Faster R-CNN系列（三）：用keras構建多輸入-多輸出-多自定義loss的模型

用keras構建多輸入-多輸出-多自定義loss的模型

keras 有兩種模型

• Sequential 順序模型
• Model 函數式模型 Model(inputs,outputs,name=None)

from keras import Sequential, Model
import keras.layers as KL
import keras.backend as K
import numpy as np

Using TensorFlow backend.

​

本文實現的模型

定義輸入

input_tensor1=KL.Input((32,32,3))
input_tensor2=KL.Input((4,))
input_target=KL.Input((2,))

input_tensor1 支路

x=KL.BatchNormalization(axis=-1)(input_tensor1)
x=KL.Conv2D(16,(3,3),padding="same")(x)
x=KL.Activation("relu")(x)
x=KL.MaxPooling2D(2)(x)
x=KL.Conv2D(32,(3,3),padding="same")(x)
x=KL.Activation("relu")(x)
x=KL.MaxPooling2D(2)(x)   #下采樣兩次後 從32->8
x=KL.Flatten()(x)
x=KL.Dense(32)(x)
x=KL.Dense(2)(x)

out2=x   #命名問題 input_tensor1 對應 out2  見上圖 

input_tensor2 支路

y=KL.Dense(32)(input_tensor2)
y=KL.Dense(2)(y)

out1=y

定義模型

model_temp=Model([input_tensor1,input_tensor2,input_target],[out1,out2])

可視化之前編寫的模型

• 查看模型結構
• 文字方法表述

model_temp.summary()

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 32, 32, 3)    0                                            
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 32, 32, 3)    12          input_1[0][0]                    
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 32, 32, 16)   448         batch_normalization_1[0][0]      
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 32, 32, 16)   0           conv2d_1[0][0]                   
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D)  (None, 16, 16, 16)   0           activation_1[0][0]               
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 16, 16, 32)   4640        max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 16, 16, 32)   0           conv2d_2[0][0]                   
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D)  (None, 8, 8, 32)     0           activation_2[0][0]               
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, 4)            0                                            
__________________________________________________________________________________________________
flatten_1 (Flatten)             (None, 2048)         0           max_pooling2d_2[0][0]            
__________________________________________________________________________________________________
dense_3 (Dense)                 (None, 32)           160         input_2[0][0]                    
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 32)           65568       flatten_1[0][0]                  
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 2)            66          dense_3[0][0]                    
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 2)            66          dense_1[0][0]                    
==================================================================================================
Total params: 70,960
Trainable params: 70,954
Non-trainable params: 6
__________________________________________________________________________________________________

圖形化可視化

• 下載graphviz 並添加 Graphviz2.38\bin 到環境變量
• pip install pydot-ng
• 管理員權限開啓jupyter notebook

from keras.utils.vis_utils import plot_model
plot_model(model_temp,to_file="images/model_temp.png",show_shapes=True)

定義loss

• 把loss 定義成一個層 比較有條理
• keras 定義層的方法(兩種)
  • 方法1 使用KL.Lambda構建層 func 爲函數
    KL.Lambda(func)(input)
  • 方法2 繼承layer並重寫類實現自定義layer
    import keras.engine as KE

#定義loss
def custom_loss1(y_true,y_pred):
    return K.mean(K.abs(y_true-y_pred))

# s使用方法1  使用KL.Lambda構建層  func 爲函數
loss1=KL.Lambda(lambda x:custom_loss1(*x),name="loss1")([out2,out1])
loss2=KL.Lambda(lambda x:custom_loss1(*x),name="loss2")([input_target,out2])

model=Model([input_tensor1,input_tensor2,input_target],[out1,out2,loss1,loss2])

plot_model(model,to_file="images/model.png",show_shapes=True)

取出loss層model.get_layer() 並取出loss層的輸出 model.get_layer(name).output

loss_layer1=model.get_layer("loss1").output
loss_layer2=model.get_layer("loss2").output
model.add_loss(loss_layer1)
model.add_loss(loss_layer2)
#已經add_loss 過了 所以一下可以設置爲None
model.compile(optimizer='sgd',loss=[None,None,None,None])