, 0.8, 0.7, 0.6, 0.5]
layers = [64, 128, 256, 512, 512]
iters = [2, 2, 3, 3]
h = X
# VGG Network Layer
for i in range(4):
for j in range(iters[i]):
with tf.variable_scope('layers%s_%s' % (i, j)) as scope:
h = F.conv(h, layers[i])
h = F.batch_norm(h, is_train)
h = F.activation(h)
h = F.dropout(h, dropout_rate[i], is_train)
h = F.max_pool(h)
# Fully Connected Layer
with tf.variable_scope('fully_connected_layer') as scope:
h = F.dense(h, layers[i + 1])
h = F.batch_norm(h, is_train)
h = F.activation(h)
h = F.dropout(h, dropout_rate[i + 1], is_train)
# Softmax Layer
with tf.variable_scope('softmax_layer') as scope:
h = F.dense(h, self._num_classes)
return h
def to_RGB(self, h, resolution):
"""To RGB layer
To RGB projects feature maps to RGB maps.
"""
with nn.parameter_scope("to_rgb_{}".format(resolution)):
h = conv(h, 3, kernel=(1, 1), pad=(0, 0), stride=(1, 1),
with_bias=True,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
return h
def from_RGB(self, h, resolution, channel):
"""From RGB layer
From RGB projects RGB maps to feature maps.
"""
with nn.parameter_scope("from_rgb_{}".format(resolution)):
h = conv(h, channel, kernel=(1, 1), pad=(0, 0), stride=(1, 1),
with_bias=True,
use_wscale=self.use_wscale,
use_he_backward=self.use_he_backward)
h = self.activation(h)
return h
def _inference(self, X, keep_prob):
h = X
h = F.activation(F.batch_normalization(F.conv(h, 16, bias_term=False)))
for i in range(self._layers):
h = self._residual(h, channels=16, strides=1)
for channels in [32, 64]:
for i in range(self._layers):
strides = 2 if i == 0 else 1
h = self._residual(h, channels, strides)
h = tf.reduce_mean(h, reduction_indices=[1,
2]) # Global Average Pooling
h = F.dense(h, self._num_classes)
return h
def _inference(self, X, keep_prob, is_train):
h = F.conv(X, 16)
for i in range(self._layers):
with tf.variable_scope(str(16*self._k)+'_layers_%s' %i):
h = self._residual(h, channels=16*self._k, strides=1, keep_prob=keep_prob, is_train=is_train)
for channels in [32*self._k, 64*self._k]:
for i in range(self._layers):
with tf.variable_scope(str(channels)+'_layers_%s' %i):
strides = 2 if i == 0 else 1
h = self._residual(h, channels, strides, keep_prob, is_train)
h = F.activation(F.batch_norm(self, 'bn', h, is_train))
h = tf.reduce_mean(h, reduction_indices=[1,2])
h = F.dense(h, self._num_classes)
return h
def _inference(self, X, keep_prob, is_train):
# Conv_layer 1
conv = F.conv(X, 192)
batch_norm = F._batch_norm(self, 'bn1', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.9, is_train)
conv = F.conv(dropout, 192)
batch_norm = F._batch_norm(self, 'bn2', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.9, is_train)
max_pool = F.max_pool(dropout) # 16 x 16
# Conv_layer 2
conv = F.conv(max_pool, 192)
batch_norm = F._batch_norm(self, 'bn3', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.8, is_train)
conv = F.conv(dropout, 192)
batch_norm = F._batch_norm(self, 'bn4', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.8, is_train)
max_pool = F.max_pool(dropout) # 8 x 8
# Conv_layer 3
conv = F.conv(max_pool, 256)
batch_norm = F._batch_norm(self, 'bn5', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.7, is_train)
conv = F.conv(dropout, 256)
batch_norm = F._batch_norm(self, 'bn6', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.7, is_train)
conv = F.conv(dropout, 256)
batch_norm = F._batch_norm(self, 'bn7', conv, is_train)
dropout = F.dropout(relu, 0.7, is_train)
max_pool = F.max_pool(dropout) # 4 x 4
# Conv_layer 4
conv = F.conv(max_pool, 512)
batch_norm = F._batch_norm(self, 'bn8', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.6, is_train)
conv = F.conv(dropout, 512)
batch_norm = F._batch_norm(self, 'bn9', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.6, is_train)
conv = F.conv(max_pool, 512)
batch_norm = F._batch_norm(self, 'bn10', conv, is_train)
relu = F.activation(batch_norm)
dropout = F.dropout(relu, 0.6, is_train)
max_pool = F.max_pool(dropout) # 2 x 2
# Fully Connected Layer
h = tf.reduce_mean(max_pool, reduction_indices=[1,2])
h = F.dropout(h, 0.5, is_train)
h = F.dense(h, 512)
h = F._batch_norm(self, 'bn11', h, is_train)
h = F.activation(h)
h = F.dropout(h, 0.5, is_train)
h = F.dense(h, self._num_classes)
return h
