7-12

• batch_size一次喂进去的data，比如喂总数据量10000，就需要迭代20次才完成一个epoch。

        train_history = model.fit(x=train_Xs,
                                        y=train_Ys,
                                        validation_split=0,
                                        epochs=30,
                                        batch_size=500,
                                        verbose=2)

• 预处理，将图像读到list，再转array然后进行归一化。

def pre_deal(file_names,satrt_num,data_lenth,input_size_1 = input_size_1, input_size_2 = input_size_2):
    files_ = file_names[satrt_num:(satrt_num+data_lenth)]

    x_imgs = []
    y_imgs = []
    for i in range(data_lenth):
        x_img = cv2.imread('dataset/img/'+files_[i]).astype("float")
        y_img = cv2.imread('dataset/contours/'+files_[i]).astype("float")
        x_img = cv2.resize(x_img,(input_size_1,input_size_2))
        y_img = cv2.resize(y_img,(input_size_1,input_size_2))
        x_img /= 255
        y_img /= 255
        x_imgs.append(x_img)
        y_imgs.append(y_img)

    print x_imgs[2].shape,len(x_imgs)

    in_imgs = np.array(x_imgs)                    #将 list 转化为一个　array　
    out_imgs = np.array(y_imgs)

    print in_imgs.shape

    IN_imgs = in_imgs.reshape(in_imgs.shape[0],input_size_1,input_size_2,3).astype('float32')
    OUT_imgs = out_imgs.reshape(out_imgs.shape[0],input_size_1,input_size_2,3).astype('float32')

    return IN_imgs,OUT_imgs

分多次进行训练（效果并不好），因为一次不能装完所有数据，内存限制

    for num in range(1):
        print ("group :",num)
        train_Xs, train_Ys = pre_deal(train_files,num*10000,10000)
        train_history = model.fit(x=train_Xs,
                                        y=train_Ys,
                                        validation_split=0,
                                        epochs=30,
                                        batch_size=500,
                                        verbose=2)
        save_model(model, '7_12unet4.h5')