Darknet: YOLOv3 support

[wkdhkr]

https://pjreddie.com/darknet/yolo/

YOLO v3 is now available. can I use v3 cfg in Windows?

[spinoza1791]

Take a look at yolov3.cfg: https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg
It has new parameters in it that I don't quite understand. Hopefully, someone can fill in the blanks. For example, the [shortcut] layer replaces the typical maxpool layers, the last detection layer seems to have a new equation (filters=255?), what is mask 0-9? and the anchored [yolo] layers are spread out. People aren't going to know how to custom this new .cfg, because the new features aren't documented anywhere.

[lesreaper]

It looks like the training procedure is the same, but uses the new Darknet53 model instead of Darknet19 or TinyYolo (currently).

The inferencing time is definitely slower though.

[AlexeyAB]

@wkdhkr @spinoza1791 @lesreaper

You can train Yolo v3 using this repo

• [shortcut] - already implemented - this is simply residual connection from ResNet

• [upsample] - is a new layer - something like old [reorg] layer but with correct premutations: https://github.com/pjreddie/darknet/blob/d3828827e70b293a3045a1eb80bfb4026095b87b/src/blas.c#L334-L349

• [yolo] - new detector layer instead of old [region] layer: https://github.com/pjreddie/darknet/blob/d3828827e70b293a3045a1eb80bfb4026095b87b/src/yolo_layer.c

[visionNoob]

Thank you @AlexeyAB

[TaihuLight]

What is the difference between YOLO v3 and YOLO v2? The new layer and new network structures are add?

[TheMikeyR]

@TaihuLight seems like he added feature pyramids https://github.com/pjreddie/darknet/issues/555 and probably more.

[DeHysolate]

YOLO v3 added. (d9ae3dd681ed1c98e807ff937dbbb9cfc4d19fe0)

If I want to use YOLO v3 in dll format in Windows, can I use the function in yolo_v2_class.hpp?
Its name is v2, so I wonder if v3 applies.

[AlexeyAB]

Yolo v3 can be trained successfully.

Just trained 5000 iterations on Windows 7 x64, CUDA 9.1, cuDNN 7.0, OpenCV 3.4.0 using this command:
darknet.exe detector train data/obj.data yolov3_obj.cfg darknet53.conv.74

---

Result accuracy:
darknet.exe detector map data/obj.data yolov3_obj.cfg backup/yolov3_obj_5000.weights

 for thresh = 0.25, precision = 0.97, recall = 0.99, F1-score = 0.98
 for thresh = 0.25, TP = 8490, FP = 271, FN = 88, average IoU = 74.62 %

 mean average precision (mAP) = 0.906889, or 90.69 %
Total Detection Time: 465.000000 Seconds

---

Detection:
darknet.exe detector test data/obj.data yolov3_obj.cfg backup/yolov3_obj_5000.weights img_4_109.jpg

img_4_109.jpg: Predicted in 0.035000 seconds.
Car-camaro: 100%

---

darknet.exe detector test data/obj.data yolov3_obj.cfg backup/yolov3_obj_5000.weights img_5_769.jpg

img_5_769.jpg: Predicted in 0.034000 seconds.
Car-lamborghini: 100%

---

Average loss during training:

[wkdhkr]

great work!!

[ahsan856jalal]

@AlexeyAB how do you add this chart showing errors in pjreddie's fork. Moreover did you modify this fork as your fork always converge faster than pjreddie's

[AlexeyAB]

@ahsan856jalal I made many changes some of them I already do not remember, some of them: https://github.com/AlexeyAB/darknet/issues/529#issuecomment-377204382

Loss chart is enabled by default in this fork:

• https://github.com/AlexeyAB/darknet/blob/a7a2e1bb4b0efa55ac2af91358e8c8d2d20076a7/src/detector.c#L107
• https://github.com/AlexeyAB/darknet/blob/a7a2e1bb4b0efa55ac2af91358e8c8d2d20076a7/src/detector.c#L176-L177

[Gotta-C]

darknet.exe detector map ，where the functions you defined ? There is no detector.c, I want to draw the presion and recall curve,but only extract map from training log.

