def build_model(self, img_input: TensorType) -> TensorType:
"""Build graph using img_input as input.
Args:
img_input: 4D Image input tensor of shape (batch, height, width, channels)
Returns:
`Tensor` holding output probabilities per class, shape (batch, num_classes)
"""
x = layers.conv(img_input,
filters_out=64,
kernel_size=7,
stride=2,
add_bias=False,
name='conv1_conv')
x = layers.norm(x, axis=-1, epsilon=1.001e-5, name='conv1_bn')
x = layers.relu(x, name='conv1_relu')
x = layers.zero_padding(x, padding=((1, 1), (1, 1)), name='pool1_pad')
x = layers.max_pool(x, kernel_size=3, name='pool1')
x = self.conv_block(x, 3, [64, 64, 256], stage=2, block='1', strides=1)
x = self.identity_block(x, 3, [64, 64, 256], stage=2, block='2')
x = self.identity_block(x, 3, [64, 64, 256], stage=2, block='3')
x = self.conv_block(x, 3, [128, 128, 512], stage=3, block='1')
x = self.identity_block(x, 3, [128, 128, 512], stage=3, block='2')
x = self.identity_block(x, 3, [128, 128, 512], stage=3, block='3')
x = self.identity_block(x, 3, [128, 128, 512], stage=3, block='4')
x = self.conv_block(x, 3, [256, 256, 1024], stage=4, block='1')
x = self.identity_block(x, 3, [256, 256, 1024], stage=4, block='2')
x = self.identity_block(x, 3, [256, 256, 1024], stage=4, block='3')
x = self.identity_block(x, 3, [256, 256, 1024], stage=4, block='4')
x = self.identity_block(x, 3, [256, 256, 1024], stage=4, block='5')
x = self.identity_block(x, 3, [256, 256, 1024], stage=4, block='6')
x = self.conv_block(x, 3, [512, 512, 2048], stage=5, block='1')
x = self.identity_block(x, 3, [512, 512, 2048], stage=5, block='2')
x = self.identity_block(x, 3, [512, 512, 2048], stage=5, block='3')
x = layers.avg_pool(x, kernel_size=7, strides=1, name='avg_pool')
x = layers.squeeze(x, axis=[1, 2], name='squeeze')
x = layers.fully_connected(x, self.num_classes, name='probs')
x = layers.softmax(x, name='output-prob')
return x
def build_model(self, img_input: TensorType) -> TensorType:
"""Build graph using img_input as input.
Args:
img_input: 4D Image input tensor of shape (batch, height, width,
channels)
Returns:
`Tensor` holding output probabilities per class, shape (batch,
num_classes)
"""
channel_axis = -1
x = conv_norm_relu(img_input,
64,
7,
strides=2,
padding='SAME',
name='InceptionV1/Conv2d_1a_7x7',
norm_suffix="/BatchNorm",
weight_suffix="weights",
conv_suffix="")
x = max_pool(x, 3, strides=2, padding='same', name='MaxPool_2a_3x3')
x = conv_norm_relu(x,
64,
1,
padding='same',
name='InceptionV1/Conv2d_2b_1x1',
weight_suffix="weights",
conv_suffix="",
norm_suffix="/BatchNorm")
x = conv_norm_relu(x,
192,
3,
padding='same',
name='InceptionV1/Conv2d_2c_3x3',
weight_suffix="weights",
conv_suffix="",
norm_suffix="/BatchNorm")
x = max_pool(x, 3, strides=2, padding='same', name='MaxPool_3a_3x3')
# Now the '3' level inception units
x = self.inception_block(x, ((64, ), (96, 128), (16, 32), (32, )),
channel_axis, 'InceptionV1/Mixed_3b')
x = self.inception_block(x, ((128, ), (128, 192), (32, 96), (64, )),
channel_axis, 'InceptionV1/Mixed_3c')
x = max_pool(x, 3, strides=2, padding='same', name='MaxPool_4a_3x3')
# Now the '4' level inception units
x = self.inception_block(x, ((192, ), (96, 208), (16, 48), (64, )),
