def AttentionResNetCifar10(shape=(32, 32, 3), n_channels=32, n_classes=10):
"""
Attention-56 ResNet for Cifar10 Dataset
https://arxiv.org/abs/1704.06904
"""
input_ = Input(shape=shape)
x = Conv2D(n_channels, (5, 5), padding='same')(input_)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x) # 16x16
x = residual_block(x, input_channels=32, output_channels=128)
x = attention_block(x, encoder_depth=2)
x = residual_block(x, input_channels=128, output_channels=256, stride=2) # 8x8
x = attention_block(x, encoder_depth=1)
x = residual_block(x, input_channels=256, output_channels=512, stride=2) # 4x4
x = attention_block(x, encoder_depth=1)
x = residual_block(x, input_channels=512, output_channels=1024)
x = residual_block(x, input_channels=1024, output_channels=1024)
x = residual_block(x, input_channels=1024, output_channels=1024)
x = AveragePooling2D(pool_size=(4, 4), strides=(1, 1))(x) # 1x1
x = Flatten()(x)
output = Dense(n_classes, activation='softmax')(x)
model = Model(input_, output)
return model
def AttentionResNetCifar10(shape=(32, 32, 3), n_channels=32, n_classes=10):
input_ = Input(shape=shape)
x = Conv2D(n_channels, (5, 5), padding='same')(input_)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = MaxPool2D(pool_size=(2, 2))(x) # 16x16
x = residual_block(x, input_channels=32, output_channels=128)
x = attention_block(x, encoder_depth=2)
x = Conv2D(x.get_shape()[-1].value, (1, 1),name='residual_attention_stage1')(x)
x = residual_block(x, input_channels=128, output_channels=256, stride=2) # 8x8
x = attention_block(x, encoder_depth=1)
x = Conv2D(x.get_shape()[-1].value, (1, 1),name='residual_attention_stage2')(x)
x = residual_block(x, input_channels=256, output_channels=512, stride=2) # 4x4
x = attention_block(x, encoder_depth=1) #第三个Attention Module在论文中是不需要要skip_connections的，所以encoder_depth在此处应该为0,不过代码为了方便还是设为1
x = Conv2D(x.get_shape()[-1].value, (1, 1),name='residual_attention_stage3')(x)
x = residual_block(x, input_channels=512, output_channels=1024)
x = residual_block(x, input_channels=1024, output_channels=1024)
x = residual_block(x, input_channels=1024, output_channels=1024)
x = AveragePooling2D(pool_size=(4, 4), strides=(1, 1))(x) # 1x1
x = Flatten()(x)
output = Dense(n_classes, activation='softmax')(x)
model = Model(input_, output)
return model
def AttentionResNet18(shape=(224, 224, 3), n_channels=64, n_classes=5, dropout=0, regularization=0.01):
"""
Attention ResNet-18 with an attention-layer, variable imput size
"""
regularizer = l2(regularization)
input_ = Input(shape=shape)
x = Conv2D(n_channels, (7, 7), strides=(2, 2),
padding='same')(input_) # 112x112
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(3, 3), strides=( 2, 2), padding='same')(x) # 56x56
x = residual_block(x, output_channels=n_channels * 16, stride=4) # 14x14
# ATTENTION LAYER
x = attention_block(x, encoder_depth=1) # bottleneck 7x7
x = residual_block(x, output_channels=n_channels * 32, stride=2) # 7x7
pool_size = (x.get_shape()[1], x.get_shape()[2])
x = AveragePooling2D(pool_size=pool_size, strides=(1, 1))(x)
x = Flatten()(x)
if dropout:
x = Dropout(dropout)(x)
output = Dense(n_classes, kernel_regularizer=regularizer,
activation='softmax')(x)
model = Model(input_, output)
return model
def AttentionResNet56(shape=(224, 224, 3),
n_channels=64,
n_classes=10,
dropout=0,
regularization=0.01):
"""
Attention-56 ResNet
https://arxiv.org/abs/1704.06904
"""
regularizer = l2(regularization)
input_ = Input(shape=shape)
x = Conv2D(n_channels, (7, 7), strides=(2, 2),
padding='same')(input_) # 112x112
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2),
padding='same')(x) # 56x56
x = residual_block(x, output_channels=n_channels * 4) # 56x56
x = attention_block(x, encoder_depth=3) # bottleneck 7x7
x = residual_block(x, output_channels=n_channels * 8, stride=2) # 28x28
x = attention_block(x, encoder_depth=2) # bottleneck 7x7
x = residual_block(x, output_channels=n_channels * 16, stride=2) # 14x14
x = attention_block(x, encoder_depth=1) # bottleneck 7x7
x = residual_block(x, output_channels=n_channels * 32, stride=2) # 7x7
x = residual_block(x, output_channels=n_channels * 32)
x = residual_block(x, output_channels=n_channels * 32)
pool_size = (x.get_shape()[1].value, x.get_shape()[2].value)
x = AveragePooling2D(pool_size=pool_size, strides=(1, 1))(x)
x = Flatten()(x)
if dropout:
x = Dropout(dropout)(x)
output = Dense(n_classes,
kernel_regularizer=regularizer,
activation='softmax')(x)
model = Model(input_, output)
return model
def attention(inputs,
skip,
depth,
depth_bottleneck,
scale=2,
stride=1,
rate=1,
outputs_collections=None,
scope=None):
assert stride == 1
if args.resize == 'bilinear': #双线性插值
resize_method = tf.image.ResizeMethod.BILINEAR
if args.resize == 'nearest': #最近邻插值
resize_method = tf.image.ResizeMethod.NEAREST_NEIGHBOR
with tf.variable_scope(scope, 'bottleneck_skip', [inputs, skip]) as sc:
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
#res_inpt = tf.image.resize_images(inputs, tf.shape(skip)[1:3], method=resize_method)
skip = attention_block(skip)
res_inpt = slim.conv2d(inputs,
depth * scale * scale, [1, 1],
stride=1,
padding='SAME',
activation_fn=None,
scope=scope)
res_inpt = tf.depth_to_space(res_inpt, scale)
res_inpt = tf.image.resize_images(res_inpt,
tf.shape(skip)[1:3],
method=resize_method)
if depth != depth_in: #
shortcut = slim.conv2d(
res_inpt,
depth, [1, 1],
stride=stride,
activation_fn=None,
scope='shortcut') #第一次resskip时，2048 != 512,要把shortcut的深度变为512
else:
shortcut = res_inpt
# print("Live from skip bottleneck block! We got %s as input and %s as skip connection" % (inputs.get_shape(), skip.get_shape()))
concat = tf.concat([res_inpt, skip], 3)
residual = slim.conv2d(concat,
depth_bottleneck, [1, 1],
stride=stride,
scope='conv1')
residual = resnet_utils.conv2d_same(residual,
depth_bottleneck,
3,
1,
rate=rate,
scope='conv2')
residual = slim.conv2d(residual,
depth, [1, 1],
stride=1,
activation_fn=None,
scope='conv3')
output = tf.nn.relu(shortcut + residual)
# print("So far in the end of bottleneck skip we have %s" % (output.get_shape()))
return slim.utils.collect_named_outputs(outputs_collections, sc.name,
output)
