MNIST数据读取-numpy版本

MNIST数据集地址：http://yann.lecun.com/exdb/mnist/

格式解释

范例

超简单的numpy读取代码：

import numpy as np

train_images_path = 'train-images.idx3-ubyte'
train_labels_path = 'train-labels.idx1-ubyte'

test_images_path = 't10k-images.idx3-ubyte'
test_labels_path = 't10k-labels.idx1-ubyte'

def get_dataset(data_path):
    train_images = read_idx_file(data_path + train_images_path)
    train_labels = read_idx_file(data_path + train_labels_path)
    
    test_images = read_idx_file(data_path + test_images_path)
    test_labels = read_idx_file(data_path + test_labels_path)

    return train_images, train_labels, test_images, test_labels

def read_idx_file(file_name):
    byte_data = np.fromfile(file_name, dtype=np.uint8)

    file_type_id = byte_data[2]
    items = byte_data[3]
    data_shape = np.zeros((items, ), dtype=np.int32)

    for i in range(items):
        dim_t = byte_data[4 + i * 4 : 4 + (i + 1) * 4]
        dim_t.dtype = np.uint32
        dim_t.byteswap(inplace=True) # convert big edian
        data_shape[i] = dim_t[0]

    data_offset = items * 4 + 4
    data = byte_data[data_offset:].reshape(data_shape)
    return data

验证结果

import matplotlib.pyplot as plt

from read_data import get_dataset

data_path = r'G:\dataset\MNIST' + '\\'

train_images, train_labels, test_images, test_labels = get_dataset(data_path)

# evaluate the codes
img0 = train_images[0]
plt.imshow(img0)
plt.show()