channel_axis, 'InceptionV1/Mixed_4b')
x = self.inception_block(x, ((160, ), (112, 224), (24, 64), (64, )),
channel_axis, 'InceptionV1/Mixed_4c')
x = self.inception_block(x, ((128, ), (128, 256), (24, 64), (64, )),
channel_axis, 'InceptionV1/Mixed_4d')
x = self.inception_block(x, ((112, ), (144, 288), (32, 64), (64, )),
channel_axis, 'InceptionV1/Mixed_4e')
x = self.inception_block(x, ((256, ), (160, 320), (32, 128), (128, )),
channel_axis, 'InceptionV1/Mixed_4f')
x = max_pool(x, 2, strides=2, padding='same', name='MaxPool_5a_2x2')
# Now the '5' level inception units
x = self.inception_block(x, ((256, ), (160, 320), (32, 128), (128, )),
channel_axis, 'InceptionV1/Mixed_5b')
x = self.inception_block(x, ((384, ), (192, 384), (48, 128), (128, )),
channel_axis, 'InceptionV1/Mixed_5c')
# Classification block
x = avg_pool(x,
kernel_size=7,
strides=1,
name='avg_pool',
padding='valid')
x = conv(x,
filters_out=self.num_classes + 1,
kernel_size=1,
padding='valid',
add_bias=True,
name='InceptionV1/Logits/Conv2d_0c_1x1',
weight_suffix="weights",
bias_suffix="biases")
x = squeeze(x, axis=[1, 2], name='squeeze')
x = softmax(x, name='output-prob')
return x
def build_model(self, img_input: TensorType) -> TensorType:
"""Build graph using img_input as input.
Args:
img_input: 4D Image input tensor of shape (batch, height, width, channels)
Returns:
`Tensor` holding output probabilities per class, shape (batch, num_classes)
"""
x = conv_norm_relu(img_input, 32, 3, strides=2, padding='VALID')
x = conv_norm_relu(x, 32, 3, padding='VALID')
x = conv_norm_relu(
x,
64,
3,
)
x = max_pool(x, 3, strides=2)
x = conv_norm_relu(x, 80, 1, padding='VALID')
x = conv_norm_relu(x, 192, 3, padding='VALID')
x = max_pool(x, 3, strides=2)
# mixed 0: 35 x 35 x 256
branch1x1 = conv_norm_relu(x, 64, 1)
branch5x5 = conv_norm_relu(x, 48, 1)
branch5x5 = conv_norm_relu(branch5x5, 64, 5)
branch3x3dbl = conv_norm_relu(x, 64, 1)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 32, 1)
x = concat([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=-1,
name='mixed0')
# mixed 1: 35 x 35 x 288
branch1x1 = conv_norm_relu(x, 64, 1)
branch5x5 = conv_norm_relu(x, 48, 1)
branch5x5 = conv_norm_relu(branch5x5, 64, 5)
branch3x3dbl = conv_norm_relu(x, 64, 1)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 64, 1)
x = concat([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=-1,
name='mixed1')
# mixed 2: 35 x 35 x 288
branch1x1 = conv_norm_relu(x, 64, 1)
branch5x5 = conv_norm_relu(x, 48, 1)
branch5x5 = conv_norm_relu(branch5x5, 64, 5)
branch3x3dbl = conv_norm_relu(x, 64, 1)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 64, 1)
x = concat([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=-1,
name='mixed2')
# mixed 3: 17 x 17 x 768
branch3x3 = conv_norm_relu(x, 384, 3, strides=2, padding='VALID')
branch3x3dbl = conv_norm_relu(x, 64, 1)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 96, 3)
branch3x3dbl = conv_norm_relu(branch3x3dbl,
96,
3,
strides=2,
padding='VALID')
branch_pool = max_pool(x, 3, 2)
x = concat([branch3x3, branch3x3dbl, branch_pool],
axis=-1,
name='mixed3')
# mixed 4: 17 x 17 x 768
branch1x1 = conv_norm_relu(x, 192, 1)
