keras讀取訓(xùn)練好的模型參數(shù)并把參數(shù)賦值給其它模型詳解

更新時間：2020年06月15日 11:15:51 作者：Tom Hardy

這篇文章主要介紹了keras讀取訓(xùn)練好的模型參數(shù)并把參數(shù)賦值給其它模型詳解，具有很好的參考價值，希望對大家有所幫助。一起跟隨小編過來看看吧

介紹

本博文中的代碼，實現(xiàn)的是加載訓(xùn)練好的模型model_halcon_resenet.h5，并把該模型的參數(shù)賦值給兩個不同的新的model。

函數(shù)式模型

官網(wǎng)上給出的調(diào)用一個訓(xùn)練好模型，并輸出任意層的feature。

model = Model(inputs=base_model.input, outputs=base_model.get_layer(‘block4_pool').output)

但是這有一個問題，就是新的model，如果輸入inputs和訓(xùn)練好的model的inputs大小不同呢？比如我想建立一個輸入是600x600x3的新model，但是訓(xùn)練好的model輸入是200x200x3，而這時我又想調(diào)用訓(xùn)練好模型的卷積核參數(shù)，這時該怎么辦呢？

其實想一下，用訓(xùn)練好的模型參數(shù)，即使輸入的尺寸不同，但是這些模型參數(shù)仍然可以處理計算，只是輸出的feature map大小不同。那到底怎么賦值呢？其實很簡單

在定義新的model時，新的model層在定義時，需要加上名字，而這個名字就是訓(xùn)練好的模型的每層名字。如下代碼所示：

inputs=Input(shape=(400,500,3))
X=Conv2D(32, (3, 3),name=“conv2d_1”)(inputs)
X=BatchNormalization(name=“batch_normalization_1”)(X)
X=Activation(‘relu',name=“activation_1”)(X)

最后通過以下代碼即可建立一個新的模型并擁有訓(xùn)練好模型的參數(shù)：

model=Model(inputs=inputs, outputs=X)
model.load_weights(‘model_halcon_resenet.h5', by_name=True)

源代碼

from keras.models import load_model
from keras.preprocessing import image
from keras.applications.vgg19 import preprocess_input
from keras.models import Model
import numpy as np
from keras.layers import Conv2D, MaxPooling2D,merge
from keras.layers import BatchNormalization,Activation
from keras.layers import Input, Dense
from PIL import Image
import numpy as np
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten,Input
from keras.layers import Conv2D, MaxPooling2D,merge,AveragePooling2D,GlobalAveragePooling2D
from keras.layers import BatchNormalization,Activation
from sklearn.model_selection import train_test_split
from keras.applications.densenet import DenseNet169, DenseNet121
from keras.applications.inception_resnet_v2 import InceptionResNetV2
from keras.applications.inception_v3 import InceptionV3
from keras.optimizers import SGD
from keras import regularizers
from keras.models import Model
import tensorflow as tf
from PIL import Image
from keras.callbacks import TensorBoard
import os
import cv2
from keras import backend as K
from model import focal_loss
import keras.losses

#ReadMe 該代碼是參考fast rcnn系列，先對整幅圖像提取特征feature map，然后從原圖對應(yīng)位置上映射到feature map，并對feature map進行
# 切片，從而提取對應(yīng)某個位置上的特征，并把該特征送進后面的識別網(wǎng)絡(luò)進行分類識別。
keras.losses.focal_loss = focal_loss#這句代碼是為了引入定義的loss
base_model=load_model('model_halcon_resenet.h5')
base_model.summary()

