AttributeError：'順序'沒有屬性'get_shape'當合並模型

Question

我想創建兩個順序模型（每個訓練不同的數據集 - 不同的圖像）。然後，我想取其平均輸出，並添加一個softmax圖層，根據兩個連續模型爲我提供單一分類輸出。我的代碼如下，但我得到一個屬性錯誤，說'順序'對象沒有屬性'get_shape'。

完整的錯誤代碼是：

Traceback (most recent call last): 
    File "Mergedmodels.pyu", line 135, in <module> 
    merged = average ([modelo, modelN1]) 
    File "G:\Anaconda\lib\site-packages\keras\layers\merge.py", line 481, in average 
    return Average(**kwargs)(inputs) 
    File "G:\Anaconda\lib\site-packages\keras\engine\topology.py", line 542, in _ call_input_shapes.append(K.int_sshape(x_elem)) 
    File "G:\Anaconda\lib\site-packages\keras\backend\tensorflow_backend.py", line 411, in int_shape 
    shape = x.get_shape() 
    AttributeError: 'Sequential' object has no attribute 'get_shape' 

就如何解決它的任何想法？

import numpy as np 
from keras.models import Sequential 
from keras.layers import Dense, Dropout, Activation, Flatten 
from keras.layers import merge 
from keras.layers import average 
from keras.layers import Convolution2D, MaxPooling2D 
from keras.utils import np_utils 
from keras.preprocessing.image import ImageDataGenerator 
from keras.datasets import mnist 
import pandas as pd 
from numpy import array 
from PIL import Image 
import matplotlib.pyplot as plt 
from keras import backend as K 
import glob 
import os 

K.set_image_dim_ordering('th') 


np.random.seed(123) #set for reproducibility 

size = 48, 48 

#IMPORTING TRAINING IMAGES FOR FIRST MODEL (ORIGINAL) 
folder = 'images' 

read = lambda imname: np.asarray(Image.open(imname).convert("RGB")) 

ims = [read(os.path.join(folder, filename)) for filename in os.listdir(folder)] 
X_train = np.array([read(os.path.join(folder, filename)) for filename in os.listdir(folder)], dtype='uint8') 
#CHECK print (X_train.shape) 

X_train = X_train.reshape(X_train.shape[0],3,48,48) 
#X_test = X_test.reshape(X_test.shape[0],1,28,28) 
X_train = X_train.astype ('float32') 
#X_test = X_test.astype ('float32') 
X_train /= 255 
#X_test /= 255 

#IMPORTING TRAINING IMAGES FOR SECOND MODEL (NORMALIZED) 
folder = 'images2' 

read = lambda imname: np.asarray(Image.open(imname).convert("RGB")) 

ims = [read(os.path.join(folder, filename)) for filename in os.listdir(folder)] 
X_training = np.array([read(os.path.join(folder, filename)) for filename in os.listdir(folder)], dtype='uint8') 
#CHECK print (X_train.shape) 

X_training = X_training.reshape(X_train.shape[0],3,48,48) 
#X_test = X_test.reshape(X_test.shape[0],1,28,28) 
X_training = X_training.astype ('float32') 
#X_test = X_test.astype ('float32') 
X_training /= 255 
#X_test /= 255 


#IMPORTING LABELS FOR 10K TRAINING IMAGES 
saved_column = pd.read_csv('labels4.csv') 

y_labels = array(saved_column) 

Y_train = np_utils.to_categorical(y_labels,501) 

#y_train = np.array ([0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1]) 
#(X_train, y_train),(X_test, y_test) = mnist.load_data() 

#COPYING LABELS FOR SECOND MODEL TRAINING IMAGES 
#Y_training = Y_train 

#IMPORTING TEST IMAGES 
folder2 = 'test' 
read = lambda imname: np.asarray(Image.open(imname).convert("RGB")) 
ims = [read(os.path.join(folder2, filename)) for filename in os.listdir(folder2)] 
X_test = np.array([read(os.path.join(folder2, filename)) for filename in os.listdir(folder2)], dtype='uint8') 

X_test = X_test.reshape(X_test.shape[0],3,48,48) 
X_test = X_test.astype ('float32') 
X_test /= 255 