branch7x7 = conv_norm_relu(x, 128, 1)
branch7x7 = conv_norm_relu(branch7x7, 128, 1, 7)
branch7x7 = conv_norm_relu(branch7x7, 192, 7, 1)
branch7x7dbl = conv_norm_relu(x, 128, 1, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 128, 1, 7)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 1, 7)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 192, 1, 1)
x = concat([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=-1,
name='mixed4')
# mixed 5, 6: 17 x 17 x 768
for i in range(2):
branch1x1 = conv_norm_relu(x, 192, 1, 1)
branch7x7 = conv_norm_relu(x, 160, 1, 1)
branch7x7 = conv_norm_relu(branch7x7, 160, 1, 7)
branch7x7 = conv_norm_relu(branch7x7, 192, 7, 1)
branch7x7dbl = conv_norm_relu(x, 160, 1, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 160, 1, 7)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 1, 7)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 192, 1, 1)
x = concat([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=-1,
name='mixed' + str(5 + i))
# mixed 7: 17 x 17 x 768
branch1x1 = conv_norm_relu(x, 192, 1, 1)
branch7x7 = conv_norm_relu(x, 192, 1, 1)
branch7x7 = conv_norm_relu(branch7x7, 192, 1, 7)
branch7x7 = conv_norm_relu(branch7x7, 192, 7, 1)
branch7x7dbl = conv_norm_relu(x, 192, 1, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 1, 7)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv_norm_relu(branch7x7dbl, 192, 1, 7)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 192, 1, 1)
x = concat([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=-1,
name='mixed7')
# mixed 8: 8 x 8 x 1280
branch3x3 = conv_norm_relu(x, 192, 1)
branch3x3 = conv_norm_relu(branch3x3,
320,
3,
strides=2,
padding='VALID')
branch7x7x3 = conv_norm_relu(x, 192, 1, 1)
branch7x7x3 = conv_norm_relu(branch7x7x3, 192, 1, 7)
branch7x7x3 = conv_norm_relu(branch7x7x3, 192, 7, 1)
branch7x7x3 = conv_norm_relu(branch7x7x3,
192,
3,
3,
strides=2,
padding='VALID')
branch_pool = max_pool(x, 3, strides=2)
x = concat([branch3x3, branch7x7x3, branch_pool],
axis=-1,
name='mixed8')
# mixed 9: 8 x 8 x 2048
for i in range(2):
branch1x1 = conv_norm_relu(x, 320, 1, 1)
branch3x3 = conv_norm_relu(x, 384, 1, 1)
branch3x3_1 = conv_norm_relu(branch3x3, 384, 1, 3)
branch3x3_2 = conv_norm_relu(branch3x3, 384, 3, 1)
branch3x3 = concat([branch3x3_1, branch3x3_2],
axis=-1,
name='mixed9_' + str(i))
branch3x3dbl = conv_norm_relu(x, 448, 1, 1)
branch3x3dbl = conv_norm_relu(branch3x3dbl, 384, 3, 3)
branch3x3dbl_1 = conv_norm_relu(branch3x3dbl, 384, 1, 3)
branch3x3dbl_2 = conv_norm_relu(branch3x3dbl, 384, 3, 1)
branch3x3dbl = concat([branch3x3dbl_1, branch3x3dbl_2], axis=-1)
branch_pool = avg_pool(x, 3, strides=1, padding='SAME')
branch_pool = conv_norm_relu(branch_pool, 192, 1, 1)
x = concat([branch1x1, branch3x3, branch3x3dbl, branch_pool],
axis=-1,
name='mixed' + str(9 + i))
# Classification block
x = avg_pool(x, kernel_size=8, strides=1, name='avg_pool')
x = squeeze(x, axis=[1, 2], name='squeeze')
x = fully_connected(x, self.num_classes, name='predictions')
x = softmax(x, name='output-prob')
return x
def build_model(self, img_input: TensorType) -> TensorType:
"""Build graph using img_input as input.
