from __future__ import absolute_import
import sugartensor as tf
__author__ = 'namju.kim@kakaocorp.com'
#
# VGG nets
#
@tf.sg_sugar_func
[docs]def sg_vgg_16(tensor, opt):
r"""Applies vgg 16 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
tensor: A `Tensor`
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
bn: True or False(default). If True, batch normal will be applied
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
opt += tf.sg_opt(num_class=1000, conv_only=False, squeeze=True, act='relu', bn=False)
# convolution layers
with tf.sg_context(name=opt.name, act=opt.act, bn=opt.bn, reuse=opt.reuse):
conv = (tensor
.sg_conv(dim=64, name='conv1_1')
.sg_conv(name='conv1_2')
.sg_pool()
.sg_conv(dim=128, name='conv2_1')
.sg_conv(name='conv2_2')
.sg_pool()
.sg_conv(dim=256, name='conv3_1')
.sg_conv(name='conv3_2')
.sg_conv(name='conv3_3')
.sg_pool()
.sg_conv(dim=512, name='conv4_1')
.sg_conv(name='conv4_2')
.sg_conv(name='conv4_3')
.sg_pool()
.sg_conv(dim=512, name='conv5_1')
.sg_conv(name='conv5_2')
.sg_conv(name='conv5_3')
.sg_pool())
# fully convolution layers
with tf.sg_context(name=opt.name, act=opt.act, pad='VALID', bn=opt.bn, reuse=opt.reuse):
fc = (conv
.sg_conv(dim=4096, size=7, dout=opt.dout, name='fc1')
.sg_conv(dim=4096, size=1, dout=opt.dout, name='fc2')
.sg_conv(dim=opt.num_class, size=1, act='linear', name='fc3'))
if opt.conv_only:
return conv
else:
if opt.squeeze:
return fc.sg_squeeze(axis=(1, 2))
else:
return fc
@tf.sg_sugar_func
[docs]def sg_vgg_19(tensor, opt):
r"""Applies vgg 19 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
tensor: A `Tensor`.
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
bn: True or False(default). If True, batch normal will be applied
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
opt += tf.sg_opt(num_class=1000, conv_only=False, squeeze=True, act='relu')
# convolution layers
with tf.sg_context(name=opt.name, act=opt.act, bn=opt.bn, reuse=opt.reuse):
conv = (tensor
.sg_conv(dim=64, name='conv1_1')
.sg_conv(name='conv1_2')
.sg_pool()
.sg_conv(dim=128, name='conv2_1')
.sg_conv(name='conv2_2')
.sg_pool()
.sg_conv(dim=256, name='conv3_1')
.sg_conv(name='conv3_2')
.sg_conv(name='conv3_3')
.sg_conv(name='conv3_4')
.sg_pool()
.sg_conv(dim=512, name='conv4_1')
.sg_conv(name='conv4_2')
.sg_conv(name='conv4_3')
.sg_conv(name='conv4_4')
.sg_pool()
.sg_conv(dim=512, name='conv5_1')
.sg_conv(name='conv5_2')
.sg_conv(name='conv5_3')
.sg_conv(name='conv5_4')
.sg_pool())
# fully convolution layers
with tf.sg_context(name=opt.name, act=opt.act, pad='VALID', bn=opt.bn, reuse=opt.reuse):
fc = (conv
.sg_conv(dim=4096, size=7, dout=opt.dout, name='fc1')
.sg_conv(dim=4096, size=1, dout=opt.dout, name='fc2')
.sg_conv(dim=opt.num_class, size=1, act='linear', name='fc3'))
if opt.conv_only:
return conv
else:
if opt.squeeze:
return fc.sg_squeeze(axis=(1, 2))
else:
return fc
#
# Residual nets ( version 2 )
#
@tf.sg_sugar_func
[docs]def sg_resnet_50(x, opt):
r"""Applies residual net 50 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _resnet_graph(x, opt, [3, 4, 6, 3])
@tf.sg_sugar_func
[docs]def sg_resnet_101(x, opt):
r"""Applies residual net 101 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _resnet_graph(x, opt, [3, 4, 23, 3])
@tf.sg_sugar_func
[docs]def sg_resnet_152(x, opt):
r"""Applies residual net 152 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _resnet_graph(x, opt, [3, 8, 36, 3])
@tf.sg_sugar_func
[docs]def sg_resnet_200(x, opt):
r"""Applies residual net 200 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _resnet_graph(x, opt, [3, 24, 36, 3])
@tf.sg_sugar_func
[docs]def sg_densenet_121(x, opt):
r"""Applies dense net 121 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
k: An integer. The Growth rate of densenet. Default is 32.