#IMPORTING LABELS FOR TEST IMAGES 
another_column = pd.read_csv('labelstest4.csv') 
test_labels = array(another_column) 
Y_test = np_utils.to_categorical(test_labels,501) 
#train_labels = np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]) 
#Y_train = np_utils.to_categorical(y_train, 2) 
#Y_test = np_utils.to_categorical(y_test,10) 


#BUILDING FIRST NN FOR ORIGINAL IMAGES 
modelo = Sequential() 

modelo.add(Convolution2D(32,3,3, activation='relu', input_shape=(3,48,48), dim_ordering='th')) 
modelo.add(Convolution2D(32,3,3, activation = 'relu')) 
modelo.add(MaxPooling2D(pool_size=(2,2))) 
modelo.add(Dropout(0.25)) 

modelo.add(Flatten()) 
modelo.add(Dense(128,activation='relu')) 
modelo.add(Dropout(0.5)) 
modelo.add(Dense(501, activation = 'sigmoid')) 

modelo.compile(loss='categorical_crossentropy', 
    optimizer = 'adam', 
    metrics = ['accuracy']) 

modelo.fit(X_train, Y_train, 
    batch_size = 5, nb_epoch= 5, verbose = 1) 

score = modelo.evaluate(X_test, Y_test, verbose=0) 

#BUILDING SECOND NN FOR NORMALIZED IMAGES 
modelN1 = Sequential() 

modelN1.add(Convolution2D(32,3,3, activation='relu', input_shape=(3,48,48), dim_ordering='th')) 
modelN1.add(Convolution2D(32,3,3, activation = 'relu')) 
modelN1.add(MaxPooling2D(pool_size=(2,2))) 
modelN1.add(Dropout(0.25)) 

modelN1.add(Flatten()) 
modelN1.add(Dense(128,activation='relu')) 
modelN1.add(Dropout(0.5)) 
modelN1.add(Dense(501, activation = 'sigmoid')) 

modelN1.compile(loss='categorical_crossentropy', 
    optimizer = 'adam', 
    metrics = ['accuracy']) 

modelN1.fit(X_training, Y_train, 
    batch_size = 5, nb_epoch= 1, verbose = 1) 

score = modelN1.evaluate(X_test, Y_test, verbose=0) 

#MERGING MODELS 
merged = average([modelo, modelN1]) 

finalmodel = Sequential() 
finalmodel.add(merged) 
finalmodel.add(Dense(501, activation = 'softmax')) 

finalmodel.compile(loss='categorical_crossentropy', 
    optimizer = 'adam', 
    metrics = ['accuracy']) 

Y_madeuplabels = np.array ([0, 1, 52, 20]) 
Y_training = np_utils.to_categorical(Y_madeuplabels, 501) 


finalmodel.fit([X_train], Y_training, 
    batch_size = 5, nb_epoch= 1, verbose = 1) 

score = finalmodel.evaluate(X_test, Y_test, verbose=0) 

print ("the code ran") 

Answer 1

的組合順序模型似乎並沒有結束張量而不是層因爲average作品Keras 2.0 這樣的工作方式。

因此，錯誤消息是說Sequential模型沒有它所不具備的方法get_shape()。 get_shape()只存在於張量。

這裏是複製錯誤的例子：

mod1 = Sequential() 
mod1.add(Dense(1, input_shape=(10,))) 

mod2 = Sequential() 
mod2.add(Dense(1, input_shape=(10,))) 

avg = average([mod1, mod2]) # throws AttributeError 

哈克的方式來解決這個問題是使用functional API來 的兩個模型的輸出組合，然後做SOFTMAX層。舉個例子：

X1 = np.random.rand(10, 10) 
X2 = np.random.rand(10, 10) 
Y = np.random.choice(2, 10) 

mod1 = Sequential() 
mod1.add(Dense(16, input_shape=(10,))) 

mod2 = Sequential() 
mod2.add(Dense(16, input_shape=(10,))) 

# so use the outputs of the models to do the average over 
# this way we do averaging over tensor __not__ models. 
avg = average([mod1.output, mod2.output]) 
dense = Dense(1, activation="sigmoid")(avg) 

# the two inputs are the inputs to the sequential models 
# and the output is the dense layer 
mod3 = Model(inputs=[mod1.input, mod2.input], outputs=[dense]) 
mod3.compile(loss='binary_crossentropy', optimizer='sgd') 
mod3.fit([X1, X2], Y)