cam_traincontinue.py

# encoding: utf-8
#used for VOC2007   double output    adding an additional conv layers
from keras.preprocessing.image import ImageDataGenerator
from keras.applications.vgg16 import preprocess_input,VGG16
from keras.layers import GlobalAveragePooling2D,Dense,Conv2D
from keras.models import Model
from keras.utils.vis_utils import plot_model
from keras.optimizers import Adagrad,Adam
import numpy as np
from keras.models import load_model  


# 数据准备
train_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,# ((x/255)-0.5)*2  归一化到±1之间
    rotation_range=30,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
)
val_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,
    rotation_range=30,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
)




train_generator = train_datagen.flow_from_directory(directory='./CUB_200_2011/CUB_200_2011/train',
                                  target_size=(224,224),#VGGNet16规定大小
                                  batch_size=32)            
val_generator = val_datagen.flow_from_directory(directory='./CUB_200_2011/CUB_200_2011/validation',
                                target_size=(224,224),
                                batch_size=32,
                                save_to_dir='./image_generator_save')   



# 构建基础模型
model =load_model('./mid_model.h5')





def setup_to_transfer_learning(model):

    GAP_LAYER_ = 17 # max_pooling_2d_2
    for layer in model.layers[:GAP_LAYER_+1]:
        layer.trainable = False
    for layer in model.layers[GAP_LAYER_+1:]:
        layer.trainable = True
    model.compile(optimizer=Adam(lr=0.0001),loss='categorical_crossentropy',metrics=['accuracy'])

def setup_to_fine_tune(model):
    GAP_LAYER = 0 # max_pooling_2d_2  13
    for layer in model.layers[:GAP_LAYER+1]:
        layer.trainable = False
    for layer in model.layers[GAP_LAYER+1:]:
        layer.trainable = True
    model.compile(optimizer=Adam(lr=0.00005),loss='categorical_crossentropy',metrics=['accuracy'])



# setup_to_transfer_learning(model)
# history_tl = model.fit_generator(generator=train_generator,
#                     steps_per_epoch=200,#800                     
#                     epochs=4,#2
#                     validation_data=val_generator,
#                     validation_steps=6,#12
#                     class_weight='auto'
#                     )

setup_to_fine_tune(model)
history_ft = model.fit_generator(generator=train_generator,
                                 steps_per_epoch=200,                
                                 epochs=6,
                                 validation_data=val_generator,
                                 validation_steps=6,
                                 class_weight='auto')

 


#保存中间模型以供继续训练
model.save('./mid_model.h5')

#提取并存储全连接层参数
w=model.layers[-1].get_weights()
w=w[0]
w=np.array(w)
np.save("./fcn_w.npy",w)

#改变model的结构，将feature map作为输出
model_save = Model(inputs=model.input,outputs=[model.layers[-3].output,model.output])
#保存模型
model_save.save('./cam_model.h5')