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _densenet_graph(x, opt + tf.sg_opt(k=32), [6, 12, 24, 16])
@tf.sg_sugar_func
[docs]def sg_densenet_169(x, opt):
r"""Applies dense net 169 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
k: An integer. The Growth rate of densenet. Default is 32.
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _densenet_graph(x, opt + tf.sg_opt(k=32), [6, 12, 32, 32])
@tf.sg_sugar_func
[docs]def sg_densenet_201(x, opt):
r"""Applies dense net 201 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
k: An integer. The Growth rate of densenet. Default is 32.
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _densenet_graph(x, opt + tf.sg_opt(k=32), [6, 12, 48, 32])
@tf.sg_sugar_func
[docs]def sg_densenet_161(x, opt):
r"""Applies dense net 161 model.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
k: An integer. The Growth rate of densenet. Default is 48.
num_class: An integer. Number of class. Default is 1000.
conv_only: Boolean. If True, fc layers are not applied. Default is False.
squeeze: Boolean. If True (default), the dimensions with size 1 in the final outputs will be removed.
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String(Optional). If provided, used as the scope name of this network
Returns:
A `Tensor`.
"""
return _densenet_graph(x, opt + tf.sg_opt(k=48), [6, 12, 36, 24])
#
# Residual net utility
#
@tf.sg_sugar_func
[docs]def sg_resnet_layer(x, opt):
r"""Applies basic architecture of residual net.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
dim: An integer. Dimension for this resnet layer
num: Number of times to repeat
act: String. 'relu' (default). the activation function name
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String. (optional) Used as convolution layer prefix
Returns:
A `Tensor`.
"""
assert opt.dim is not None, 'dim is mandatory.'
assert opt.num is not None, 'num is mandatory.'
# default stride
opt += tf.sg_opt(stride=1, act='relu')
# format convolutional layer name
def cname(index):
return opt.name if opt.name is None else opt.name + '_%d' % index
with tf.sg_context(bias=False, reuse=opt.reuse):
# 1st block
out = (x
.sg_bypass(act=opt.act, bn=True, name=cname(0))
.sg_conv(dim=opt.dim, size=1, act=opt.act, stride=opt.stride, bn=True, name=cname(1))
.sg_conv(dim=opt.dim, size=3, act=opt.act, bn=True, name=cname(2))
.sg_conv(dim=opt.dim*4, size=1, name=cname(3)))
out += x.sg_conv(dim=opt.dim * 4, size=1, stride=opt.stride, name=cname(4))
# middle blocks
for i in range(1, opt.num):
out_new = (out
.sg_bypass(act=opt.act, name=cname(4 * i + 1))
.sg_conv(dim=opt.dim, size=1, act=opt.act, bn=True, name=cname(4 * i + 2))
.sg_conv(dim=opt.dim, size=3, act=opt.act, bn=True, name=cname(4 * i + 3))
.sg_conv(dim=opt.dim*4, size=1, name=cname(4 * i + 4)))
out += out_new
return out
# construct residual network graphs
def _resnet_graph(x, opt, nums):
# default option
opt += tf.sg_opt(num_class=1000, conv_only=False, squeeze=True, act='relu')
# convolution layers ( residual net v2 arch )
with tf.sg_context(name=opt.name):
conv = (x
.sg_conv(dim=64, size=7, stride=2, bias=False, reuse=opt.reuse, name='conv1')
.sg_pool(size=3, stride=2, pad='SAME')
.sg_resnet_layer(dim=64, num=nums[0], stride=1, act=opt.act, reuse=opt.reuse, name='conv2')
.sg_resnet_layer(dim=128, num=nums[1], stride=2, act=opt.act, reuse=opt.reuse, name='conv3')
.sg_resnet_layer(dim=256, num=nums[2], stride=2, act=opt.act, reuse=opt.reuse, name='conv4')
.sg_resnet_layer(dim=512, num=nums[3], stride=2, act=opt.act, reuse=opt.reuse, name='conv5')
.sg_bypass(act=opt.act, bn=True, reuse=opt.reuse, name='final_act') # final activation
.sg_pool(size=7, stride=1, avg=True)) # global average pool
# fully convolution layers
fc = conv.sg_conv(dim=opt.num_class, size=1, act='linear', bn=False, reuse=opt.reuse, name='fc')
# return selectively
if opt.conv_only:
return conv
else:
if opt.squeeze:
return fc.sg_squeeze(axis=(1, 2))
else:
return fc
#
# Dense net utility
#
@tf.sg_sugar_func
[docs]def sg_densenet_layer(x, opt):
r"""Applies basic architecture of densenet layer.
