Python 批量将raw数据或图片一一存成.npy文件

深度学习神经网络训练读入的都是numpy数组，一般直接读取图片在做预处理送入模型训练时很浪费时间的，我们可以提前将数据转成.npy格式的文件，并且数据的预处理也可以在这里完成。那么，将预处理好的数据直接送入网络就可以节省很多的时间了。

关于下面有关raw数据读取看不懂的可以查看我的另外一篇专门敢于读raw数据转png图片的博客，里面有解释。

# -*- coding:utf8 -*-

import numpy as np
import cv2
import os

def oneimg2npy_(root_dir, out_path):
    filelist = os.listdir(root_dir)

    "读取raw数据时使用"
    # 文件排序
    filelist.sort(key=lambda x: int(x.split('_')[5]))
    for file in filelist:
        bayer = np.zeros(shape=2073600, dtype='uint16')
        file_path = os.path.join(root_dir, file)

        with open(file_path, "rb") as f:
            for i in range(0, len(f.read()), 2):
                f.seek(i)
                raw = f.read(2)
                a1 = int((raw[0] / 16) % 16)
                a2 = int(raw[0] % 16)
                a3 = int((raw[1] / 16) % 16)
                a4 = int(raw[1] % 16)
                value = a3 * 256 + a4 * 16 + a1 * 1
                bayer[int(i / 2)] = value

        bayer = bayer.reshape(1080, 1920)

        height, width = bayer.shape[0], bayer.shape[1]
        # print(height, width)
        h = height // 2
        w = width // 2

        R = np.zeros(shape=(h, w), dtype='uint16', order='C')
        Gr = np.zeros(shape=(h, w), dtype='uint16', order='C')
        Gb = np.zeros(shape=(h, w), dtype='uint16', order='C')
        B = np.zeros(shape=(h, w), dtype='uint16', order='C')

        for x in range(height):
            for y in range(0, width, 2):
                if x % 2 == 0:
                    R[int(x / 2)][int(y / 2)] = bayer[x][y]
                    Gr[int(x / 2)][int(y / 2)] = bayer[x][y + 1]
                elif x % 2 == 1:
                    Gb[int(x / 2)][int(y / 2)] = bayer[x][y]
                    B[int(x / 2)][int(y / 2)] = bayer[x][y + 1]

        out = np.stack((R, Gr, Gb, B))
        out = out.astype(np.float32)

        if not os.path.exists(out_path):
            os.makedirs(out_path)
        np.save(os.path.join(out_path, './{}.npy'.format(file.split(".")[0])), out)
        print("saved: {}".format(file))

        f.close()
    print("raw finished")

    "读取rgb数据时使用"
    filelist.sort(key=lambda x: int((x.split('_')[5]))
    for file in filelist:
    	file_path = os.path.join(root_dir, file)
    	label = cv2.imread(file_path)
        # label = cv2.resize(label, (960, 540), interpolation=cv2.INTER_NEAREST).astype(np.float32)
        label = np.transpose(label, (2, 0, 1))
        if not os.path.exists(out_path):
        	os.makedirs(out_path)
        np.save(os.path.join(out_path, './{}.npy'.format(file.split(".")[0])), label)
    print("rgb finished")

if __name__ == '__main__':
    _ = oneimg2npy_(root_dir='', out_path='')