Darknet: How to train tiny YOLO?

[smitshilu]

Hello,
I want to train tiny YOLO on my own dataset. How can I train it?

Thank you,

[AlexeyAB]

Hi,

• How to train tiny-yolo: https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects

• How to train (to detect your custom objects): https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

• How to train (Pascal VOC Data): https://github.com/AlexeyAB/darknet#how-to-train-pascal-voc-data

[dexception]

@AlexeyAB
Can you tell me on how many images yolov2-tiny model was trained ?

[chinmay5]

@AlexeyAB for training the tiny-yolo model having 42 classes, how many iterations are expected? The conversion rate for tiny Yolo is pretty slow. Should I consider trying to change few hyper-parameters?
yolo-tiny-obj.cfg.txt
I am attaching the config file here. I just updated the filter size and am able to start the training.

I am also attaching portion of the output for the training phase. I think the model is taking really long to train because the "non-tiny" version of YOLO had started giving me good result in only 1500 iterations while here I am through 18K iterations and still struggling

[AlexeyAB]

for training the tiny-yolo model having 42 classes, how many iterations are expected?

You should train about 42 x 2000 ~= 84 000 iterations (if you have about 84 000 images)

Should I consider trying to change few hyper-parameters?

It depends on what kind of object do you try to detect.

In general, better to use yolov3-tiny.cfg instead of yolov2-tiny.cfg it has ~the same performance (speed) but much more accuracy.
How to train yolov3-tiny: https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects

[chinmay5]

Hi @AlexeyAB , my only problem is, I have around 600 distinct annotated training images. This is the number that was provided in the competition as well so not sure if I should try running so many iterations with the same data size

[AlexeyAB]

@chinmay5 Yes, for 600 images, you should train about 4000 - 8000 iterations.
And more if you increased data augmentation parameters: jitter, hue, saturation, exposure

[chinmay5]

@AlexeyAB I am able to perform the training on the model and thank you so much for that. Next, I want to convert this tiny-yolo-v3 model into a coreml model so that I can use it on the iPhone. Can someone direct me to some references. More importantly, is there a way for a direct conversion somehow?

[chinmay5]

@AlexeyAB is there a way to obtain graphs or other measure to figure out which model shall I use. I see that error values are generated but I did not get a graph where they had been plotted. I tried with the IoU value but I really want to know if it was the best model and whether I can improve. I have a separate test set of some 300 images as well

[AlexeyAB]

@chinmay5 Try to compare mAP of your models: https://github.com/AlexeyAB/darknet#when-should-i-stop-training
Or just take the latest weights-file with the lowest training avg loss.

[keides2]

Hi @AlexeyAB ,
I trained with yolov3-tiny_obj.cfg and yolov3-tiny.weights with the following command.
$ ./darknet partial cfg / yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15
$ ./darknet detector train data / obj.data yolov3-tiny_obj.cfg yolov3-tiny.conv.15

I want to convert the generated .weights file to pb file using darkflow's following command and run it on android mobile phone, but I got an error. (I will add the details of the error tomorrow)
$ ./flow --module cfg/yolov3-tiny_obj.cfg --load bin/yolov3-tiny_obj-10000.weights --savepb

Can darkflow convert yolov3-tiny_obj_10000.weights to pb file?

If possible, can I use this pb file with android-yolo-v2-master?

Addition:
(tensorflow) [shimatani@bslpc168 ~/darkflow]$./flow --model cfg/yolov3-tiny_obj.cfg --load bin/yolov
3-tiny_obj_10000.weights --savepb

/home/shimatani/darkflow/darkflow/dark/darknet.py:54: UserWarning: ./cfg/yolov3-tiny_obj_10000.cfg not found, use cfg/yolov3-tiny_obj.cfg instead
cfg_path, FLAGS.model))
Parsing cfg/yolov3-tiny_obj.cfg
Layer [yolo] not implemented

Correction:
If possible, can I use this pb file with android-yolo-v2 of szaza?

