我正在尝试为VGG中的自动编码器实现分池。取消缓冲基本上是跟踪从编码器中提取最大池的历史，然后在解码器中应用该历史。如图we save the mask for pooling history for each maxpool所示。
这是我已经完成的简单实现。
from keras.layers import Input, Conv2D, UpSampling2D
import keras
from keras.layers import Lambda
import keras.backend as K
import tensorflow as tf
from keras.models import Model
from keras.layers import Conv2D, MaxPooling2D, GlobalMaxPooling2D, Input
from keras.layers import Input, Conv2D, UpSampling2D
from keras.utils.data_utils import get_file
import keras.backend as K
import numpy as np
import tensorflow as tf
from keras.utils import plot_model
def unpool(args):
mask, x = args
#print(mask.shape, x.shape)
return keras.layers.multiply([mask, x])
def mask_make(xt):
t = MaxPooling2D((2, 2), strides=(2, 2), padding = 'SAME')(xt)
t = UpSampling2D()(t)
bool_mask = K.greater_equal(xt, t)
mask = K.cast(bool_mask, dtype='float32')
#print (mask.shape)
mask_input = Input(tensor=mask) # Makes the mask to a Keras tensor to use as input
return mask_input
def vgg_layers(inputs, target_layer):
# Block 1
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv1')(inputs)
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x)
m1 = mask_make(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
# Block 2
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv1')(x)
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv2')(x)
m2 = mask_make(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
# Block 3
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv1')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv2')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv3')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv4')(x)
m3 = mask_make(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
# Block 4
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv1')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv2')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv3')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv4')(x)
m4 = mask_make(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
# Block 5
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv1')(x)
# Decoder
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='decoder_block5_conv1')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='decoder_block4_conv4')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='decoder_block4_conv3')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='decoder_block4_conv2')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='decoder_block4_conv1')(x)
x = UpSampling2D((2, 2), name='decoder_block4_upsample')(x)
x = Lambda(unpool)([m4, x])
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='decoder_block3_conv4')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='decoder_block3_conv3')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='decoder_block3_conv2')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='decoder_block3_conv1')(x)
x = UpSampling2D((2, 2), name='decoder_block3_upsample')(x)
x = Lambda(unpool)([m3, x])
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='decoder_block2_conv2')(x)
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='decoder_block2_conv1')(x)
x = UpSampling2D((2, 2), name='decoder_block2_upsample')(x)
x = Lambda(unpool)([m2, x])
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='decoder_block1_conv2')(x)
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='decoder_block1_conv1')(x)
x = UpSampling2D((2, 2), name='decoder_block1_upsample')(x)
x = Lambda(unpool)([m1, x])
return x, masks
def VGG19(input_tensor=None, input_shape=(32,32,3), target_layer=1):
"""
VGG19, up to the target layer (1 for relu1_1, 2 for relu2_1, etc.)
"""
if input_tensor is None:
inputs = Input(shape=input_shape)
else:
inputs = Input(tensor=input_tensor, shape=input_shape)
layers, masks = vgg_layers(inputs, target_layer)
model = Model(inputs, layers, name='vgg19')
plot_model(model, to_file='model.png')
#load_weights(model)
#return model, masks
VGG19()
但是，我收到“ ValueError：图形已断开连接：”
请帮助我解决这个问题。
答案 0 :(得分：1)
使用串联解决了该问题
def mask_make(x, orig):
t = UpSampling2D()(x)
_,a,b,c = orig.shape
#print (a,b,c)
xReshaped = Reshape((1,a*b*c))(t)
origReshaped = Reshape((1,a*b*c))(orig)
#print (xReshaped.shape)
#print (origReshaped.shape)
together = Concatenate(axis = -1)([origReshaped,xReshaped])
togReshaped = Reshape((2,a,b,c))(together)
#print (togReshaped.shape)
bool_mask = Lambda(lambda t: K.greater_equal(t[:,0], t[:,1]))(togReshaped)
mask = Lambda(lambda t: K.cast(t, dtype='float32'))(bool_mask)
#mask = Reshape((a,b,c))(mask)
#print (mask.shape)
return mask
def vgg_layers(inputs, target_layer):
masks = []
# Block 1
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv1')(inputs)
if target_layer == 1:
return x
x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x)
orig = x
x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
masks.append(mask_make(x, orig))
# Block 2
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv1')(x)
if target_layer == 2:
return x
x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv2')(x)
orig = x
x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
masks.append(mask_make(x, orig))
# Block 3
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv1')(x)
if target_layer == 3:
return x
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv2')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv3')(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv4')(x)
orig = x
x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
masks.append(mask_make(x, orig))
# Block 4
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv1')(x)
if target_layer == 4:
return x
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv2')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv3')(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv4')(x)
orig = x
x = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
masks.append(mask_make(x, orig))
# Block 5
x = Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv1')(x)
return x, masks
def load_weights(model):
f = h5py.File(WEIGHTS_PATH)
layer_names = [name for name in f.attrs['layer_names']]
for layer in model.layers:
b_name = layer.name.encode()
if b_name in layer_names:
g = f[b_name]
weights = [g[name] for name in g.attrs['weight_names']]
layer.set_weights(weights)
layer.trainable = False
f.close()
def VGG19(input_tensor=None, input_shape=None, target_layer=1):
"""
VGG19, up to the target layer (1 for relu1_1, 2 for relu2_1, etc.)
"""
if input_tensor is None:
inputs = Input(shape=input_shape)
else:
inputs = Input(tensor=input_tensor, shape=input_shape)
layers, masks = vgg_layers(inputs, target_layer)
model = Model(inputs, layers, name='vgg19')
load_weights(model)
return model, masks