19 Commits

Author SHA1 Message Date
Dimitrii Lopanov
630ded761f Update README.md 2020-02-25 18:11:31 +03:00
Dimitrii Lopanov
674ff110d5 Update README.md 2020-02-25 08:26:00 +03:00
Dimitrii
f70cb0f1e6 upd readme 2020-02-21 15:52:39 +03:00
Dimitrii
49f6b85ba6 gitignore 2020-02-21 15:51:24 +03:00
Dimitrii
1f72563ea0 gitignore 2020-02-21 15:51:12 +03:00
Dimitrii
2c8e594cc7 vscode 2020-02-21 15:50:41 +03:00
Dimitrii
8e0371f19a upd readme 2020-02-21 15:50:10 +03:00
Dimitrii
380db18c1e upd gitignore 2020-02-21 15:48:02 +03:00
Dimitrii
feb1a0353d upd readme 2020-02-21 15:47:21 +03:00
Dimitrii
e486364a2b excess files 2020-02-20 18:10:34 +03:00
Dimitrii
bc4192f727 GOD SAVE STACKOVERFLOW 2020-02-20 18:10:11 +03:00
Dimitrii
48544ab339 got segfault 2020-02-20 17:02:08 +03:00
Dimitrii
fd2b030c94 ? 2020-02-20 13:17:29 +03:00
Dimitrii
f9e8a32a9e ? 2020-02-20 13:16:58 +03:00
Dimitrii
1fad0385f3 readme 2020-02-20 09:38:36 +03:00
Dimitrii
9d89162235 need to check what happens to im.data 2020-02-20 09:35:57 +03:00
Dimitrii
2b2f3f159c need to deal with image 2020-02-19 14:52:19 +03:00
Dimitrii
30fbbb9949 add network 2020-02-18 15:36:17 +03:00
Dimitrii
87175172a2 upd git ignore 2020-02-18 13:21:18 +03:00
21 changed files with 230 additions and 217 deletions

13
.gitignore vendored
View File

@@ -1,10 +1,13 @@
example/main
example/sample.jpg
example/coco.data
example/coco.names
example/yolov3-320.cfg
example/yolov3-320.weights
example/yolov3-416.cfg
example/yolov3-416.weights
example/yolov3.cfg
example/yolov3.weights
darknet.h
*.so
predictions.png
predictions.jpg
main
bad.list
data
.vscode

