手寫體識別例程的mnist數據集包含了4個文件，分別爲train-images-idx3-ubyte、train-labels-idx1-ubyte、t10k-images-idx3-ubyte、t10k-labels-idx1-ubyte。這四個文件分別是訓練集的圖片文件、訓練集的標籤文件、測試集的圖片文件、測試集的標籤文件，其中，訓練集中有60000張圖片，測試集有10000張圖片。這4個文件是二進制文件，那麼這麼圖片是如何在這些文件存儲的呢？
我們已測試集的文件爲例進行分析，即t10k-images-idx3-ubyte和t10k-labels-idx1-ubyte。文件的定義如下
t10k-images-idx3-ubyte定義
| 地址偏移 | 數據類型 | 取值 | 描述 |
| 0000 | uint32 | 2051 | 魔數（大端存儲） |
| 0004 | uint32 | 10000 | 文件包含的條目總數 |
| 0008 | uint32 | 28 | 行數 |
| 000c | uint32 | 28 | 列數 |
| 0010 | uint8 | ？ | 像素灰度值（0~255） |
| 0011 | uint8 | ？ | 像素灰度值（0~255） |
| ... | ... | ... | ... |
t10k-labels-idx1-ubyte定義
| 地址偏移 | 數據類型 | 取值 | 描述 |
| 0000 | uint32 | 2049 | 魔數（大端存儲） |
| 0004 | uint32 | 10000 | 文件包含的條目總數 |
| 0008 | uint8 | ? | 標籤值（0~9） |
| 000c | uint8 | ? | 標籤值（0~9） |
| ... | ... | ... | ... |
訓練集的兩個文件與測試集的文件定義是完全一樣的。
從上面的表格我們可以看出來t10k-images-idx3-ubyte中存儲的是10000張圖片，每個圖片是28X28的像素分辨率，每個像素佔一個字節數據，數據描述的是像素的灰度值，0表示白色，255表示黑色。
t10k-labels-idx1-ubyte中存儲的是每張圖片對應的標籤值，即0~9的數字值，每張圖片與每個標籤是一一對應的。爲了詳細分析，我們看一下這兩個文件的內部取值，由於是二進制文件，我們可以用hexdump軟件打開，如果計算機中沒有這個軟件，可以用下面的命令進行安裝
sudo apt-get install hexdump
用下面的命令打開文件
hexdump -Cv t10k-images-idx3-ubyte | more
打開文件後如下所示
00000000 00 00 08 03 00 00 27 10 00 00 00 1c 00 00 00 1c |......'.........|
00000010 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000020 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000030 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000040 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000050 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000060 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000070 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000080 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000090 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000000a0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000000b0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000000c0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000000d0 00 00 00 00 00 00 00 00 00 00 54 b9 9f 97 3c 24 |..........T...<$|
000000e0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000000f0 00 00 00 00 00 00 de fe fe fe fe f1 c6 c6 c6 c6 |................|
00000100 c6 c6 c6 c6 aa 34 00 00 00 00 00 00 00 00 00 00 |.....4..........|
00000110 00 00 43 72 48 72 a3 e3 fe e1 fe fe fe fa e5 fe |..CrHr..........|
00000120 fe 8c 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000130 00 00 00 11 42 0e 43 43 43 3b 15 ec fe 6a 00 00 |....B.CCC;...j..|
00000140 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000150 00 00 00 00 00 00 53 fd d1 12 00 00 00 00 00 00 |......S.........|
00000160 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000170 00 16 e9 ff 53 00 00 00 00 00 00 00 00 00 00 00 |....S...........|
00000180 00 00 00 00 00 00 00 00 00 00 00 00 00 81 fe ee |................|
00000190 2c 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |,...............|
000001a0 00 00 00 00 00 00 00 00 3b f9 fe 3e 00 00 00 00 |........;..>....|
000001b0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000001c0 00 00 00 00 85 fe bb 05 00 00 00 00 00 00 00 00 |................|
