Keras，Theano - Cropping2D在GPU上引入「輸入尺寸不匹配」

Question

我試圖用keras創建一個自編碼器。 （參考：https://blog.keras.io/building-autoencoders-in-keras.html）

下面的測試代碼工作在我的Mac就好了：當我測試這段代碼在AWS EC2 g2.2xlarge

import numpy as np 
from keras.layers import Input, Dense, Convolution2D, MaxPooling2D, UpSampling2D, Cropping2D, ZeroPadding2D 
from keras.models import Model 

input_img = Input(shape=(1, 30, 32)) 

testx = np.random.rand(10,1,30,32) 

x = ZeroPadding2D((1, 0))(input_img) 
x = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(x) 
x = MaxPooling2D((2, 2), border_mode='same')(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
x = MaxPooling2D((2, 2), border_mode='same')(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
encoded = MaxPooling2D((2, 2), border_mode='same')(x) 

x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(encoded) 
x = UpSampling2D((2, 2))(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
x = UpSampling2D((2, 2))(x) 
x = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(x) 
x = UpSampling2D((2, 2))(x) 
x = Cropping2D(cropping=((1,1),(0,0)))(x) 
decoded = Convolution2D(1, 3, 3, activation='sigmoid', border_mode='same')(x) 

autoencoder = Model(input_img, decoded) 
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy') 
autoencoder.fit(testx, testx, nb_epoch = 5, batch_size =1) 

但是，（我的配置信息是在這篇文章的結束），我遇到了以下錯誤：

Using Theano backend. 
Using gpu device 0: GRID K520 (CNMeM is enabled with initial size: 95.0% of memory, cuDNN 5005) 
____________________________________________________________________________________________________ 
Layer (type)      Output Shape   Param #  Connected to      
==================================================================================================== 
input_1 (InputLayer)    (None, 1, 30, 32)  0            
____________________________________________________________________________________________________ 
zeropadding2d_1 (ZeroPadding2D) (None, 1, 32, 32)  0   input_1[0][0]      
____________________________________________________________________________________________________ 
convolution2d_1 (Convolution2D) (None, 16, 32, 32) 160   zeropadding2d_1[0][0]    
____________________________________________________________________________________________________ 
maxpooling2d_1 (MaxPooling2D) (None, 16, 16, 16) 0   convolution2d_1[0][0]    
____________________________________________________________________________________________________ 
convolution2d_2 (Convolution2D) (None, 8, 16, 16)  1160  maxpooling2d_1[0][0]    
____________________________________________________________________________________________________ 
maxpooling2d_2 (MaxPooling2D) (None, 8, 8, 8)  0   convolution2d_2[0][0]    
____________________________________________________________________________________________________ 
convolution2d_3 (Convolution2D) (None, 8, 8, 8)  584   maxpooling2d_2[0][0]    
____________________________________________________________________________________________________ 
maxpooling2d_3 (MaxPooling2D) (None, 8, 4, 4)  0   convolution2d_3[0][0]    
____________________________________________________________________________________________________ 
convolution2d_4 (Convolution2D) (None, 8, 4, 4)  584   maxpooling2d_3[0][0]    
____________________________________________________________________________________________________ 
upsampling2d_1 (UpSampling2D) (None, 8, 8, 8)  0   convolution2d_4[0][0]    
____________________________________________________________________________________________________ 
convolution2d_5 (Convolution2D) (None, 8, 8, 8)  584   upsampling2d_1[0][0]    
____________________________________________________________________________________________________ 
upsampling2d_2 (UpSampling2D) (None, 8, 16, 16)  0   convolution2d_5[0][0]    
____________________________________________________________________________________________________ 
convolution2d_6 (Convolution2D) (None, 16, 16, 16) 1168  upsampling2d_2[0][0]    
____________________________________________________________________________________________________ 
upsampling2d_3 (UpSampling2D) (None, 16, 32, 32) 0   convolution2d_6[0][0]    
____________________________________________________________________________________________________ 
cropping2d_1 (Cropping2D)  (None, 16, 30, 32) 0   upsampling2d_3[0][0]    
____________________________________________________________________________________________________ 
convolution2d_7 (Convolution2D) (None, 1, 30, 32)  145   cropping2d_1[0][0]    
==================================================================================================== 
Total params: 4,385 
Trainable params: 4,385 
Non-trainable params: 0 
____________________________________________________________________________________________________ 
Epoch 1/5 
Traceback (most recent call last): 
    File "testtest.py", line 32, in <module> 
    autoencoder.fit(testx, testx, nb_epoch = 5, batch_size =1) 
    File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 1144, in fit 
    initial_epoch=initial_epoch) 
    File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 844, in _fit_loop 
    outs = f(ins_batch) 
    File "/usr/local/lib/python2.7/dist-packages/keras/backend/theano_backend.py", line 953, in __call__ 
    return self.function(*inputs) 
    File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 886, in __call__ 
    storage_map=getattr(self.fn, 'storage_map', None)) 
    File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 325, in raise_with_op 
    reraise(exc_type, exc_value, exc_trace) 
    File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 873, in __call__ 
    self.fn() if output_subset is None else\ 
ValueError: Input dimension mis-match. (input[0].shape[3] = 32, input[1].shape[3] = 0) 
Apply node that caused the error: Elemwise{mul,no_inplace}(convolution2d_7_target, Elemwise{log,no_inplace}.0) 
Toposort index: 703 
Inputs types: [TensorType(float32, 4D), TensorType(float32, 4D)] 
Inputs shapes: [(1, 1, 30, 32), (1, 1, 30, 0)] 
Inputs strides: [(3840, 3840, 128, 4), (120, 120, 4, 4)] 
Inputs values: ['not shown', array([], shape=(1, 1, 30, 0), dtype=float32)] 
Outputs clients: [[Elemwise{add,no_inplace}(Elemwise{mul,no_inplace}.0, Elemwise{mul,no_inplace}.0)]] 