inputs=Input(shape=(400,500,3))
X=Conv2D(32, (3, 3),name="conv2d_1")(inputs)
X=BatchNormalization(name="batch_normalization_1")(X)
X=Activation('relu',name="activation_1")(X)
#第一個殘差模塊
X_1=Conv2D(32, (3, 3),padding='same',name="conv2d_2")(X)
X_1=BatchNormalization(name="batch_normalization_2")(X_1)
X_1= Activation('relu',name="activation_2")(X_1)
X_1 = Conv2D(32, (3, 3),padding='same',name="conv2d_3")(X_1)
X_1 = BatchNormalization(name="batch_normalization_3")(X_1)
merge_data = merge([X_1, X], mode='sum',name="merge_1")
X = Activation('relu',name="activation_3")(merge_data)
#第一個殘差模塊結(jié)束
X=MaxPooling2D(pool_size=(2, 2),strides=(2,2),name="max_pooling2d_1")(X)
X=Conv2D(64, (3, 3),kernel_regularizer=regularizers.l2(0.01),name="conv2d_4")(X)
X=BatchNormalization(name="batch_normalization_4")(X)
X=Activation('relu',name="activation_4")(X)
#第二個殘差模塊
X_2=Conv2D(64, (3, 3),padding='same',name="conv2d_5")(X)
X_2=BatchNormalization(name="batch_normalization_5")(X_2)
X_2= Activation('relu',name="activation_5")(X_2)
X_2 = Conv2D(64, (3, 3),padding='same',name="conv2d_6")(X_2)
X_2 = BatchNormalization(name="batch_normalization_6")(X_2)
merge_data = merge([X_2, X], mode='sum',name="merge_2")
X = Activation('relu',name="activation_6")(merge_data)
#第二個殘差模塊結(jié)束
X = MaxPooling2D(pool_size=(2, 2), strides=(2, 2),name="max_pooling2d_2")(X)
X=Conv2D(64, (3, 3),name="conv2d_7")(X)
X=BatchNormalization(name="batch_normalization_7")(X)
X=Activation('relu',name="activation_7")(X)
X=MaxPooling2D(pool_size=(2, 2),strides=(2,2),name="max_pooling2d_3")(X)
#第三個殘差模塊開始
X_3=Conv2D(64, (3, 3),padding='same',name="conv2d_8")(X)
X_3=BatchNormalization(name="batch_normalization_8")(X_3)
X_3= Activation('relu',name="activation_8")(X_3)
X_3 = Conv2D(64, (3, 3),padding='same',name="conv2d_9")(X_3)
X_3 = BatchNormalization(name="batch_normalization_9")(X_3)
merge_data = merge([X_3, X], mode='sum',name="merge_3")
X = Activation('relu',name="activation_9")(merge_data)
#第三個殘差模塊結(jié)束
X=Conv2D(32, (3, 3),kernel_regularizer=regularizers.l2(0.01),name="conv2d_10")(X)
X=BatchNormalization(name="batch_normalization_10")(X)
X=Activation('relu',name="activation_10")(X)
#第四個殘差模塊開始
X_4=Conv2D(32, (3, 3),padding='same',name="conv2d_11")(X)
X_4=BatchNormalization(name="batch_normalization_11")(X_4)
X_4= Activation('relu',name="activation_11")(X_4)
X_4 = Conv2D(32, (3, 3),padding='same',name="conv2d_12")(X_4)
X_4 = BatchNormalization(name="batch_normalization_12")(X_4)
merge_data = merge([X_4, X], mode='sum',name="merge_4")
X = Activation('relu',name="activation_12")(merge_data)
#第四個殘差模塊結(jié)束
X = MaxPooling2D(pool_size=(2, 2), strides=(2, 2),name="max_pooling2d_4")(X)
X = Conv2D(64, (3, 3),name="conv2d_13")(X)
X = BatchNormalization(name="batch_normalization_13")(X)
X = Activation('relu',name="activation_13")(X)
#第五個殘差模塊開始
X_5=Conv2D(64, (3, 3),padding='same',name="conv2d_14")(X)