Args:
img_input: 4D Image input tensor of shape (batch, height, width, channels)
Returns:
`Tensor` holding output probabilities per class, shape (batch, num_classes)
"""
x = layers.conv(img_input, filters_out=32, kernel_size=3, stride=2, add_bias=False, name='block1_conv1')
x = layers.norm(x, name='block1_conv1_bn')
x = layers.relu(x, name='block1_conv1_act')
x = layers.conv(x, filters_out=64, kernel_size=3, add_bias=False, name='block1_conv2')
x = layers.norm(x, name='block1_conv2_bn')
x = layers.relu(x, name='block1_conv2_act')
residual = layers.conv(x, filters_out=128, kernel_size=1, stride=2, padding='same', add_bias=False)
residual = layers.norm(residual, name="batch_normalization")
x = layers.separable_conv(x, filters_out=128, kernel_size=3, padding='same', add_bias=False,
name='block2_sepconv1')
x = layers.norm(x, name='block2_sepconv1_bn')
x = layers.relu(x, name='block2_sepconv2_act')
x = layers.separable_conv(x, filters_out=128, kernel_size=3, padding='same', add_bias=False,
name='block2_sepconv2')
x = layers.norm(x, name='block2_sepconv2_bn')
x = layers.max_pool(x, 3, strides=2, padding='same', name='block2_pool')
x += residual
residual = layers.conv(x, filters_out=256, kernel_size=1, stride=2, padding='same', add_bias=False)
residual = layers.norm(residual, name="batch_normalization")
x = layers.relu(x, name='block3_sepconv1_act')
x = layers.separable_conv(x, filters_out=256, kernel_size=3,
padding='same',
add_bias=False,
name='block3_sepconv1')
x = layers.norm(x, name='block3_sepconv1_bn')
x = layers.relu(x, name='block3_sepconv2_act')
x = layers.separable_conv(x, filters_out=256, kernel_size=3,
padding='same',
add_bias=False,
name='block3_sepconv2')
x = layers.norm(x, name='block3_sepconv2_bn')
x = layers.max_pool(x, 3, strides=2,
padding='same',
name='block3_pool')
x += residual
residual = layers.conv(x, filters_out=728, kernel_size=1,
stride=2,
padding='same',
add_bias=False)
residual = layers.norm(residual, name="batch_normalization")
x = layers.relu(x, name='block4_sepconv1_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name='block4_sepconv1')
x = layers.norm(x, name='block4_sepconv1_bn')
x = layers.relu(x, name='block4_sepconv2_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name='block4_sepconv2')
x = layers.norm(x, name='block4_sepconv2_bn')
x = layers.max_pool(x, 3, strides=2,
padding='same',
name='block4_pool')
x += residual
for i in range(8):
residual = x
prefix = 'block' + str(i + 5)
x = layers.relu(x, name=prefix + '_sepconv1_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name=prefix + '_sepconv1')
x = layers.norm(x, name=prefix + '_sepconv1_bn')
x = layers.relu(x, name=prefix + '_sepconv2_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name=prefix + '_sepconv2')
x = layers.norm(x, name=prefix + '_sepconv2_bn')
x = layers.relu(x, name=prefix + '_sepconv3_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name=prefix + '_sepconv3')
x = layers.norm(x, name=prefix + '_sepconv3_bn')
x += residual
residual = layers.conv(x, filters_out=1024, kernel_size=1, stride=2,
padding='same', add_bias=False)
residual = layers.norm(residual, name="batch_normalization")
x = layers.relu(x, name='block13_sepconv1_act')
x = layers.separable_conv(x, filters_out=728, kernel_size=3,
padding='same',
add_bias=False,
name='block13_sepconv1')
x = layers.norm(x, name='block13_sepconv1_bn')
x = layers.relu(x, name='block13_sepconv2_act')
x = layers.separable_conv(x, filters_out=1024, kernel_size=3,
padding='same',
add_bias=False,
name='block13_sepconv2')
x = layers.norm(x, name='block13_sepconv2_bn')
x = layers.max_pool(x, 3,
strides=2,
padding='same',
name='block13_pool')
x += residual
x = layers.separable_conv(x, filters_out=1536, kernel_size=3,
padding='same',
add_bias=False,
name='block14_sepconv1')
x = layers.norm(x, name='block14_sepconv1_bn')
x = layers.relu(x, name='block14_sepconv1_act')
x = layers.separable_conv(x, filters_out=2048, kernel_size=3,
padding='same',
add_bias=False,
name='block14_sepconv2')
x = layers.norm(x, name='block14_sepconv2_bn')
x = layers.relu(x, name='block14_sepconv2_act')
# Classification block
x = layers.avg_pool(x, kernel_size=10, strides=1, name='avg_pool')
x = layers.squeeze(x, axis=[1, 2], name='squeeze')
x = layers.fully_connected(x, self.num_classes, name='predictions')
x = layers.softmax(x, name='output-prob')
return x
def build_model(self, img_input: TensorType) -> TensorType:
"""Build graph using img_input as input.
