淺談keras使用預(yù)訓(xùn)練模型vgg16分類,損失和準(zhǔn)確度不變

更新時(shí)間：2020年07月02日 08:39:13 作者：我的九八卡

這篇文章主要介紹了淺談keras使用預(yù)訓(xùn)練模型vgg16分類,損失和準(zhǔn)確度不變，具有很好的參考價(jià)值，希望對(duì)大家有所幫助。一起跟隨小編小編過來看看吧

問題keras使用預(yù)訓(xùn)練模型vgg16分類，損失和準(zhǔn)確度不變。

細(xì)節(jié)：使用keras訓(xùn)練一個(gè)兩類數(shù)據(jù)，正負(fù)比例1：3，在vgg16后添加了幾個(gè)全鏈接并初始化了。并且對(duì)所有層都允許訓(xùn)練。

但是準(zhǔn)確度一直是0.75.

數(shù)據(jù)預(yù)先處理已經(jīng)檢查過格式正確

再將模型中relu改成sigmoid就正常了。

數(shù)據(jù)處理程序

import os
import pickle
import numpy as np
 
import DataFile
import SelectiveSearch
import Generator
import IoU
import Model_CRNN_VGG16
 
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint
 
def data_generator(gen1,gen0):
 while True:
 data_pos = next(gen1)
 data_neg = next(gen0)
 ret_X = np.vstack((data_pos[0],data_neg[0]))
 ret_y = np.vstack((data_pos[1],data_neg[1]))
 
 index = np.arange(ret_y.shape[0])
 np.random.shuffle(index)
 
 ret_X = ret_X[index, :, :, :] # X_train是訓(xùn)練集，y_train是訓(xùn)練標(biāo)簽
 ret_y = ret_y[index]
 yield ret_X,ret_y
 
if __name__ == "__main__":
 type = "train"
 
 # 數(shù)據(jù)生成器,每個(gè)mini-batch包含32個(gè)正樣本(屬于VOC 20個(gè)類別),96個(gè)負(fù)樣本(background)
 RESIZE = (224, 224)
 path = "category_images"
 categories = os.listdir(path)
 categories.append('background')
 print(categories)
 
 train_1_datagen = ImageDataGenerator(
 rescale=1.0/255,
 #shear_range=0.2,
 #zoom_range=0.2,
 horizontal_flip=True)
 
 train_1_generator = train_1_datagen.flow_from_directory(
 'category_images',
 target_size=RESIZE,
 batch_size=32,
 classes = categories)
 
 train_0_datagen = ImageDataGenerator(
 rescale=1.0 / 255,
 #shear_range=0.2,
 #zoom_range=0.2,
 horizontal_flip=True)
 
 train_0_generator = train_0_datagen.flow_from_directory(
 'category_background',
 target_size=RESIZE,
 batch_size=32*3,
 classes=categories)
 
 generator = data_generator(train_1_generator,train_0_generator)
 
 # 創(chuàng)建模型
 model = Model_CRNN_VGG16.CRNN_Model(input_shape=(*RESIZE,3))
 cnn = model.CNN(len(categories))
 if os.path.exists('weights-cnn.hdf5'):
 cnn.load_weights('weights-cnn.hdf5')
 if type == "train":
 checkpoint = ModelCheckpoint('weights-cnn.hdf5',save_weights_only=True)
 cnn.fit_generator(generator = generator,steps_per_epoch=200,epochs=1000,callbacks=[checkpoint])
 else:
 img = next(generator)[0]
 result = cnn.predict(img)
 print(result)
 
 # 訓(xùn)練SVM
 # 非極大值抑制
 # 預(yù)測(cè)

模型程序：

from keras.applications.vgg16 import VGG16
from keras.layers import *
from keras.models import Model
from keras.optimizers import SGD,Adam
 
class CRNN_Model():
 def __init__(self,input_shape,trainable=True):
 vgg16 = VGG16(include_top=False,weights="imagenet", input_shape=input_shape)
 for layer in vgg16.layers:
  layer.trainable = trainable
 self.base_model = vgg16
 
 def CNN(self,classes):
 img_input = self.base_model.input
 x = self.base_model.get_layer('block5_conv3').output
 
 x = Flatten(name='crnn_flatten')(x)
 
 x = Dense(512,activation='relu', kernel_initializer='he_normal', name='crnn_fc1')(x)
 x = Dense(512,activation='relu', kernel_initializer='he_normal',name='crnn_fc2')(x)
 x = Dense(classes, activation='softmax', kernel_initializer='he_normal', name='crnn_predictions')(x)
 
 model = Model(img_input,x)
 
 sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
 adam = Adam()
 model.compile(optimizer=adam,
   loss='categorical_crossentropy',
   metrics=['accuracy'])
 
 model.summary()
 return model
if __name__ == "__main__":
 pass

補(bǔ)充知識(shí)：val_acc一直不變

val_loss一直不變的原因

之前用keras編寫了LSTM模型，做圖片分類，自己劃分了測(cè)試集和訓(xùn)練集，但是得到的結(jié)果是每個(gè)epoch訓(xùn)練的準(zhǔn)確率都不變。