X_5=BatchNormalization(name="batch_normalization_14")(X_5)
X_5= Activation('relu',name="activation_14")(X_5)
X_5 = Conv2D(64, (3, 3),padding='same',name="conv2d_15")(X_5)
X_5 = BatchNormalization(name="batch_normalization_15")(X_5)
merge_data = merge([X_5, X], mode='sum',name="merge_5")
X = Activation('relu',name="activation_15")(merge_data)
#第五個殘差模塊結(jié)束
model=Model(inputs=inputs, outputs=X)
model.load_weights('model_halcon_resenet.h5', by_name=True)
#讀取指定圖像數(shù)據(jù)
image_dir='C:/Users/18301/Desktop/blister/new/blister_mixed_11.png'
img = image.load_img(image_dir, target_size=(400, 500))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
#利用第一個模型預(yù)測出特征數(shù)據(jù)，并對特征數(shù)據(jù)進行切片
feature_map=model.predict(x)
T=np.array(feature_map)
f_1=T[:,16:21,0:10,:]
print(f_1.shape)
print(feature_map.shape)
#第一個模型沒有問題
#定義第二個模型
inputs_sec=Input(shape=(1,5,10,64))
X_= Flatten(name="flatten_1")(inputs_sec)
X_ = Dense(256, activation='relu',name="dense_1")(X_)
X_ = Dropout(0.5,name="dropout_1")(X_)
predictions = Dense(6, activation='softmax',name="dense_2")(X_)
model_sec=Model(inputs=inputs_sec, outputs=predictions)
model_sec.load_weights('model_halcon_resenet.h5', by_name=True)
#第二個模型定義結(jié)束
model_sec.summary()
#開始對整幅圖像進行切片，并記錄坐標位置
pic=cv2.imread(image_dir)
cor_list=[]
name_list=['blank','green_blank','red_blank','yellow','yellow_balnk','yellow_blue']
font = cv2.FONT_HERSHEY_SIMPLEX
for i in range(3):
 for j in range(5):
 if(i==2):
  cut_feature = T[:, 4 * j:4 * j + 5, 17:27, :]
  data = np.expand_dims(cut_feature, axis=0)
  result = model_sec.predict(data)
  print(result)
  result_data=result[0].tolist()
  #如果置信度過低，則舍棄
  # if(max(result_data)<=0.7):
  # continue
  index_num = result_data.index(max(result_data))
  name=name_list[index_num]
  cor_list = [i * 160 + 6, j * 80] # 每個切片數(shù)據(jù)，映射到原圖上，檢測框?qū)?yīng)的左上角坐標
  x=cor_list[0]
  y=cor_list[1]
  cv2.rectangle(pic, (160 * i + 6, 80 * j), ((i + 1) * 160 + 6, 80 * (j+ 1)), (0, 255, 0), 2)
  cv2.putText(pic, name, (x + 40, y + 40), font, 0.5, (0, 0, 255), 1)
 else:
  cut_feature = T[:, 4 * j:4 * j + 5, 9 * i:9 * i + 10, :]
  data = np.expand_dims(cut_feature, axis=0)
  result = model_sec.predict(data)
  print(result)
  result_data = result[0].tolist()
  #如果置信度過低，則舍棄
  # if (max(result_data) <= 0.7):
  # continue
  index_num = result_data.index(max(result_data))
  name = name_list[index_num]
  cor_list = [i * 160 + 6, j * 80] # 每個切片數(shù)據(jù)，映射到原圖上，檢測框?qū)?yīng)的左上角坐標
  x = cor_list[0]
  y = cor_list[1]
  cv2.rectangle(pic, (160 * i + 6, 80 * j), ((i + 1) * 160 + 6, 80 * (j + 1)), (0, 255, 0), 2)
  cv2.putText(pic, name, (x + 40, y + 40), font, 0.5, (0, 0, 255), 1)

cv2.imshow('pic',pic)
cv2.waitKey(0)
cv2.destroyAllWindows()
# data= np.expand_dims(f_1, axis=0)
# result=model_sec.predict(data)
# print(result)
#第二個模型可以完全預(yù)測，沒有問題

補充知識：加載訓(xùn)練好的模型參數(shù)，但是權(quán)重一直變化