一、官方文档（中文）

二、官方文档（英文）

三、为什么要用npy文件保存文件

• A LOT faster, also notice that we didn’t need to reshape the data since that information was contained in the .npy file.（速度更快）
• Another “minor” feature of using .npy files is the reduced storage the file occupies. In this case it’s more than a 50% reduction in size. This can wary a lot though but in general the .npy files are more storage friendly.（占用内存更少）

四、读取和保存

（1）加载npy文件，并将npy文件写入一个txt文件

import numpy as np
test=np.load('./bvlc_alexnet.npy',encoding = "latin1")  #加载文件
doc = open('1.txt', 'a')  #打开一个存储文件，并依次写入
print(test, file=doc)  #将打印内容写入文件中

（2）保存npy文件

#显示字典
print(train_dataset.class_to_idx)
idx_to_labels = {y:x for x,y in train_dataset.class_to_idx.items()}#获取dict:train_dataset.class_to_idx的keys和values
print(idx_to_labels)

np.save('idx_to_labels.npy',idx_to_labels)
np.save('labels_to index.npy',train_dataset.class_to_idx)

（3）结构解析

import numpy as np
from numpy import *  #使用numpy的属性且不需要在前面加上numpy
import tensorflow as tf
 
#模型文件（.npy）部分内容如下：由一个字典组成，字典中的每一个键对应一层网络模型参数。（包括权重w和偏置b）
a = {'conv1':[array([[1,2],[3,4]],dtype=float32),array([5,6],dtype=float32)],'conv2':[array([[1,2],[3,4]],dtype=float32),array([5,6],dtype=float32)]}
 
conv1_w = a['conv1'][0]
conv1_b = a['conv1'][1]
conv2_w = a['conv2'][0]
conv2_b = a['conv2'][1]
 
print(conv1_w)
print(tf.Variable(conv1_w))
print(conv1_b)
print(tf.Variable(conv1_b))