[dfsaw]

is opencv necessary for this chart to get created?

[AlexeyAB]

@dfsaw Yes.

[dfsaw]

I am getting this error while running make file with open cv=1

cannot find -lippicv .

Can please someone help me out.

[AlexeyAB]

@dfsaw

• What OpenCV version do you use?

Try to do

sudo cp 3rdparty/ippicv/unpack/ippicv/lib/intel64/libippicv.a /usr/local/lib/

Or try to install OpenCV 3.3.0: https://github.com/opencv/opencv/archive/3.3.0.zip

---

May it can help to solve your issue:

• http://answers.opencv.org/question/84265/compiling-error-with-lippicv/
• https://github.com/opencv/opencv/issues/5852
• https://stackoverflow.com/questions/34401117/compiling-code-with-opencv-usr-bin-ld-cannot-find-lippicv
• https://askubuntu.com/questions/720528/opencv-compiling-error-ippicv

[dfsaw]

i am using 3.3.0.10 I am trying to run on GPU. it is working fine without Opencv but since I need to generate training loss chart i need open cv

[dfsaw]

Is there any other way to generate avg loss for training set without opencv. It would be really great if someone could help.

I tried running this script but the log file under scripts folder is not generated

/darknet detector train build/darknet/x64/data/obj.data cfg/yolov3-tiny-obj.cfg yolov3-tiny.conv.15 > test.log

[vikasmishra591]

can anyone tell me how to plot the graph of loss function of yolov3.

[jalaldev1980]

Dear
@dfsaw & @vikasmishra591

did you get any solutions ?

how can I get the .log file ?

before or after the training is there any why to save the log outputs
I want to plot the Accuracy vs epoch , loss vs epoch

I am training my own dataset using darknet , tiny yolo v2

thank you

[Rahul-Venugopal]

Hi @jalaldev1980 ,

Got any solution ?

[AlexeyAB]

@Rahul-Venugopal

You can save the log output by using | tee log.txt at the end of command:
./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 | tee log.txt

(on windows you should download wtee.exe: https://code.google.com/archive/p/wintee/ and use command darknet.exe detector train cfg/coco.data yolov3.cfg darknet53.conv.74 | wtee log.txt)

Then you can use this script to draw Loss-chart by using saved log.txt file: https://github.com/AlexeyAB/darknet/tree/master/scripts/log_parser

---

Also i the latest version of darknet the Loss-chart will be saved automatically to the file chart.png for each 100 iterations, if you compiled Darknet with OpenCV.
(even if you use flag -dont_show if you train the model on Amazon EC2)

If you use flag -map then will be saved Loss-chart & accuracy mAP-chart to the chart.png

---

Also you can run training on the remote server:
./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 -dont_show -mjpeg_port 8090 -map
And then see Loss & mAP charts by using web-browser Chrome/Firefox by using URL with ip-address of remote server: http://ip-address:8090

More: https://github.com/AlexeyAB/darknet#how-to-use-on-the-command-line

[Rahul-Venugopal]

Hi @AlexeyAB ,

Thank you so much for such a quick response . It saved my day :)

[kocica]

Hi @AlexeyAB

I am training on remote cluster where I can not use graphics therefore I am force to use "dont_show" parameter. Is there any option to save the chart instead of printing it while training or do I have to write it my self? Thank you in advance.

[AlexeyAB]

@kocica Hi,

Compile with OPENCV=1 and chart.png will be saved automatically for each 100 iterations

• ./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 -dont_show will be saved Loss-chart

• ./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 -dont_show -map will be saved mAP & Loss-chart

[anthonymg1994]

Darknet.exe detector train ..... -map,
Doesnt show map progress in graph and only extract map from training log.

[AlexeyAB]

@anthonymg1994
darknet.exe detector train ..... -map - it shows map progress in graph during training. It doesn't extract map from training log.