101
README.md
View File

@@ -1,19 +1,22 @@
# FORK of go-darknet https://github.com/gyonluks/go-darknet
# go-darknet: Go bindings for Darknet
### This is fork of go-darknet https://github.com/gyonluks/go-darknet applied to FORK of Darknet https://github.com/AlexeyAB/darknet
[![GoDoc](https://godoc.org/github.com/LdDl/go-darknet?status.svg)](https://godoc.org/github.com/LdDl/go-darknet)
go-darknet is a Go package, which uses Cgo to enable Go applications to use
YOLO in [Darknet].
go-darknet is a Go package, which uses Cgo to enable Go applications to use YOLO in [Darknet].
## License
## Table of Contents
go-darknet follows [Darknet]'s [license].
- [Requirements](#requirements)
- [Installation](#installation)
- [Usage](#usage)
- [Documentation](#documentation)
- [License](#license)
## Requirements
For proper codebase please use fork of [darknet](https://github.com/LdDl/darknet)
There are instructions for defining GPU/CPU + function for loading image from memory.
For proper codebase please use fork of [darknet](https://github.com/AlexeyAB/darknet). Latest commit I've tested [here](https://github.com/AlexeyAB/darknet/commit/64fb042c63637038671ae9d53c06165599b28912)
In order to use go-darknet, `libdarknet.so` should be available in one of
the following locations:
@@ -26,53 +29,85 @@ Also, [darknet.h] should be available in one of the following locations:
* /usr/include
* /usr/local/include
## Install
To achieve it, after Darknet compilation (via make) execute following command:
```shell
sudo cp libdarknet.so /usr/lib/libdarknet.so && sudo cp include/darknet.h /usr/local/include/darknet.h
```
Note: do not forget to set LIBSO=1 in Makefile before executing 'make':
```Makefile
LIBSO=1
```
## Installation
```shell
go get github.com/LdDl/go-darknet
```
The package name is `darknet`.
## Usage
## Use
Example Go program is provided in the [example] directory. Please refer to the code on how to use this Go package.
Example Go code/program is provided in the [example] directory. Please
refer to the code on how to use this Go package.
Building and running the example program is easy:
Building and running program:
Navigate to [example] folder
```shell
cd $GOPATH/github.com/LdDl/go-darknet/example
#download dataset (coco.names, coco.data, weights and configuration file)
```
Download dataset (sample of image, coco.names, yolov3.cfg, yolov3.weights).
```shell
./download_data.sh
#build program
go build main.go
#run it
./main -configFile yolov3.cfg --dataConfigFile coco.data -imageFile sample.jpg -weightsFile yolov3.weights
```
Note: you don't need *coco.data* file anymore, because sh-script above does insert *coco.names* into 'names' filed in *yolov3.cfg* file (so AlexeyAB's fork can deal with it properly)
So last rows in yolov3.cfg file will look like:
```bash
......
[yolo]
mask = 0,1,2
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
names = coco.names # this is path to coco.names file
```
Build and run program
```
go build main.go && ./main --configFile=yolov3.cfg --weightsFile=yolov3.weights --imageFile=sample.jpg
```
Output should be something like this:
```shell
truck (7): 95.6232% | start point: (78,69) | end point: (222, 291)
truck (7): 81.5451% | start point: (0,114) | end point: (90, 329)
car (2): 99.8129% | start point: (269,192) | end point: (421, 323)
car (2): 99.6615% | start point: (567,188) | end point: (743, 329)
car (2): 99.5795% | start point: (425,196) | end point: (544, 309)
car (2): 96.5765% | start point: (678,185) | end point: (797, 320)
car (2): 91.5156% | start point: (391,209) | end point: (441, 291)
car (2): 88.1737% | start point: (507,193) | end point: (660, 324)
car (2): 83.6209% | start point: (71,199) | end point: (102, 281)
bicycle (1): 59.4000% | start point: (183,276) | end point: (257, 407)
person (0): 96.3393% | start point: (142,119) | end point: (285, 356)
truck (7): 49.5197% | start point: (0,136) | end point: (85, 311)
car (2): 36.3747% | start point: (95,152) | end point: (186, 283)
truck (7): 48.4384% | start point: (95,152) | end point: (186, 283)
truck (7): 45.6590% | start point: (694,178) | end point: (798, 310)
car (2): 76.8379% | start point: (1,145) | end point: (84, 324)
truck (7): 25.5731% | start point: (107,89) | end point: (215, 263)
car (2): 99.8783% | start point: (511,185) | end point: (748, 328)
car (2): 99.8194% | start point: (261,189) | end point: (427, 322)
car (2): 99.6408% | start point: (426,197) | end point: (539, 311)
car (2): 74.5610% | start point: (692,186) | end point: (796, 316)
car (2): 72.8053% | start point: (388,206) | end point: (437, 276)
bicycle (1): 72.2932% | start point: (178,270) | end point: (268, 406)
person (0): 97.3026% | start point: (143,135) | end point: (268, 343)
```
## Documentation
See go-darknet's API documentation at [GoDoc].
## License
go-darknet follows [Darknet]'s [license].
[Darknet]: https://github.com/pjreddie/darknet
[license]: https://github.com/pjreddie/darknet/blob/master/LICENSE
[darknet.h]: https://github.com/pjreddie/darknet/blob/master/include/darknet.h
[include/darknet.h]: https://github.com/pjreddie/darknet/blob/master/include/darknet.h
[Makefile]: https://github.com/pjreddie/darknet/blob/master/Makefile
[darknet.h]: https://github.com/AlexeyAB/darknet/blob/master/include/darknet.h
[include/darknet.h]: https://github.com/AlexeyAB/darknet/blob/master/include/darknet.h
[Makefile]: https://github.com/alexeyab/darknet/blob/master/Makefile
[example]: /example
[GoDoc]: https://godoc.org/github.com/LdDl/go-darknet