Args:
img_input: 4D Image input tensor of shape (batch, height, width, channels)
Returns:
`Tensor` holding output probabilities per class, shape (batch, num_classes)
"""
filters = self.penultimate_filters // 24
x = layers.conv(img_input,
filters_out=self.stem_block_filters,
kernel_size=3,
stride=2,
padding='valid',
add_bias=False,
name='stem_conv1')
x = layers.norm(x,
axis=-1,
momentum=0.9997,
epsilon=1e-3,
name='stem_bn1')
p = None
x, p = self.reduction_a_cell(x,
p,
filters // (self.filter_multiplier**2),
block_id='stem_1')
x, p = self.reduction_a_cell(x,
p,
filters // self.filter_multiplier,
block_id='stem_2')
for i in range(self.num_blocks):
x, p = self.normal_a_cell(x, p, filters, block_id='%d' % i)
x, p0 = self.reduction_a_cell(x,
p,
filters * self.filter_multiplier,
block_id='reduce_%d' % self.num_blocks)
p = p0 if not self.skip_reduction else p
for i in range(self.num_blocks):
x, p = self.normal_a_cell(x,
p,
filters * self.filter_multiplier,
block_id='%d' %
(self.num_blocks + i + 1))
x, p0 = self.reduction_a_cell(x,
p,
filters * self.filter_multiplier**2,
block_id='reduce_%d' %
(2 * self.num_blocks))
p = p0 if not self.skip_reduction else p
for i in range(self.num_blocks):
x, p = self.normal_a_cell(x,
p,
filters * self.filter_multiplier**2,
block_id='%d' %
(2 * self.num_blocks + i + 1))
x = layers.relu(x, 'relu')
# Classification block
x = layers.avg_pool(x, kernel_size=7, strides=1, name='avg_pool')
x = layers.squeeze(x, axis=[1, 2], name='squeeze')
x = layers.fully_connected(x, self.num_classes, name='predictions')
x = layers.softmax(x, name='output-prob')
return x
def reduction_a_cell(ip, p, filters, block_id=None):
"""Adds a Reduction cell for NASNet-A (Fig. 4 in the paper).
Args:
ip: Input tensor `x`
p: Input tensor `p`
filters: Number of output filters
block_id: String block_id
Returns:
A tf tensor
"""
channel_dim = -1
with tf.name_scope('reduction_A_block_%s' % block_id):
p = NASNetMobile.adjust_block(p, ip, filters, block_id)
h = layers.relu(ip)
h = layers.conv(h,
filters_out=filters,
kernel_size=(1, 1),
stride=1,
padding='same',
name='reduction_conv_1_%s' % block_id,
add_bias=False)
h = layers.norm(h,
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='reduction_bn_1_%s' % block_id)
h3 = layers.zero_padding(h,
padding=NASNetMobile.correct_pad(
h, (3, 3)),
name='reduction_pad_1_%s' % block_id)
with tf.name_scope('block_1'):
x1_1 = NASNetMobile.separable_conv_block(
h,
filters=filters,
kernel_size=5,
strides=2,
block_id='reduction_left1_%s' % block_id)
x1_2 = NASNetMobile.separable_conv_block(
p,
filters=filters,
kernel_size=7,
strides=2,
block_id='reduction_right1_%s' % block_id)
x1 = x1_1 + x1_2
with tf.name_scope('block_2'):
x2_1 = layers.max_pool(h3,
3,
strides=2,
padding='valid',
name='reduction_left2_%s' % block_id)
x2_2 = NASNetMobile.separable_conv_block(
p,
filters=filters,
kernel_size=7,
strides=2,
block_id='reduction_right2_%s' % block_id)
x2 = x2_1 + x2_2
with tf.name_scope('block_3'):
x3_1 = layers.avg_pool(h3,
3,
strides=2,
padding='valid',
name='reduction_left3_%s' % block_id)
x3_2 = NASNetMobile.separable_conv_block(
p,
filters,
5,
strides=2,
block_id='reduction_right3_%s' % block_id)
x3 = x3_1 + x3_2
with tf.name_scope('block_4'):
x4 = layers.avg_pool(x1,
3,
strides=1,
padding='same',
name='reduction_left4_%s' % block_id)
x4 += x2
with tf.name_scope('block_5'):
x5_1 = NASNetMobile.separable_conv_block(
x1, filters, 3, block_id='reduction_left4_%s' % block_id)
x5_2 = layers.max_pool(h3,
3,
strides=2,
padding='valid',
name='reduction_right5_%s' % block_id)
x5 = x5_1 + x5_2
x = layers.concat([x2, x3, x4, x5],
axis=channel_dim,
name='reduction_concat_%s' % block_id)
return x, ip
def normal_a_cell(ip, p, filters, block_id=None):
"""Adds a Normal cell for NASNet-A (Fig. 4 in the paper).