[Leprechault]

@ahsan856jalal I made many changes some of them I already do not remember, some of them: #529 (comment)

Loss chart is enabled by default in this fork:

• https://github.com/AlexeyAB/darknet/blob/a7a2e1bb4b0efa55ac2af91358e8c8d2d20076a7/src/detector.c#L107
• https://github.com/AlexeyAB/darknet/blob/a7a2e1bb4b0efa55ac2af91358e8c8d2d20076a7/src/detector.c#L176-L177

Please, @ahsan856jalal for me doesn't work.

I add in line 107:

img = draw_train_chart(max_img_loss, net.max_batches, number_of_lines, img_size);

and in lines 176 to 177:

if(!dont_show)

`draw_train_loss(img, img_size, avg_loss, max_img_loss, i, net.max_batches);`

and the detector.c was changed by (split some changes):

        /*
           int zz;
           for(zz = 0; zz < train.X.cols; ++zz){
           image im = float_to_image(net->w, net->h, 3, train.X.vals[zz]);
           int k;
           for(k = 0; k < l.max_boxes; ++k){
           box b = float_to_box(train.y.vals[zz] + k*5, 1);
           printf("%f %f %f %f\n", b.x, b.y, b.w, b.h);
           draw_bbox(im, b, 1, 1,0,0);
           }
           show_image(im, "truth11");
           img = draw_train_chart(max_img_loss, net.max_batches, number_of_lines, img_size); 
           save_image(im, "truth11");
           }
         */

and

        if (xmin < 0) xmin = 0;
        if (ymin < 0) ymin = 0;
        if (xmax > w) xmax = w;
        if (ymax > h) ymax = h;
        if(!dont_show) 
        draw_train_loss(img, img_size, avg_loss, max_img_loss, i, net.max_batches); 

Any tips?

[LuizBoina]

@Rahul-Venugopal

You can save the log output by using | tee log.txt at the end of command:
./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 | tee log.txt

(on windows you should download wtee.exe: https://code.google.com/archive/p/wintee/ and use command darknet.exe detector train cfg/coco.data yolov3.cfg darknet53.conv.74 | wtee log.txt)

Then you can use this script to draw Loss-chart by using saved log.txt file: https://github.com/AlexeyAB/darknet/tree/master/scripts/log_parser

Also i the latest version of darknet the Loss-chart will be saved automatically to the file chart.png for each 100 iterations, if you compiled Darknet with OpenCV.
(even if you use flag -dont_show if you train the model on Amazon EC2)

If you use flag -map then will be saved Loss-chart & accuracy mAP-chart to the chart.png

Also you can run training on the remote server:
./darknet detector train cfg/coco.data yolov3.cfg darknet53.conv.74 -dont_show -mjpeg_port 8090 -map
And then see Loss & mAP charts by using web-browser Chrome/Firefox by using URL with ip-address of remote server: http://ip-address:8090

More: https://github.com/AlexeyAB/darknet#how-to-use-on-the-command-line

Hi, I trained the darknet with -map and have the information about map progress during the training. However, when I use this script with the log the output only show the loss graph, do you know how can I use this script or similar to plot map too?

Thanks.

[stephanecharette]

Did you forget to use the -map parameter?

This is the command that I personally use:

~/darknet/darknet detector -map -dont_show train cars.data cars.cfg 2>&1 | tee --append output.log

[LuizBoina]

Hi @stephanecharette, no, I dont.
I used this line of command on my remote server to train the network:

./darknet detector train data/obj.data cfg/yolov3_custom.cfg darknet53.conv.74 -map -dont_show | tee log.txt

The log.txt have a lot of map progress like this:

(next mAP calculation at 8305 iterations)
Last accuracy [email protected] = 93.41 %, best = 93.41 %

And the few last lines have similar information as when I run detector map, but when I use log_parser this only give me the graph with loss and nothing about the map.

[stephanecharette]

I don't know if it matters, but I have -map -dont_show at the start of the command, while you have it at the end. Note that the mAP% doesn't appear on the chart until 1000 iterations have been completed.