8
classes.c Normal file
View File

@@ -0,0 +1,8 @@
#include <stdlib.h>
char *get_class_name(char **names, int index, int names_len) {
if (index >= names_len) {
return NULL;
}
return names[index];
}

14
classes.go Normal file
View File

@@ -0,0 +1,14 @@
package darknet
// #include <stdlib.h>
// #include "classes.h"
import "C"
func makeClassNames(names **C.char, classes int) []string {
out := make([]string, classes)
for i := 0; i < classes; i++ {
n := C.get_class_name(names, C.int(i), C.int(classes))
out[i] = C.GoString(n)
}
return out
}

3
classes.h Normal file
View File

@@ -0,0 +1,3 @@
#pragma once
extern char *get_class_name(char **names, int index, int names_len);

View File

@@ -1,24 +0,0 @@
#include <stdlib.h>
#include <darknet.h>
void free_class_names(char **names)
{
free(names);
}
char ** read_class_names(char *data_cfg)
{
list *options = read_data_cfg(data_cfg);
char *name_list = option_find_str(options, "names", "data/names.list");
return get_labels(name_list);
}
char * get_class_name(char **names, int index, int names_len)
{
if (index >= names_len) {
return NULL;
}
return names[index];
}

View File

@@ -1,28 +0,0 @@
package darknet
// #include <stdlib.h>
// #include "classnames.h"
import "C"
import "unsafe"
func freeClassNames(names **C.char) {
C.free_class_names(names)
}
func loadClassNames(dataConfigFile string) **C.char {
d := C.CString(dataConfigFile)
defer C.free(unsafe.Pointer(d))
return C.read_class_names(d)
}
func makeClassNames(names **C.char, classes int) []string {
out := make([]string, classes)
for i := 0; i < classes; i++ {
n := C.get_class_name(names, C.int(i), C.int(classes))
s := C.GoString(n)
out[i] = s
}
return out
}

View File

@@ -1,5 +0,0 @@
#pragma once
extern void free_class_names(char **names);
extern char ** read_class_names(char *data_cfg);
extern char * get_class_name(char **names, int index, int names_len);

View File

@@ -1,5 +1,5 @@
package darknet
// #cgo CFLAGS: -I/usr/local/include -I/usr/local/cuda/include
// #cgo LDFLAGS: -L/usr/local/lib -ldarknet -lm
// #cgo LDFLAGS: -L/usr/lib -ldarknet -lm
import "C"

View File

@@ -1,19 +1,17 @@
#include <darknet.h>
detection * get_detection(detection *dets, int index, int dets_len)
{
#include "detection.h"
detection *get_detection(detection *dets, int index, int dets_len) {
if (index >= dets_len) {
return NULL;
}
return dets + index;
}
float get_detection_probability(detection *det, int index, int prob_len)
{
float get_detection_probability(detection *det, int index, int prob_len) {
if (index >= prob_len) {
return .0;
}
return det->prob[index];
}

View File

@@ -25,19 +25,15 @@ type DetectionResult struct {
OverallTimeTaken time.Duration
}
func makeDetection(img *Image, det *C.detection, threshold float32, classes int,
classNames []string) *Detection {
func makeDetection(img *DarknetImage, det *C.detection, threshold float32, classes int, classNames []string) *Detection {
if det == nil {
return &Detection{}
}
dClassIDs := make([]int, 0)
dClassNames := make([]string, 0)
dProbs := make([]float32, 0)
for i := 0; i < int(classes); i++ {
dProb := float32(
C.get_detection_probability(det, C.int(i), C.int(classes)))
dProb := float32(C.get_detection_probability(det, C.int(i), C.int(classes)))
if dProb > threshold {
dClassIDs = append(dClassIDs, i)
cN := classNames[i]
@@ -45,7 +41,6 @@ func makeDetection(img *Image, det *C.detection, threshold float32, classes int,
dProbs = append(dProbs, dProb*100)
}
}
fImgW := C.float(img.Width)
fImgH := C.float(img.Height)
halfRatioW := det.bbox.w / 2.0
@@ -66,19 +61,16 @@ func makeDetection(img *Image, det *C.detection, threshold float32, classes int,
ClassNames: dClassNames,
Probabilities: dProbs,
}
return &out
}
func makeDetections(img *Image, detections *C.detection, detectionsLength int,
threshold float32, classes int, classNames []string) []*Detection {
func makeDetections(img *DarknetImage, detections *C.detection, detectionsLength int, threshold float32, classes int, classNames []string) []*Detection {
// Make list of detection objects.
ds := make([]*Detection, detectionsLength)
for i := 0; i < int(detectionsLength); i++ {
det := C.get_detection(detections, C.int(i), C.int(classes))
det := C.get_detection(detections, C.int(i), C.int(detectionsLength))
d := makeDetection(img, det, threshold, classes, classNames)
ds[i] = d
}
return ds
}