Backtrace when the node is created(use Theano flag traceback.limit=N to make it longer): 
    File "testtest.py", line 30, in <module> 
    autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy') 
    File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 619, in compile 
    sample_weight, mask) 
    File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 307, in weighted 
    score_array = fn(y_true, y_pred) 
    File "/usr/local/lib/python2.7/dist-packages/keras/objectives.py", line 49, in binary_crossentropy 
    return K.mean(K.binary_crossentropy(y_pred, y_true), axis=-1) 
    File "/usr/local/lib/python2.7/dist-packages/keras/backend/theano_backend.py", line 1235, in binary_crossentropy 
    return T.nnet.binary_crossentropy(output, target) 

如果去除Cropping2D功能，我可以運行在GPU代碼：

import numpy as np 
from keras.layers import Input, Dense, Convolution2D, MaxPooling2D, UpSampling2D, Cropping2D, ZeroPadding2D 
from keras.models import Model 

input_img = Input(shape=(1, 30, 32)) 

testx = np.random.rand(10,1,30,32) 
testy = np.random.rand(10,1,32,32) 

x = ZeroPadding2D((1, 0))(input_img) 
x = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(x) 
x = MaxPooling2D((2, 2), border_mode='same')(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
x = MaxPooling2D((2, 2), border_mode='same')(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
encoded = MaxPooling2D((2, 2), border_mode='same')(x) 

# at this point the representation is (8, 4, 4) i.e. 128-dimensional 

x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(encoded) 
x = UpSampling2D((2, 2))(x) 
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x) 
x = UpSampling2D((2, 2))(x) 
x = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(x) 
x = UpSampling2D((2, 2))(x) 
# x = Cropping2D(cropping=((1,1),(0,0)))(x) 
decoded = Convolution2D(1, 3, 3, activation='sigmoid', border_mode='same')(x) 

autoencoder = Model(input_img, decoded) 
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy') 
autoencoder.summary() 
autoencoder.fit(testx, testy, nb_epoch = 5, batch_size =1) 

在類似的情況下，也Cropping3D提高類似的錯誤... :-(

我的問題是：

我怎麼能解決這個問題，並在GPU上運行Cropping2D/Cropping3D？

ps。我的AWS EC2 g2.2xlarge配置如下：

• Ubuntu的14.04
• Cuda的7.5
• Cudnn V5
• Keras（1.2.0）
• Theano（0.9.0dev4）

謝謝！

Answer 1

Keras似乎有一個bug。 它一直在照顧。案件結案！