例程1 获取中间某一层的权重和偏置,并打印
from keras.models import Sequential,Model
from keras.layers import Dense
import numpy as np
model = Sequential()
model.add(Dense(32,activation="relu",input_dim=100))
model.add(Dense(16,activation="relu",name="Dense_1"))
model.add(Dense(1, activation='sigmoid',name="Dense_2"))
model.compile(optimizer='rmsprop',loss='binary_crossentropy',metrics=['accuracy'])
#假设训练和测试使用同一组数据
data = np.random.random((1000, 100))
labels = np.random.randint(2, size=(1000, 1))
model.fit(data,labels,epochs=10,batch_size=32)
#取某一层的输出为输出新建为model,采用函数模型
dense1_layer_model = Model(inputs=model.input,outputs=model.get_layer('Dense_1').output)
dense1_output = dense1_layer_model.predict(data)
print (dense1_output.shape)
#获得某一层的权重和偏置
weight_Dense_1,bias_Dense_1 = model.get_layer('Dense_1').get_weights()
print(weight_Dense_1.shape)
print(bias_Dense_1.shape)
例程2 模型训练完毕后,根据输入,预测中间层某一个变量对应的值
方法1、将输入和中间层变量构建一个新模型
x_in = Input(shape=(784,))
x = x_in
x = Dense(512, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(256, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(num_classes, activation='softmax')(x)
model = Model(x_in, x)
假如模型训练完成后,我想要获取x = Dense(256, activation='relu')(x)
对应的输出,那可以在定义模型的时候,先把对应的变量存起来,然后重新定义一个模型:
x_in = Input(shape=(784,))
x = x_in
x = Dense(512, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(256, activation='relu')(x)
y = x
x = Dropout(0.2)(x)
x = Dense(num_classes, activation='softmax')(x)
model = Model(x_in, x)
model2 = Model(x_in, y)
将model
训练完成后,直接用model2.predict
就可以查看对应的256维的输出了。这样做的前提是y
必须是某个层的输出,不能是随意一个张量。
相关参考::https://spaces.ac.cn/archives/6575
方法2、使用get_layer获取中间层变量,并根据输入得到其对应的输出
'''
Created on 2018-4-16
'''
import keras
from keras.models import Sequential
from keras.layers import Dense
from keras.models import Model
from keras.callbacks import ModelCheckpoint,Callback
import numpy as np
import tflearn
import tflearn.datasets.mnist as mnist
x_train, y_train, x_test, y_test = mnist.load_data(one_hot=True)
x_valid = x_test[:5000]
y_valid = y_test[:5000]
x_test = x_test[5000:]
y_test = y_test[5000:]
print(x_valid.shape)
print(x_test.shape)
model = Sequential()
model.add(Dense(units=64, activation='relu', input_dim=784))
model.add(Dense(units=10, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
filepath = 'D:\\machineTest\\model-ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'
# filepath = 'D:\\machineTest\\model-ep{epoch:03d}-loss{loss:.3f}.h5'
checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min')
print(model.get_config())
# [{'class_name': 'Dense', 'config': {'bias_regularizer': None, 'use_bias': True, 'kernel_regularizer': None, 'batch_input_shape': (None, 784), 'trainable': True, 'kernel_constraint': None, 'bias_constraint': None, 'kernel_initializer': {'class_name': 'VarianceScaling', 'config': {'scale': 1.0, 'distribution': 'uniform', 'mode': 'fan_avg', 'seed': None}}, 'activity_regularizer': None, 'units': 64, 'dtype': 'float32', 'bias_initializer': {'class_name': 'Zeros', 'config': {}}, 'activation': 'relu', 'name': 'dense_1'}}, {'class_name': 'Dense', 'config': {'bias_regularizer': None, 'use_bias': True, 'kernel_regularizer': None, 'bias_initializer': {'class_name': 'Zeros', 'config': {}}, 'kernel_constraint': None, 'bias_constraint': None, 'kernel_initializer': {'class_name': 'VarianceScaling', 'config': {'scale': 1.0, 'distribution': 'uniform', 'mode': 'fan_avg', 'seed': None}}, 'activity_regularizer': None, 'trainable': True, 'units': 10, 'activation': 'softmax', 'name': 'dense_2'}}]
# model.fit(x_train, y_train, epochs=1, batch_size=128, callbacks=[checkpoint],validation_data=(x_valid, y_valid))
model.fit(x_train, y_train, epochs=1,validation_data=(x_valid, y_valid),steps_per_epoch=10,validation_steps=1)
# score = model.evaluate(x_test, y_test, batch_size=128)
# print(score)
# #获取模型结构状况
# model.summary()
# _________________________________________________________________
# Layer (type) Output Shape Param #
# =================================================================
# dense_1 (Dense) (None, 64) 50240(784*64+64(b))
# _________________________________________________________________
# dense_2 (Dense) (None, 10) 650(64*10 + 10 )
# =================================================================
# #根据下标和名称返回层对象
# layer = model.get_layer(index = 0)
# 获取模型权重,设置权重model.set_weights()
weights = np.array(model.get_weights())
print(weights.shape)
# (4,)权重由4部分组成
print(weights[0].shape)
# (784, 64)dense_1 w1
print(weights[1].shape)
# (64,)dense_1 b1
print(weights[2].shape)
# (64, 10)dense_2 w2
print(weights[3].shape)
# (10,)dense_2 b2
# # 保存权重和加载权重
# model.save_weights("D:\\xxx\\weights.h5")
# model.load_weights("D:\\xxx\\weights.h5", by_name=False)#by_name=True,可以根据名字匹配和层载入权重
# 查看中间结果,必须要先声明个函数式模型
dense1_layer_model = Model(inputs=model.input,outputs=model.get_layer('dense_1').output)
out = dense1_layer_model.predict(x_test)
print(out.shape)
# (5000, 64)
# 如果是函数式模型,则可以直接输出
# import keras
# from keras.models import Model
# from keras.callbacks import ModelCheckpoint,Callback
# import numpy as np
# from keras.layers import Input,Conv2D,MaxPooling2D
# import cv2
#
# image = cv2.imread("D:\\machineTest\\falali.jpg")
# print(image.shape)
# cv2.imshow("1",image)
#
# # 第一层conv
# image = image.reshape([-1, 386, 580, 3])
# img_input = Input(shape=(386, 580, 3))
# x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv1')(img_input)
# x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x)
# x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
# model = Model(inputs=img_input, outputs=x)
# out = model.predict(image)
# print(out.shape)
# out = out.reshape(193, 290,64)
# image_conv1 = out[:,:,1].reshape(193, 290)
# image_conv2 = out[:,:,20].reshape(193, 290)
# image_conv3 = out[:,:,40].reshape(193, 290)
# image_conv4 = out[:,:,60].reshape(193, 290)
# cv2.imshow("conv1",image_conv1)
# cv2.imshow("conv2",image_conv2)
# cv2.imshow("conv3",image_conv3)
# cv2.imshow("conv4",image_conv4)
# cv2.waitKey(0)
中间结果输出可以查看conv过之后的图像:
原始图像:
经过一层conv以后,输出其中4张图片:
相关参考:
https://blog.csdn.net/u011311291/article/details/79963831