View File

@@ -1,6 +1,5 @@
wget --output-document=sample.jpg https://cdn-images-1.medium.com/max/800/1*EYFejGUjvjPcc4PZTwoufw.jpeg
wget --output-document=coco.names https://raw.githubusercontent.com/pjreddie/darknet/master/data/coco.names
wget --output-document=coco.data https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/coco.data
sed -i 's#data/coco.names#coco.names#g' coco.data
wget --output-document=yolov3.cfg https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3.cfg
wget --output-document=coco.names https://raw.githubusercontent.com/AlexeyAB/darknet/master/data/coco.names
wget --output-document=yolov3.cfg https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov3.cfg
sed -i -e "\$anames = coco.names" yolov3.cfg
wget --output-document=yolov3.weights https://pjreddie.com/media/files/yolov3.weights

View File

@@ -1,15 +1,17 @@
package main
import (
"bytes"
"flag"
"fmt"
"image"
"image/jpeg"
"log"
"os"
darknet "github.com/LdDl/go-darknet"
)
var dataConfigFile = flag.String("dataConfigFile", "",
"Path to data configuration file. Example: cfg/coco.data")
var configFile = flag.String("configFile", "",
"Path to network layer configuration file. Example: cfg/yolov3.cfg")
var weightsFile = flag.String("weightsFile", "",
@@ -24,7 +26,7 @@ func printError(err error) {
func main() {
flag.Parse()
if *dataConfigFile == "" || *configFile == "" || *weightsFile == "" ||
if *configFile == "" || *weightsFile == "" ||
*imageFile == "" {
flag.Usage()
@@ -33,10 +35,9 @@ func main() {
n := darknet.YOLONetwork{
GPUDeviceIndex: 0,
DataConfigurationFile: *dataConfigFile,
NetworkConfigurationFile: *configFile,
WeightsFile: *weightsFile,
Threshold: .5,
Threshold: .25,
}
if err := n.Init(); err != nil {
printError(err)
@@ -44,14 +45,23 @@ func main() {
}
defer n.Close()
img, err := darknet.ImageFromPath(*imageFile)
infile, err := os.Open(*imageFile)
if err != nil {
printError(err)
return
panic(err.Error())
}
defer infile.Close()
src, err := jpeg.Decode(infile)
if err != nil {
panic(err.Error())
}
defer img.Close()
dr, err := n.Detect(img)
imgDarknet, err := darknet.Image2Float32(src)
if err != nil {
panic(err.Error())
}
defer imgDarknet.Close()
dr, err := n.Detect(imgDarknet)
if err != nil {
printError(err)
return
@@ -60,7 +70,6 @@ func main() {
log.Println("Network-only time taken:", dr.NetworkOnlyTimeTaken)
log.Println("Overall time taken:", dr.OverallTimeTaken, len(dr.Detections))
for _, d := range dr.Detections {
for i := range d.ClassIDs {
bBox := d.BoundingBox
fmt.Printf("%s (%d): %.4f%% | start point: (%d,%d) | end point: (%d, %d)\n",
@@ -72,3 +81,9 @@ func main() {
}
}
}
func imageToBytes(img image.Image) ([]byte, error) {
buf := new(bytes.Buffer)
err := jpeg.Encode(buf, img, nil)
return buf.Bytes(), err
}

13
image.c Normal file
View File

@@ -0,0 +1,13 @@
#include <darknet.h>
void fill_image_f32(image* im, int w, int h, int c, float* data) {
int i;
for (i = 0; i < w*h*c; i++) {
im->data[i] = data[i];
}
}
void set_data_f32_val(float* data, int index, float value) {
data[index] = value;
}

