Cyrille Nofficial
456f327788
* add flip-image option * add command to list models * add option to override image size when training is launched * add option to disable aws spot instance
439 lines
12 KiB
Go
439 lines
12 KiB
Go
// Copyright 2012 The Go Authors. All rights reserved.
|
|
// Use of this source code is governed by a BSD-style
|
|
// license that can be found in the LICENSE file.
|
|
|
|
package tiff
|
|
|
|
import (
|
|
"bytes"
|
|
"compress/zlib"
|
|
"encoding/binary"
|
|
"image"
|
|
"io"
|
|
"sort"
|
|
)
|
|
|
|
// The TIFF format allows to choose the order of the different elements freely.
|
|
// The basic structure of a TIFF file written by this package is:
|
|
//
|
|
// 1. Header (8 bytes).
|
|
// 2. Image data.
|
|
// 3. Image File Directory (IFD).
|
|
// 4. "Pointer area" for larger entries in the IFD.
|
|
|
|
// We only write little-endian TIFF files.
|
|
var enc = binary.LittleEndian
|
|
|
|
// An ifdEntry is a single entry in an Image File Directory.
|
|
// A value of type dtRational is composed of two 32-bit values,
|
|
// thus data contains two uints (numerator and denominator) for a single number.
|
|
type ifdEntry struct {
|
|
tag int
|
|
datatype int
|
|
data []uint32
|
|
}
|
|
|
|
func (e ifdEntry) putData(p []byte) {
|
|
for _, d := range e.data {
|
|
switch e.datatype {
|
|
case dtByte, dtASCII:
|
|
p[0] = byte(d)
|
|
p = p[1:]
|
|
case dtShort:
|
|
enc.PutUint16(p, uint16(d))
|
|
p = p[2:]
|
|
case dtLong, dtRational:
|
|
enc.PutUint32(p, uint32(d))
|
|
p = p[4:]
|
|
}
|
|
}
|
|
}
|
|
|
|
type byTag []ifdEntry
|
|
|
|
func (d byTag) Len() int { return len(d) }
|
|
func (d byTag) Less(i, j int) bool { return d[i].tag < d[j].tag }
|
|
func (d byTag) Swap(i, j int) { d[i], d[j] = d[j], d[i] }
|
|
|
|
func encodeGray(w io.Writer, pix []uint8, dx, dy, stride int, predictor bool) error {
|
|
if !predictor {
|
|
return writePix(w, pix, dy, dx, stride)
|
|
}
|
|
buf := make([]byte, dx)
|
|
for y := 0; y < dy; y++ {
|
|
min := y*stride + 0
|
|
max := y*stride + dx
|
|
off := 0
|
|
var v0 uint8
|
|
for i := min; i < max; i++ {
|
|
v1 := pix[i]
|
|
buf[off] = v1 - v0
|
|
v0 = v1
|
|
off++
|
|
}
|
|
if _, err := w.Write(buf); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func encodeGray16(w io.Writer, pix []uint8, dx, dy, stride int, predictor bool) error {
|
|
buf := make([]byte, dx*2)
|
|
for y := 0; y < dy; y++ {
|
|
min := y*stride + 0
|
|
max := y*stride + dx*2
|
|
off := 0
|
|
var v0 uint16
|
|
for i := min; i < max; i += 2 {
|
|
// An image.Gray16's Pix is in big-endian order.
|
|
v1 := uint16(pix[i])<<8 | uint16(pix[i+1])
|
|
if predictor {
|
|
v0, v1 = v1, v1-v0
|
|
}
|
|
// We only write little-endian TIFF files.
