Last active
August 25, 2021 00:48
-
-
Save higuma/dbcd006546eb844c01e5102b4d0bcc93 to your computer and use it in GitHub Desktop.
MNIST database reader with Golang
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
// MNIST database reader | |
// http://yann.lecun.com/exdb/mnist/ | |
package mnist | |
import ( | |
"fmt" | |
"os" | |
"path/filepath" | |
) | |
const ( | |
TrainImagesFile = "train-images-idx3-ubyte" | |
TrainLabelsFile = "train-labels-idx1-ubyte" | |
TestImagesFile = "t10k-images-idx3-ubyte" | |
TestLabelsFile = "t10k-labels-idx1-ubyte" | |
labelsFileMagic = 0x00000801 | |
imagesFileMagic = 0x00000803 | |
msgInvalidFormat = "Invalid format: %s" | |
msgSizeUnmatch = "Data size does not match: %s %s" | |
) | |
func fileError(f *os.File) error { | |
return fmt.Errorf(msgInvalidFormat, f.Name()) | |
} | |
// internal: Read 4 bytes and convert to big endian integer | |
func readInt32(f *os.File) (int, error) { | |
buf := make([]byte, 4) | |
n, e := f.Read(buf) | |
switch { | |
case e != nil: | |
return 0, e | |
case n != 4: | |
return 0, fileError(f) | |
} | |
v := 0 | |
for _, x := range buf { | |
v = v * 256 + int(x) | |
} | |
return v, nil | |
} | |
// internal: raw image data | |
type imageData struct { | |
N int | |
W int | |
H int | |
Data []uint8 | |
} | |
func readImagesFile(path string) (*imageData, error) { | |
f, e := os.Open(path) | |
if e != nil { | |
return nil, e | |
} | |
defer f.Close() | |
magic, e := readInt32(f) | |
if e != nil || magic != imagesFileMagic { | |
return nil, fileError(f) | |
} | |
n, e := readInt32(f) | |
if e != nil { | |
return nil, fileError(f) | |
} | |
w, e := readInt32(f) | |
if e != nil { | |
return nil, fileError(f) | |
} | |
h, e := readInt32(f) | |
if e != nil { | |
return nil, fileError(f) | |
} | |
sz := n * w * h | |
data := &imageData{n, w, h, make([]uint8, sz)} | |
len, e := f.Read(data.Data) | |
if e != nil || len != sz { | |
return nil, fileError(f) | |
} | |
return data, nil | |
} | |
// internal: raw label data | |
type labelData struct { | |
N int | |
Data []uint8 | |
} | |
func readLabelsFile(path string) (*labelData, error) { | |
f, e := os.Open(path) | |
if e != nil { | |
return nil, e | |
} | |
defer f.Close() | |
magic, e := readInt32(f) | |
if e != nil || magic != labelsFileMagic { | |
return nil, fileError(f) | |
} | |
n, e := readInt32(f) | |
if e != nil { | |
return nil, fileError(f) | |
} | |
data := &labelData{n, make([]uint8, n)} | |
len, e := f.Read(data.Data) | |
if e != nil || len != n { | |
return nil, fileError(f) | |
} | |
return data, nil | |
} | |
// Single digit+image datum | |
type DigitImage struct { | |
Digit int | |
Image [][]uint8 | |
} | |
// Data set | |
type DataSet struct { | |
N int | |
W int | |
H int | |
Data []DigitImage | |
} | |
// Database readers | |
func ReadDataSet(imagesPath, labelsPath string) (*DataSet, error) { | |
images, e := readImagesFile(imagesPath) | |
if e != nil { | |
return nil, e | |
} | |
labels, e := readLabelsFile(labelsPath) | |
if e != nil { | |
return nil, e | |
} | |
if images.N != labels.N { | |
return nil, fmt.Errorf(msgSizeUnmatch, labelsPath, imagesPath) | |
} | |
dataSet := &DataSet{N: images.N, W: images.W, H: images.H} | |
dataSet.Data = make([]DigitImage, dataSet.N) | |
rows := splitToRows(images.Data, images.N, images.H) | |
for i := 0; i < dataSet.N; i++ { | |
data := &dataSet.Data[i] | |
data.Digit = int(labels.Data[i]) | |
data.Image = rows[0:dataSet.H] | |
rows = rows[dataSet.H:] | |
} | |
return dataSet, nil | |
} | |
func ReadTrainSet(dir string) (*DataSet, error) { | |
imagesPath := filepath.Join(dir, TrainImagesFile) | |
labelsPath := filepath.Join(dir, TrainLabelsFile) | |
return ReadDataSet(imagesPath, labelsPath) | |
} | |
func ReadTestSet(dir string) (*DataSet, error) { | |
imagesPath := filepath.Join(dir, TestImagesFile) | |
labelsPath := filepath.Join(dir, TestLabelsFile) | |
return ReadDataSet(imagesPath, labelsPath) | |
} | |
func splitToRows(data []uint8, N, H int) [][]uint8 { | |
nR := N * H | |
rows := make([][]uint8, nR) | |
for i := 0; i < nR; i++ { | |
rows[i] = data[0:H] | |
data = data[H:] | |
} | |
return rows | |
} | |
// (debugging utility) | |
func PrintImage(image [][]uint8) { | |
for _, row := range image { | |
for _, pix := range row { | |
if pix == 0 { | |
fmt.Print(" ") | |
} else { | |
fmt.Printf("%X", pix / 16) | |
} | |
} | |
fmt.Println() | |
} | |
} | |
/*** | |
// SAMPLE USAGE | |
// ============ | |
import ( | |
"fmt" | |
"./mnist" | |
) | |
func printData(dataSet *mnist.DataSet, index int) { | |
data := dataSet.Data[index] | |
fmt.Println(data.Digit) // print Digit (label) | |
mnist.PrintImage(data.Image) // print Image | |
} | |
func main() { | |
dataSet, err := mnist.ReadTrainSet("./mnist") | |
// or dataSet, err := mnist.ReadTestSet("./mnist") | |
if err != nil { | |
fmt.Println(err) | |
return | |
} | |
fmt.Println(dataSet.N) // number of data | |
fmt.Println(dataSet.W) // image width [pixel] | |
fmt.Println(dataSet.H) // image height [pixel] | |
for i := 0; i < 10; i++ { | |
printData(dataSet, i) | |
} | |
printData(dataSet, dataSet.N-1) | |
} | |
***/ |
Oh, what a mistake!
As you know, previous code is harmless as long as MNIST data set is correct.
But I have corrected line 136.
Thank you very much for your suggestion.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
if images.N != images.N {
shoud be
if images.N != labels.N {