000001d0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 09 |................|
000001e0 cd f8 3a 00 00 00 00 00 00 00 00 00 00 00 00 00 |..:.............|
000001f0 00 00 00 00 00 00 00 00 00 00 00 7e fe b6 00 00 |...........~....|
00000200 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000210 00 00 00 00 00 00 4b fb f0 39 00 00 00 00 00 00 |......K..9......|
00000220 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000230 00 13 dd fe a6 00 00 00 00 00 00 00 00 00 00 00 |................|
00000240 00 00 00 00 00 00 00 00 00 00 00 00 03 cb fe db |................|
00000250 23 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |#...............|
00000260 00 00 00 00 00 00 00 00 26 fe fe 4d 00 00 00 00 |........&..M....|
00000270 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000280 00 00 00 1f e0 fe 73 01 00 00 00 00 00 00 00 00 |......s.........|
00000290 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 85 |................|
000002a0 fe fe 34 00 00 00 00 00 00 00 00 00 00 00 00 00 |..4.............|
000002b0 00 00 00 00 00 00 00 00 00 00 3d f2 fe fe 34 00 |..........=...4.|
000002c0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000002d0 00 00 00 00 00 00 79 fe fe db 28 00 00 00 00 00 |......y...(.....|
000002e0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000002f0 00 00 79 fe cf 12 00 00 00 00 00 00 00 00 00 00 |..y.............|
00000300 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000310 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000320 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000330 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000340 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000350 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000360 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000370 00 00 00 00 00 00 00 00 00 00 00 00 00 00 74 7d |..............t}|
00000380 ab ff ff 96 5d 00 00 00 00 00 00 00 00 00 00 00 |....]...........|
00000390 00 00 00 00 00 00 00 00 00 a9 fd fd fd fd fd fd |................|
000003a0 da 1e 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
000003b0 00 00 00 00 a9 fd fd fd d5 8e b0 fd fd 7a 00 00 |.............z..|
000003c0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 34 |...............4|
對於多字節的數據，是按大端模式存儲的，我們可以根據上面表格的定義對文件內容進行分析。
首先我們看前4個字節 00000803，它轉換爲十進制爲2051，正是表格中定義的魔數。
接下來4個字節是00002710，它轉換爲十進制爲10000，正是表格中定義的條目數。
再接下來就是兩個4字節的0000001c，都是十進制的28，分別是圖片像素的行數和列數。
再下面的就是圖片的像素值，28X28=784個字節爲一張圖片，我們從上面的數據中可以看到00比較多，代表白色的背景，其它值則代表有顏色的像素，我們把這些784個像素值描繪在28X28的圖片上，如下圖所示。
很明顯，這是一個數字7，文件中其它的數據是剩下的9999張和上圖類似的圖片的數據，有興趣可以描幾個看看。至此t10k-images-idx3-ubyte文件已經淋漓盡致的展現在我們的面前。
接下來，我們再來看看t10k-labels-idx1-ubyte文件，它存儲的是標籤值，用下面的命令打開文件
hexdump -Cv t10k-labels-idx1-ubyte | more
打開後如下圖所示。
00000000 00 00 08 01 00 00 27 10 07 02 01 00 04 01 04 09 |......'.........|
00000010 05 09 00 06 09 00 01 05 09 07 03 04 09 06 06 05 |................|
00000020 04 00 07 04 00 01 03 01 03 04 07 02 07 01 02 01 |................|
00000030 01 07 04 02 03 05 01 02 04 04 06 03 05 05 06 00 |................|
00000040 04 01 09 05 07 08 09 03 07 04 06 04 03 00 07 00 |................|