View File

@@ -1,66 +1,64 @@
package darknet
// #include <darknet.h>
// #include <stdlib.h>
// #include "image.h"
import "C"
import (
"errors"
"image"
"unsafe"
"golang.org/x/image/draw"
)
// Image represents the image buffer.
type Image struct {
// DarknetImage represents the image buffer.
type DarknetImage struct {
Width int
Height int
image C.image
}
var (
errUnableToLoadImage = errors.New("unable to load image")
)
// Close and release resources.
func (img *Image) Close() error {
func (img *DarknetImage) Close() error {
C.free_image(img.image)
return nil
}
// ImageFromPath reads image file specified by path.
func ImageFromPath(path string) (*Image, error) {
p := C.CString(path)
defer C.free(unsafe.Pointer(p))
img := Image{
image: C.load_image_color(p, 0, 0),
// https://stackoverflow.com/questions/33186783/get-a-pixel-array-from-from-golang-image-image/59747737#59747737
func imgTofloat32(src image.Image) []float32 {
bounds := src.Bounds()
width, height := bounds.Max.X, bounds.Max.Y
srcRGBA := image.NewRGBA(src.Bounds())
draw.Copy(srcRGBA, image.Point{}, src, src.Bounds(), draw.Src, nil)
ans := []float32{}
red := []float32{}
green := []float32{}
blue := []float32{}
for y := 0; y < height; y++ {
for x := 0; x < width; x++ {
idxSource := (y*width + x) * 4
pix := srcRGBA.Pix[idxSource : idxSource+4]
rpix, gpix, bpix := float32(pix[0])/257.0, float32(pix[1])/257.0, float32(pix[2])/257.0
red = append(red, rpix)
green = append(green, gpix)
blue = append(blue, bpix)
}
}
ans = append(ans, red...)
ans = append(ans, green...)
ans = append(ans, blue...)
return ans
}
if img.image.data == nil {
return nil, errUnableToLoadImage
// Image2Float32 Returns []float32 representation of image.Image
func Image2Float32(img image.Image) (*DarknetImage, error) {
ans := imgTofloat32(img)
width := img.Bounds().Dx()
height := img.Bounds().Dy()
imgDarknet := &DarknetImage{
Width: width,
Height: height,
image: C.make_image(C.int(width), C.int(height), 3),
}
img.Width = int(img.image.w)
img.Height = int(img.image.h)
return &img, nil
}
// ImageFromMemory reads image file data represented by the specified byte
// slice.
func ImageFromMemory(buf []byte) (*Image, error) {
cBuf := C.CBytes(buf)
defer C.free(cBuf)
img := Image{
image: C.load_image_from_memory_color((*C.uchar)(cBuf),
C.int(len(buf)), 0, 0),
}
if img.image.data == nil {
return nil, errUnableToLoadImage
}
img.Width = int(img.image.w)
img.Height = int(img.image.h)
return &img, nil
C.fill_image_f32(&imgDarknet.image, C.int(width), C.int(height), 3, (*C.float)(unsafe.Pointer(&ans[0])))
return imgDarknet, nil
}

6
image.h Normal file
View File

@@ -0,0 +1,6 @@
#pragma once
#include <darknet.h>
extern void fill_image_f32(image *im, int w, int h, int c, float* data);
extern void set_data_f32_val(float* data, int index, float value);

View File

@@ -4,19 +4,25 @@
#include "network.h"
int get_network_layer_classes(network *n, int index) {
return n->layers[index].classes;
}
#include "detection.h"
struct network_box_result perform_network_detect(network *n, image *img, int classes, float thresh, float hier_thresh, float nms) {
image sized = letterbox_image(*img, n->w, n->h);
struct network_box_result perform_network_detect(network *n, image *img, int classes, float thresh, float hier_thresh, float nms, int letter_box) {
image sized;
if (letter_box) {
sized = letterbox_image(*img, n->w, n->h);
} else {
sized = resize_image(*img, n->w, n->h);
}
struct network_box_result result = { NULL };
float *X = sized.data;
network_predict(n, X);
result.detections = get_network_boxes(n, img->w, img->h,thresh, hier_thresh, 0, 1, &result.detections_len);
network_predict_ptr(n, X);
int nboxes = 0;
detection *dets = get_network_boxes(n, img->w, img->h, thresh, hier_thresh, 0, 1, &nboxes, letter_box);
result.detections = get_network_boxes(n, img->w, img->h, thresh, hier_thresh, 0, 1, &result.detections_len, letter_box);
if (nms) {
do_nms_sort(result.detections, result.detections_len, classes, nms);
}
free_image(sized);
return result;
}

