2017-05-12 43 views
1

目標:評估一組5幅圖像並生成一幅圖像作爲輸出。Keras錯誤擬合模型:提供的元素過多

問題:我目前收到錯誤提供的元素太多。

聲明:我是新來的Keras和深度學習作爲一個整體,我完全相信,這種做法是錯誤的,但我想知道爲什麼我得到這個錯誤

輸出的形狀並輸入看起來正確的給我。我試過讓輸出只是一個形狀爲(無,6912​​)的密集層。
我試過讓輸出成爲Conv2d,但後來我得到以下錯誤,我不確定輸出是爲什麼(46,46,3),而不是(48,48,3)

Error when checking target: expected conv2d_1 to have shape (None, 46, 46, 3) but got array with shape (379, 48, 48, 3) 

代碼:

width = 48 
height = 48 
png = [] 

for image_path in glob.glob(r"D:\temp\*.png"): 
    png.append(misc.imread(image_path)) 

im = np.asarray(png) 
print ('dataset: ', im.shape) 

window = 6 
dataset = np.zeros([len(im) - window, window,width,height,3]) 
for i in range(len(dataset)): 
    dataset[i, :] = im[i:i + window] 
x_train = dataset[:,:-1] 
y_train = dataset[:,-1] 
y_train1 = y_train.reshape(-1,width*height*3) 

print("x_train: ", x_train.shape) 
print("y_train:" ,y_train.shape) 
print("y_train1:" ,y_train1.shape) 

model = Sequential() 
model.add(Conv3D(filters=40, 
       kernel_size=(5,10,10), 
       input_shape=(5,width,height,3), 
       padding='same', 
       activation='relu')) 
model.add(Activation('relu')) 
model.add(Conv3D(filters=40, 
       kernel_size=(3,3,3), 
       padding='same', 
       activation='relu')) 

model.add(Flatten()) 
model.add(Dropout(0.5)) 
model.add(Dense(width * height * 3, activation='softmax')) 
model.add(Reshape((48,48,3))) 

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

print("model Input: " ,model.input_shape) 
print("model output:", model.output_shape) 

model.fit(x_train, y_train, batch_size=10, epochs=300, validation_split=0.05) 

輸出:

Using TensorFlow backend. 
dataset: (385, 48, 48, 3) 
x_train: (379, 5, 48, 48, 3) 
y_train: (379, 48, 48, 3) 
y_train1: (379, 6912) 
model Input: (None, 5, 48, 48, 3) 
model output: (None, 48, 48, 3) 
Traceback (most recent call last): 
    .... 
    File "D:\Program Files\Python35\lib\site-packages\tensorflow\python\framework\tensor_util.py", line 424, in make_tensor_proto 
    (shape_size, nparray.size)) 
ValueError: Too many elements provided. Needed at most -1109917696, but received 1 

模型摘要:

_________________________________________________________________ 
Layer (type)     Output Shape    Param # 
================================================================= 
conv3d_1 (Conv3D)   (None, 5, 48, 48, 40)  60040  
_________________________________________________________________ 
activation_1 (Activation) (None, 5, 48, 48, 40)  0   
_________________________________________________________________ 
conv3d_2 (Conv3D)   (None, 5, 48, 48, 40)  43240  
_________________________________________________________________ 
flatten_1 (Flatten)   (None, 460800)   0   
_________________________________________________________________ 
dropout_1 (Dropout)   (None, 460800)   0   
_________________________________________________________________ 
dense_1 (Dense)    (None, 6912)    -110991078 
_________________________________________________________________ 
reshape_1 (Reshape)   (None, 48, 48, 3)   0   
================================================================= 

感謝先進。

+0

在conv2D的情況下,您可能忘記使用'padding ='same''。使用(3,3)內核將丟棄2個像素。 ----現在,除了使用'categorical_crossentropy'之外,在代碼中看不到任何問題。你只在分類問題中使用它,爲了以防萬一,我建議你試試'mse'。 –

+0

'model.summary()'的輸出是什麼? –

+0

我已更新模型摘要。我確實發現了「-110991078」這個有趣的數字。關於conv2D你是對的,我沒有'padding ='same''。我也會嘗試'mse' –

回答

0

Conv2D的情況下,您可能已經忘記使用padding = 'same'。 A(3,3)內核在每個維度中刪除2個像素。

並且通過在密集層中顯示的負數參數,我相信你的模型比支持的要大。

也許單層的參數有最大限制。在那種情況下,我會減少卷積中的濾波器數量,或者在平坦層之前對通道進行求和。