|
|
buf[off+0] = byte(v1)
|
|
buf[off+1] = byte(v1 >> 8)
|
|
off += 2
|
|
}
|
|
if _, err := w.Write(buf); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func encodeRGBA(w io.Writer, pix []uint8, dx, dy, stride int, predictor bool) error {
|
|
if !predictor {
|
|
return writePix(w, pix, dy, dx*4, stride)
|
|
}
|
|
buf := make([]byte, dx*4)
|
|
for y := 0; y < dy; y++ {
|
|
min := y*stride + 0
|
|
max := y*stride + dx*4
|
|
off := 0
|
|
var r0, g0, b0, a0 uint8
|
|
for i := min; i < max; i += 4 {
|
|
r1, g1, b1, a1 := pix[i+0], pix[i+1], pix[i+2], pix[i+3]
|
|
buf[off+0] = r1 - r0
|
|
buf[off+1] = g1 - g0
|
|
buf[off+2] = b1 - b0
|
|
buf[off+3] = a1 - a0
|
|
off += 4
|
|
r0, g0, b0, a0 = r1, g1, b1, a1
|
|
}
|
|
if _, err := w.Write(buf); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func encodeRGBA64(w io.Writer, pix []uint8, dx, dy, stride int, predictor bool) error {
|
|
buf := make([]byte, dx*8)
|
|
for y := 0; y < dy; y++ {
|
|
min := y*stride + 0
|
|
max := y*stride + dx*8
|
|
off := 0
|
|
var r0, g0, b0, a0 uint16
|
|
for i := min; i < max; i += 8 {
|
|
// An image.RGBA64's Pix is in big-endian order.
|
|
r1 := uint16(pix[i+0])<<8 | uint16(pix[i+1])
|
|
g1 := uint16(pix[i+2])<<8 | uint16(pix[i+3])
|
|
b1 := uint16(pix[i+4])<<8 | uint16(pix[i+5])
|
|
a1 := uint16(pix[i+6])<<8 | uint16(pix[i+7])
|
|
if predictor {
|
|
r0, r1 = r1, r1-r0
|
|
g0, g1 = g1, g1-g0
|
|
b0, b1 = b1, b1-b0
|
|
a0, a1 = a1, a1-a0
|
|
}
|
|
// We only write little-endian TIFF files.
|
|
buf[off+0] = byte(r1)
|
|
buf[off+1] = byte(r1 >> 8)
|
|
buf[off+2] = byte(g1)
|
|
buf[off+3] = byte(g1 >> 8)
|
|
buf[off+4] = byte(b1)
|
|
buf[off+5] = byte(b1 >> 8)
|
|
buf[off+6] = byte(a1)
|
|
buf[off+7] = byte(a1 >> 8)
|
|
off += 8
|
|
}
|
|
if _, err := w.Write(buf); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func encode(w io.Writer, m image.Image, predictor bool) error {
|
|
bounds := m.Bounds()
|
|
buf := make([]byte, 4*bounds.Dx())
|
|
for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
|
|
off := 0
|
|
if predictor {
|
|
var r0, g0, b0, a0 uint8
|
|
for x := bounds.Min.X; x < bounds.Max.X; x++ {
|
|
r, g, b, a := m.At(x, y).RGBA()
|
|
r1 := uint8(r >> 8)
|
|
g1 := uint8(g >> 8)
|
|
b1 := uint8(b >> 8)
|
|
a1 := uint8(a >> 8)
|
|
buf[off+0] = r1 - r0
|
|
buf[off+1] = g1 - g0
|
|
buf[off+2] = b1 - b0
|
|
buf[off+3] = a1 - a0
|
|
off += 4
|
|
r0, g0, b0, a0 = r1, g1, b1, a1
|
|
}
|
|
} else {
|
|
for x := bounds.Min.X; x < bounds.Max.X; x++ {
|
|
r, g, b, a := m.At(x, y).RGBA()
|
|
buf[off+0] = uint8(r >> 8)
|
|
buf[off+1] = uint8(g >> 8)
|
|
buf[off+2] = uint8(b >> 8)
|
|
buf[off+3] = uint8(a >> 8)
|
|
off += 4
|
|
}
|
|
}
|
|
if _, err := w.Write(buf); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
return nil
|
|
}
|
|
|
|
// writePix writes the internal byte array of an image to w. It is less general
|
|
// but much faster then encode. writePix is used when pix directly
|
|
// corresponds to one of the TIFF image types.
|
|
func writePix(w io.Writer, pix []byte, nrows, length, stride int) error {
|
|
if length == stride {
|
|
_, err := w.Write(pix[:nrows*length])
|
|
return err
|
|
}
|
|
for ; nrows > 0; nrows-- {
|
|
if _, err := w.Write(pix[:length]); err != nil {
|
|
return err
|
|
}
|
|
pix = pix[stride:]
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func writeIFD(w io.Writer, ifdOffset int, d []ifdEntry) error {
|
|
var buf [ifdLen]byte
|
|
// Make space for "pointer area" containing IFD entry data
|
|
// longer than 4 bytes.
|
|
parea := make([]byte, 1024)
|
|
pstart := ifdOffset + ifdLen*len(d) + 6
|
|
var o int // Current offset in parea.
|
|
|
|
// The IFD has to be written with the tags in ascending order.
|
|
sort.Sort(byTag(d))
|
|
|
|
// Write the number of entries in this IFD.
|
|
if err := binary.Write(w, enc, uint16(len(d))); err != nil {
|
|
return err
|
|
}
|
|
for _, ent := range d {
|
|
enc.PutUint16(buf[0:2], uint16(ent.tag))
|
|
enc.PutUint16(buf[2:4], uint16(ent.datatype))
|
|
count := uint32(len(ent.data))
|
|
if ent.datatype == dtRational {
|
|
count /= 2
|
|
}
|
|
enc.PutUint32(buf[4:8], count)
|
|
datalen := int(count * lengths[ent.datatype])
|
|
if datalen <= 4 {
|
|
ent.putData(buf[8:12])
|
|
} else {
|
|
if (o + datalen) > len(parea) {
|
|
newlen := len(parea) + 1024
|
|
for (o + datalen) > newlen {
|
|
newlen += 1024
|
|
}
|
|
newarea := make([]byte, newlen)
|
|
copy(newarea, parea)
|
|
parea = newarea
|
|
}
|
|
ent.putData(parea[o : o+datalen])
|
|
enc.PutUint32(buf[8:12], uint32(pstart+o))
|
|
o += datalen
|
|
}
|
|
if _, err := w.Write(buf[:]); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
// The IFD ends with the offset of the next IFD in the file,
|
|
// or zero if it is the last one (page 14).
|
|
if err := binary.Write(w, enc, uint32(0)); err != nil {
|
|
return err
|
|
}
|
|
_, err := w.Write(parea[:o])
|
|
return err
|
|
}
|
|
|
|
// Options are the encoding parameters.
|
|
type Options struct {
|
|
// Compression is the type of compression used.
|
|
Compression CompressionType
|
|
// Predictor determines whether a differencing predictor is used;
|
|
// if true, instead of each pixel's color, the color difference to the
|
|
// preceding one is saved. This improves the compression for certain
|
|
// types of images and compressors. For example, it works well for
|
|
// photos with Deflate compression.
|
|
Predictor bool
|
|
}
|
|
|
|
// Encode writes the image m to w. opt determines the options used for
|
|
// encoding, such as the compression type. If opt is nil, an uncompressed
|
|
// image is written.
|
|
func Encode(w io.Writer, m image.Image, opt *Options) error {
|
|
d := m.Bounds().Size()
|
|
|
|
compression := uint32(cNone)
|
|
predictor := false
|
|
if opt != nil {
|
|
compression = opt.Compression.specValue()
|
|
// The predictor field is only used with LZW. See page 64 of the spec.
|
|
predictor = opt.Predictor && compression == cLZW
|
|
}
|
|
|
|
_, err := io.WriteString(w, leHeader)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
// Compressed data is written into a buffer first, so that we
|
|
// know the compressed size.
|
|
var buf bytes.Buffer
|
|
// dst holds the destination for the pixel data of the image --
|
|
// either w or a writer to buf.
|
|
var dst io.Writer
|
|
// imageLen is the length of the pixel data in bytes.
|
|
// The offset of the IFD is imageLen + 8 header bytes.
|
|
var imageLen int
|
|
|
|
switch compression {
|
|
case cNone:
|
|
dst = w
|
|
// Write IFD offset before outputting pixel data.
