Question

我正在研究区分伪造签名和原始签名的计算机视觉问题。网络的输入是两个图像
为此，我从VGG 16的“ block3_pool”层（形状（None，28,28,256））中提取了特征，并创建了一个层来计算两个图像的256个滤镜的绝对差的平方根< / p>

但是，网络无法学习，并且对于每个时期都给出相同的训练和验证错误，调整学习速率不起作用，更改体系结构也不起作用。网络的输入是两个图像：锚点和形状为（224,224,3）的数据

我已经在非常小的数据集上对其进行了训练，预计它会过拟合，但是即使在修改数据集大小后网络也不会学习

数据的格式为每个用户有24个原始签名和24个伪造签名。从原始集合中随机选择一个将获得锚定数据，并在原始伪造和伪造签名之间随机选择将获得数据数组。我有一个锚点数组（同一样本的24张图像）和数据（24张不同原始和伪造样本的总图像），锚点的形状和一个用户的数据数组是：（24,224,224,3）



#model hyper parameters ,using sgd optimizer
epochs=50
learning_rate=0.1
decay=learning_rate/epochs
batch_size=8
keep_prob=0.8





#This is the function for lambda layer 'Layer_distance'
def root_diff(x):
    diff=K.sqrt(K.sum(K.abs(x[:,:,:,:,0]-x[:,:,:,:,1]),axis=(1,2)))
    return diff


#This creates an instance of a pre trained VGG-16 model
def base_model(input_dims=(224,224,3),output_dims=128):



 base_model=VGG16(include_top=False,weights='imagenet',input_shape=input_dims)

 for layers in base_model.layers:
   layers.trainable=False

 x=base_model.get_layer('block3_pool').output



 model=Model(inputs=base_model.input,outputs=x)
 return model



def siamese_model(anchor,data,label,anchor_valid,data_valid,label_valid,input_shape=(224,224,3)):

    anchor_input=Input(input_shape)
    data_input=Input(input_shape)



    #----------------------------Model begins from here-------------------------------------------------------#

    model_resnet=base_model(input_dims=input_shape)



    encodings_anchor=model_resnet(anchor_input)
    encodings_data=model_resnet(data_input)








    layer_expand_dims=Lambda(lambda x:K.expand_dims(x,axis=4))





    anchor_expanded=layer_expand_dims(encodings_anchor)
    data_expanded=layer_expand_dims(encodings_data)



    encodings=concatenate([anchor_expanded,data_expanded],axis=4) #gives the shape as (None,28,28,256,2)

    Layer_distance=Lambda(root_diff)(encodings) #Should give a vector of (256)







    dense_1=Dense(256,activation=None,kernel_initializer='glorot_uniform',bias_initializer='zeros')(Layer_distance)


    prediction=Dense(1,activation='sigmoid',kernel_initializer='glorot_uniform')(dense_1)




    # Connect the inputs with the outputs
    siamese_net = Model(inputs=[anchor_input,data_input],outputs=prediction)

    print(siamese_net.summary())

    for layer in siamese_net.layers:
        print("Input shape: "+str(layer.input_shape)+". Output shape: "+str(layer.output_shape))


    sgd= optimizers.SGD(lr=learning_rate, decay=1e-9, momentum=0.9, nesterov=True)
    siamese_net.compile(loss='binary_crossentropy', optimizer=sgd,metrics=['accuracy'])

    history=siamese_net.fit(x=[anchor,data],y=label,batch_size=batch_size,epochs=epochs

                            ,validation_data=([anchor_valid,data_valid],label_valid))

模型摘要（siamese_net）

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 224, 224, 3)  0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, 224, 224, 3)  0                                            
__________________________________________________________________________________________________
model_1 (Model)                 (None, 28, 28, 256)  1735488     input_1[0][0]                    
                                                                 input_2[0][0]                    
__________________________________________________________________________________________________
lambda_1 (Lambda)               (None, 28, 28, 256,  0           model_1[1][0]                    
                                                                 model_1[2][0]                    
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 28, 28, 256,  0           lambda_1[0][0]                   
                                                                 lambda_1[1][0]                   
__________________________________________________________________________________________________
lambda_2 (Lambda)               (None, 256)          0           concatenate_1[0][0]              
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 256)          65792       lambda_2[0][0]                   
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 1)            257         dense_1[0][0]                    
==================================================================================================
Total params: 1,801,537
Trainable params: 66,049
Non-trainable params: 1,735,488
__________________________________________________________________________________________________

训练结果

WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.cast instead.
Train on 48 samples, validate on 48 samples
Epoch 1/50
2019-04-21 06:10:00.354542: I tensorflow/stream_executor/dso_loader.cc:152] successfully opened CUDA library libcublas.so.10.0 locally
48/48 [==============================] - 4s 90ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 2/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 3/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 4/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 5/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 6/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 7/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 8/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 9/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 10/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 11/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 12/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 13/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 14/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 15/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 16/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 17/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 18/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 19/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 20/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 21/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 22/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 23/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 24/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 25/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 26/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 27/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 28/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 29/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 30/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 31/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 32/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 33/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 34/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 35/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 36/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 37/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 38/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 39/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 40/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 41/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 42/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 43/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 44/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 45/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 46/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 47/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 48/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 49/50
48/48 [==============================] - 1s 18ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Epoch 50/50
48/48 [==============================] - 1s 19ms/step - loss: 8.9676 - acc: 0.4375 - val_loss: 8.6355 - val_acc: 0.4583
Saved model to disk

Answer 1

@Paras doshi

我认为您正在冻结整个基础网络。因此，训练过程没有更新参数。您正在使用第三个VGG块的特征向量，因此您需要解冻这3个块中的至少一个。

def base_model(input_dims=(224,224,3),output_dims=128):



 base_model=VGG16(include_top=False,weights='imagenet',input_shape=input_dims)

 for layers in base_model.layers:
   layers.trainable=False    <---- Freeze training

希望有帮助

在50个纪元给出相同的训练错误和相同的验证错误之后，神经网络不学习

1 个答案: