參考caffe做迴歸問題，這裡主要是改caffe的迴歸程式碼為keras程式碼，順便熟悉keras的流程。caffe的訓練資料主要是通過這篇部落格網址得到訓練資料。這裡主要是通過keras書寫網路，同時做相應的預測。

首先看一下caffe的網路結構圖：

這裡按照這個網路搭建keras的網路：

# coding=utf-8
from keras.layers import merge
from keras.layers.convolutional import Convolution2D, MaxPooling2D,ZeroPadding2D,AveragePooling2D
from keras.layers.core import Dense, Activation,Flatten,Dropout
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras.layers import Input
from keras.utils import plot_model
from keras import losses
import h5py
import keras
#輸入資料的塊大小
inp = Input(shape=(1,100,100))
#網路結構
out = Convolution2D(96,kernel_size=(5,5),strides=(2,2))(inp)
out = Activation('relu')(out)
out = MaxPooling2D(pool_size=(3,3),strides=2)(out)
out = ZeroPadding2D(padding=(2,2))(out)


out = Convolution2D(filters=96,kernel_size=(3,3),strides=(1,1))(out)
out = Activation('relu')(out)
out = MaxPooling2D(pool_size=(3,3),strides=2)(out)
out = ZeroPadding2D(padding=(1,1))(out)


out = Convolution2D(filters=128,kernel_size=(3,3),strides=(1,1))(out)
out = Activation('relu')(out)
out = MaxPooling2D(pool_size=(3,3),strides=2)(out)

out = Flatten()(out)
out = Dense(units=192,activation='softmax')(out)
out = Activation('relu')(out)
out = Dropout(rate=0.5)(out)
out = Dense(units=1,activation='sigmoid')(out)
#net
model = Model(inp,out)

#做迴歸的問題採用mse損失函式
model.compile(loss=losses.mean_squared_error, optimizer='sgd')
#載入需要進行訓練的資料

x_data = keras.utils.io_utils.HDF5Matrix('train.h5', 'data')
y_data = keras.utils.io_utils.HDF5Matrix('train.h5', 'score')
#shuffle the data
model.fit(x_data, y_data,validation_split=0.3,epochs=10, batch_size=16)
model.save('regression_model.h5')
plot_model(model, to_file='model2.png',show_shapes=True)

 

然後用來預測的程式碼如下：

# coding=utf-8
import numpy as np
from keras.layers import Dense
from keras.models import Model
from keras.preprocessing import image

from keras.models import load_model
model = load_model('regression_model.h5')
# model.load_weights('weight.h5')
print model.get_config()
image_path = './samples/00001_0.028184228311.jpg'
# 載入影象
img = image.load_img(image_path,  grayscale=True, target_size=(100, 100))
# 影象預處理
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)

# 對影象進行分類
preds = model.predict(x)
# 輸出預測概率
print 'Predicted:', preds