Args:
ip: Input tensor `x`
p: Input tensor `p`
filters: Number of output filters
block_id: String block_id
Returns:
A tensorflow tensor
"""
channel_dim = -1
with tf.name_scope('normal_A_block_%s' % block_id):
p = NASNetMobile.adjust_block(p, ip, filters, block_id)
h = layers.relu(ip)
h = layers.conv(h,
filters_out=filters,
kernel_size=(1, 1),
stride=1,
padding='same',
name='normal_conv_1_%s' % block_id,
add_bias=False)
h = layers.norm(h,
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='normal_bn_1_%s' % block_id)
with tf.name_scope('block_1'):
x1_1 = NASNetMobile.separable_conv_block(
h,
filters,
kernel_size=5,
block_id='normal_left1_%s' % block_id)
x1_2 = NASNetMobile.separable_conv_block(
p, filters, block_id='normal_right1_%s' % block_id)
x1 = x1_1 + x1_2
with tf.name_scope('block_2'):
x2_1 = NASNetMobile.separable_conv_block(
p, filters, 5, block_id='normal_left2_%s' % block_id)
x2_2 = NASNetMobile.separable_conv_block(
p, filters, 3, block_id='normal_right2_%s' % block_id)
x2 = x2_1 + x2_2
with tf.name_scope('block_3'):
x3 = layers.avg_pool(h,
3,
strides=1,
padding='same',
name='normal_left3_%s' % block_id)
x3 = x3 + p
with tf.name_scope('block_4'):
x4_1 = layers.avg_pool(p,
3,
strides=1,
padding='same',
name='normal_left4_%s' % block_id)
x4_2 = layers.avg_pool(p,
3,
strides=1,
padding='same',
name='normal_right4_%s' % block_id)
x4 = x4_1 + x4_2
with tf.name_scope('block_5'):
x5 = NASNetMobile.separable_conv_block(
h, filters, block_id='normal_left5_%s' % block_id)
x5 = x5 + h
x = layers.concat([p, x1, x2, x3, x4, x5],
axis=channel_dim,
name='normal_concat_%s' % block_id)
return x, ip
def adjust_block(p, ip, filters, block_id=None):
"""Adjusts the input `previous path` to match the shape of the `input`.
Used in situations where the output number of filters needs to be changed.
Args:
p: Input tensor which needs to be modified
ip: Input tensor whose shape needs to be matched
filters: Number of output filters to be matched
block_id: String block_id
Returns:
Adjusted tf tensor.
"""
channel_dim = -1
img_dim = -2
ip_shape = ip.get_shape().as_list()
if p is not None:
p_shape = p.get_shape().as_list()
else:
p_shape = ip_shape
with tf.name_scope('adjust_block'):
if p is None:
p = ip
elif p_shape[img_dim] != ip_shape[img_dim]:
with tf.name_scope('adjust_reduction_block_%s' % block_id):
p = layers.relu(p, name='adjust_relu_1_%s' % block_id)
p1 = layers.avg_pool(p,
1,
strides=2,
padding='valid',
name='adjust_avg_pool_1_%s' %
block_id)
p1 = layers.conv(p1,
filters_out=filters // 2,
kernel_size=(1, 1),
padding='same',
add_bias=False,
name='adjust_conv_1_%s' % block_id)
p2 = layers.zero_padding(p, padding=((0, 1), (0, 1)))
p2 = layers.crop(p2, cropping=((1, 0), (1, 0)))
p2 = layers.avg_pool(p2,
1,
strides=2,
padding='valid',
name='adjust_avg_pool_2_%s' %
block_id)
p2 = layers.conv(p2,
filters_out=filters // 2,
kernel_size=(1, 1),
padding='same',
add_bias=False,
name='adjust_conv_2_%s' % block_id)
p = layers.concat([p1, p2], axis=channel_dim)
p = layers.norm(p,
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)
elif p_shape[channel_dim] != filters:
with tf.name_scope('adjust_projection_block_%s' % block_id):
p = layers.relu(p)
p = layers.conv(p,
filters_out=filters,
kernel_size=(1, 1),
stride=1,
padding='same',
name='adjust_conv_projection_%s' %
block_id,
add_bias=False)
p = layers.norm(p,
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='adjust_bn_%s' % block_id)
return p