View File

@@ -15,7 +15,6 @@ import (
// YOLONetwork represents a neural network using YOLO.
type YOLONetwork struct {
GPUDeviceIndex int
DataConfigurationFile string
NetworkConfigurationFile string
WeightsFile string
Threshold float32
@@ -29,8 +28,8 @@ type YOLONetwork struct {
}
var (
errNetworkNotInit = errors.New("network not initialised")
errUnableToInitNetwork = errors.New("unable to initialise")
errNetworkNotInit = errors.New("Network not initialised")
errUnableToInitNetwork = errors.New("Unable to initialise")
)
// Init the network.
@@ -39,29 +38,18 @@ func (n *YOLONetwork) Init() error {
defer C.free(unsafe.Pointer(nCfg))
wFile := C.CString(n.WeightsFile)
defer C.free(unsafe.Pointer(wFile))
// GPU device ID must be set before `load_network()` is invoked.
C.cuda_set_device(C.int(n.GPUDeviceIndex))
n.cNet = C.load_network(nCfg, wFile, 0)
if n.cNet == nil {
return errUnableToInitNetwork
}
C.set_batch_network(n.cNet, 1)
C.srand(2222222)
// Currently, hierarchal threshold is always 0.5.
n.hierarchalThreshold = .5
// Currently NMS is always 0.4.
n.nms = .4
n.Classes = int(C.get_network_layer_classes(n.cNet, n.cNet.n-1))
cClassNames := loadClassNames(n.DataConfigurationFile)
defer freeClassNames(cClassNames)
n.ClassNames = makeClassNames(cClassNames, n.Classes)
n.hierarchalThreshold = 0.5
n.nms = 0.45
metadata := C.get_metadata(nCfg)
n.Classes = int(metadata.classes)
n.ClassNames = makeClassNames(metadata.names, n.Classes)
return nil
}
@@ -70,34 +58,26 @@ func (n *YOLONetwork) Close() error {
if n.cNet == nil {
return errNetworkNotInit
}
C.free_network(n.cNet)
C.free_network(*n.cNet)
n.cNet = nil
return nil
}
// Detect specified image.
func (n *YOLONetwork) Detect(img *Image) (*DetectionResult, error) {
// Detect specified image
func (n *YOLONetwork) Detect(img *DarknetImage) (*DetectionResult, error) {
if n.cNet == nil {
return nil, errNetworkNotInit
}
startTime := time.Now()
result := C.perform_network_detect(n.cNet, &img.image, C.int(n.Classes),
C.float(n.Threshold), C.float(n.hierarchalThreshold), C.float(n.nms))
result := C.perform_network_detect(n.cNet, &img.image, C.int(n.Classes), C.float(n.Threshold), C.float(n.hierarchalThreshold), C.float(n.nms), C.int(0))
endTime := time.Now()
defer C.free_detections(result.detections, result.detections_len)
ds := makeDetections(img, result.detections, int(result.detections_len),
n.Threshold, n.Classes, n.ClassNames)
ds := makeDetections(img, result.detections, int(result.detections_len), n.Threshold, n.Classes, n.ClassNames)
endTimeOverall := time.Now()
out := DetectionResult{
Detections: ds,
NetworkOnlyTimeTaken: endTime.Sub(startTime),
OverallTimeTaken: endTimeOverall.Sub(startTime),
}
return &out, nil
}

View File

@@ -8,4 +8,4 @@ struct network_box_result {
};
extern int get_network_layer_classes(network *n, int index);
extern struct network_box_result perform_network_detect(network *n, image *img, int classes, float thresh, float hier_thresh, float nms);
extern struct network_box_result perform_network_detect(network *n, image *img, int classes, float thresh, float hier_thresh, float nms, int letter_box);