Note that the fc layers in the original architecture
will be replaced with fully convolutional layers.
For convenience, We still call them fc layers, though.
Args:
x: A `Tensor`.
opt:
dim: An integer. Dimension for this resnet layer
num: Number of times to repeat
act: String. 'relu' (default). the activation function name
trans: Boolean. If True(default), transition layer will be applied.
reuse: Boolean(Optional). If True, all variables will be loaded from previous network.
name: String. (optional) Used as convolution layer prefix
Returns:
A `Tensor`.
"""
assert opt.dim is not None, 'dim is mandatory.'
assert opt.num is not None, 'num is mandatory.'
# default stride
opt += tf.sg_opt(stride=1, act='relu', trans=True)
# format convolutional layer name
def cname(index):
return opt.name if opt.name is None else opt.name + '_%d' % index
# dense layer
with tf.sg_context(bias=False, reuse=opt.reuse):
out = x
for i in range(opt.num):
# dense block
out_new = (out
.sg_bypass(act=opt.act, bn=True, name=cname(3 * i + 1))
.sg_conv(dim=opt.dim // 4, size=1, act=opt.act, bn=True, name=cname(3 * i + 2))
.sg_conv(dim=opt.dim, size=3, name=cname(3 * i + 3)))
out = tf.concat([out_new, out], 3)
# transition layer
if opt.trans:
out = (out
.sg_bypass(act=opt.act, bn=True, name=cname(3 * i + 4))
.sg_conv(size=1, name=cname(3 * i + 5))
.sg_pool(avg=True))
return out
# construct dense network graphs
def _densenet_graph(x, opt, nums):
# default option
opt += tf.sg_opt(num_class=1000, k=32, conv_only=False, squeeze=True, act='relu')
# convolution layers ( dense net arch )
with tf.sg_context(name=opt.name):
conv = (x
.sg_conv(dim=opt.k, size=7, stride=2, bias=False, reuse=opt.reuse, name='conv1')
.sg_pool(size=3, stride=2, pad='SAME')
.sg_densenet_layer(dim=opt.k, num=nums[0], act=opt.act, reuse=opt.reuse, name='conv2')
.sg_densenet_layer(dim=opt.k, num=nums[1], act=opt.act, reuse=opt.reuse, name='conv3')
.sg_densenet_layer(dim=opt.k, num=nums[2], act=opt.act, reuse=opt.reuse, name='conv4')
.sg_densenet_layer(dim=opt.k, num=nums[3], act=opt.act, trans=False, reuse=opt.reuse, name='conv5')
.sg_bypass(act=opt.act, bn=True, reuse=opt.reuse, name='final_act') # final activation
.sg_pool(size=7, stride=1, avg=True)) # global average pool
# fully convolution layers
fc = conv.sg_conv(dim=opt.num_class, size=1, act='linear', bn=False, reuse=opt.reuse, name='fc')
# return selectively
if opt.conv_only:
return conv
else:
if opt.squeeze:
return fc.sg_squeeze(axis=(1, 2))
else:
return fc