00000050 02 09 01 07 03 02 09 07 07 06 02 07 08 04 07 03 |................|
00000060 06 01 03 06 09 03 01 04 01 07 06 09 06 00 05 04 |................|
00000070 09 09 02 01 09 04 08 07 03 09 07 04 04 04 09 02 |................|
00000080 05 04 07 06 07 09 00 05 08 05 06 06 05 07 08 01 |................|
00000090 00 01 06 04 06 07 03 01 07 01 08 02 00 02 09 09 |................|
000000a0 05 05 01 05 06 00 03 04 04 06 05 04 06 05 04 05 |................|
000000b0 01 04 04 07 02 03 02 07 01 08 01 08 01 08 05 00 |................|
000000c0 08 09 02 05 00 01 01 01 00 09 00 03 01 06 04 02 |................|
000000d0 03 06 01 01 01 03 09 05 02 09 04 05 09 03 09 00 |................|
000000e0 03 06 05 05 07 02 02 07 01 02 08 04 01 07 03 03 |................|
000000f0 08 08 07 09 02 02 04 01 05 09 08 07 02 03 00 04 |................|
00000100 04 02 04 01 09 05 07 07 02 08 02 06 08 05 07 07 |................|
00000110 09 01 08 01 08 00 03 00 01 09 09 04 01 08 02 01 |................|
00000120 02 09 07 05 09 02 06 04 01 05 08 02 09 02 00 04 |................|
00000130 00 00 02 08 04 07 01 02 04 00 02 07 04 03 03 00 |................|
00000140 00 03 01 09 06 05 02 05 09 02 09 03 00 04 02 00 |................|
00000150 07 01 01 02 01 05 03 03 09 07 08 06 05 06 01 03 |................|
00000160 08 01 00 05 01 03 01 05 05 06 01 08 05 01 07 09 |................|
00000170 04 06 02 02 05 00 06 05 06 03 07 02 00 08 08 05 |................|
00000180 04 01 01 04 00 03 03 07 06 01 06 02 01 09 02 08 |................|
00000190 06 01 09 05 02 05 04 04 02 08 03 08 02 04 05 00 |................|
000001a0 03 01 07 07 05 07 09 07 01 09 02 01 04 02 09 02 |................|
000001b0 00 04 09 01 04 08 01 08 04 05 09 08 08 03 07 06 |................|
000001c0 00 00 03 00 02 06 06 04 09 03 03 03 02 03 09 01 |................|
000001d0 02 06 08 00 05 06 06 06 03 08 08 02 07 05 08 09 |................|
000001e0 06 01 08 04 01 02 05 09 01 09 07 05 04 00 08 09 |................|
000001f0 09 01 00 05 02 03 07 08 09 04 00 06 03 09 05 02 |................|
00000200 01 03 01 03 06 05 07 04 02 02 06 03 02 06 05 04 |................|
00000210 08 09 07 01 03 00 03 08 03 01 09 03 04 04 06 04 |................|
00000220 02 01 08 02 05 04 08 08 04 00 00 02 03 02 07 07 |................|
00000230 00 08 07 04 04 07 09 06 09 00 09 08 00 04 06 00 |................|
00000240 06 03 05 04 08 03 03 09 03 03 03 07 08 00 08 02 |................|
00000250 01 07 00 06 05 04 03 08 00 09 06 03 08 00 09 09 |................|
00000260 06 08 06 08 05 07 08 06 00 02 04 00 02 02 03 01 |................|
00000270 09 07 05 01 00 08 04 06 02 06 07 09 03 02 09 08 |................|
00000280 02 02 09 02 07 03 05 09 01 08 00 02 00 05 02 01 |................|
00000290 03 07 06 07 01 02 05 08 00 03 07 02 04 00 09 01 |................|
我們依然對照這上面的表格來研究，
前4個字節爲00000801，它轉換爲十進制爲2049，正是表格中定義的魔數。
接下來4個字節爲00002710，它轉換爲十進制爲10000，正是表格中定義的條目數。
再後面就是標籤值，我們看到第一個值就是07，正是我們從t10k-images-idx3-ubyte文件中描出來的圖片的取值。這個文件中剩下的就是9999個0~9的標籤值。
對於訓練集的train-images-idx3-ubyte、train-labels-idx1-ubyte文件，我們這類就不做過多解釋了，定義和測試集的文件是完全一樣的。