|
|
switch m.(type) {
|
|
case *image.Paletted:
|
|
imageLen = d.X * d.Y * 1
|
|
case *image.Gray:
|
|
imageLen = d.X * d.Y * 1
|
|
case *image.Gray16:
|
|
imageLen = d.X * d.Y * 2
|
|
case *image.RGBA64:
|
|
imageLen = d.X * d.Y * 8
|
|
case *image.NRGBA64:
|
|
imageLen = d.X * d.Y * 8
|
|
default:
|
|
imageLen = d.X * d.Y * 4
|
|
}
|
|
err = binary.Write(w, enc, uint32(imageLen+8))
|
|
if err != nil {
|
|
return err
|
|
}
|
|
case cDeflate:
|
|
dst = zlib.NewWriter(&buf)
|
|
}
|
|
|
|
pr := uint32(prNone)
|
|
photometricInterpretation := uint32(pRGB)
|
|
samplesPerPixel := uint32(4)
|
|
bitsPerSample := []uint32{8, 8, 8, 8}
|
|
extraSamples := uint32(0)
|
|
colorMap := []uint32{}
|
|
|
|
if predictor {
|
|
pr = prHorizontal
|
|
}
|
|
switch m := m.(type) {
|
|
case *image.Paletted:
|
|
photometricInterpretation = pPaletted
|
|
samplesPerPixel = 1
|
|
bitsPerSample = []uint32{8}
|
|
colorMap = make([]uint32, 256*3)
|
|
for i := 0; i < 256 && i < len(m.Palette); i++ {
|
|
r, g, b, _ := m.Palette[i].RGBA()
|
|
colorMap[i+0*256] = uint32(r)
|
|
colorMap[i+1*256] = uint32(g)
|
|
colorMap[i+2*256] = uint32(b)
|
|
}
|
|
err = encodeGray(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.Gray:
|
|
photometricInterpretation = pBlackIsZero
|
|
samplesPerPixel = 1
|
|
bitsPerSample = []uint32{8}
|
|
err = encodeGray(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.Gray16:
|
|
photometricInterpretation = pBlackIsZero
|
|
samplesPerPixel = 1
|
|
bitsPerSample = []uint32{16}
|
|
err = encodeGray16(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.NRGBA:
|
|
extraSamples = 2 // Unassociated alpha.
|
|
err = encodeRGBA(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.NRGBA64:
|
|
extraSamples = 2 // Unassociated alpha.
|
|
bitsPerSample = []uint32{16, 16, 16, 16}
|
|
err = encodeRGBA64(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.RGBA:
|
|
extraSamples = 1 // Associated alpha.
|
|
err = encodeRGBA(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
case *image.RGBA64:
|
|
extraSamples = 1 // Associated alpha.
|
|
bitsPerSample = []uint32{16, 16, 16, 16}
|
|
err = encodeRGBA64(dst, m.Pix, d.X, d.Y, m.Stride, predictor)
|
|
default:
|
|
extraSamples = 1 // Associated alpha.
|
|
err = encode(dst, m, predictor)
|
|
}
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
if compression != cNone {
|
|
if err = dst.(io.Closer).Close(); err != nil {
|
|
return err
|
|
}
|
|
imageLen = buf.Len()
|
|
if err = binary.Write(w, enc, uint32(imageLen+8)); err != nil {
|
|
return err
|
|
}
|
|
if _, err = buf.WriteTo(w); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
|
|
ifd := []ifdEntry{
|
|
{tImageWidth, dtShort, []uint32{uint32(d.X)}},
|
|
{tImageLength, dtShort, []uint32{uint32(d.Y)}},
|
|
{tBitsPerSample, dtShort, bitsPerSample},
|
|
{tCompression, dtShort, []uint32{compression}},
|
|
{tPhotometricInterpretation, dtShort, []uint32{photometricInterpretation}},
|
|
{tStripOffsets, dtLong, []uint32{8}},
|
|
{tSamplesPerPixel, dtShort, []uint32{samplesPerPixel}},
|
|
{tRowsPerStrip, dtShort, []uint32{uint32(d.Y)}},
|
|
{tStripByteCounts, dtLong, []uint32{uint32(imageLen)}},
|
|
// There is currently no support for storing the image
|
|
// resolution, so give a bogus value of 72x72 dpi.
|
|
{tXResolution, dtRational, []uint32{72, 1}},
|
|
{tYResolution, dtRational, []uint32{72, 1}},
|
|
{tResolutionUnit, dtShort, []uint32{resPerInch}},
|
|
}
|
|
if pr != prNone {
|
|
ifd = append(ifd, ifdEntry{tPredictor, dtShort, []uint32{pr}})
|
|
}
|
|
if len(colorMap) != 0 {
|
|
ifd = append(ifd, ifdEntry{tColorMap, dtShort, colorMap})
|
|
}
|
|
if extraSamples > 0 {
|
|
ifd = append(ifd, ifdEntry{tExtraSamples, dtShort, []uint32{extraSamples}})
|
|
}
|
|
|
|
return writeIFD(w, imageLen+8, ifd)
|
|
}
|