[AlexeyAB]

@keides2

As I see https://github.com/szaza/android-yolo-v2 and https://github.com/thtrieu/darkflow don't support Yolo v3. It supports only Yolo v2.

[keides2]

I got it. Thank you so much, @AlexeyAB .

[Sweta02018]

I am trying to train yolov3 with darknet53 on 12289 images on GPU with 2GB space.
I get cuda out of memory error. I have changed all possible values of batch and subdivision still I am getting the same error after some 80th iterations. Can anyone help me to know about what should be the sufficient space of GPU for Darknet53?
Thanks in advance!

[chinmay5]

@AlexeyAB correct me if I am wrong, but I think the OOM is dependent upon the resolution and size of the images. Although I would say, that 2 GB memory is way too less. In my opinion, 4GB should be the minimum starting point

[AlexeyAB]

@chinmay5

• OOM doesn't depend on size of images
• OOM depends on network resolution (width= height= in cfg-file) and on ratio batch/subdivisions in cfg-file. GPU_memory_usage ~ width * height * batch / subdivisions

[sleebapaul]

@chinmay5 Have a look at here.

[123alaa]

@AlexeyAB Hey, I'm trying to train yolov3 (reproduce your training) on coco dataset, so i have these configuration correct me if something wrong:

1. i used batch_size = 64 subdivision=16
2. i added also the images that don't have annotations (you say that it is help)
3. i started the learning rate 0.001 but i see that the is decays, but it is not changes in the iterations
4. the image size is 608 with the same anchors for 416, is it ok?
   i try to do this three weeks ago, with various config, the problem that the training avg loss stuck on ~5.00 after 10000 iteration, this is valid for convergence??

Can you help.
Best

[Malouke]

hi,
i am confuse you see OOM not depand on size of your memory but about image size ?
what about resizing in top of code because all input images don't have the same shape? No?
please give us more details because actully i dont undertstand

i have arround 1000 images without any kind of data augmentation it just original ones (balls colors billards) if i should to do data augmenatation tell me .

• i have a machine with gtx960 gpu 2Gb and cpu ( i7 )
  can i ask you to telle me about the best preactices abouts size of batch an subdivisions....
  thx

[visshvesh]

I got it. Thank you so much, @AlexeyAB .

@keides2 How did you manage to make .pb file out of .weights file for yolov3 in darknet?

[BaijuMishra]

@AlexeyAB , I'm trying to train yolov2 (reproduce your training) on custom dataset(class2),
Actual Image Size is 1106*620 (6000 Images)

in config - 416*416
batch_size = 64 subdivision=8
learning rate 0.01
the problem is that avg loss stuck on ~1.00 after 5000 iteration,

Kindly help me out

[AlexeyAB]

@BaijuMishra It is normal. Train 5000 - 8000 iterations, then check mAP by using this repository: https://github.com/AlexeyAB/darknet#when-should-i-stop-training

[BaijuMishra]

@AlexeyAB Thank you for the Response.. :), Definitely I will check for mAP with Higher Iteration but before proceeding, I have some confusion on few parameters-
1- I have a image with dimension 1106 x 620 , Can I use 960x618(which is closer to my image dimension) in config file as width and Height ?
2- When I set Random = 1 , I am getting cuda malloc error .

Please guide me on the same .

and do we have any option to add new Objects using transfer learning from existing Darknet Model .?

Regards,
Baiju

[BaijuMishra]

@AlexeyAB ,In Yolov3,. Please let me know ,How to save *.Weights file into *.pb and *.cfg file to *.metadata ?

Thank you ,

Regards,
Baiju

[taojin1992]

@chinmay5 Yes, for 600 images, you should train about 4000 - 8000 iterations.
And more if you increased data augmentation parameters: jitter, hue, saturation, exposure

@AlexeyAB What is your rationale to predetermine this range of potential iteration number? Thank you!

[AlainMindana]

Hello :) How can I train tiny YOLOv3 and install it in Android? @AlexeyAB

[abdalkhalik]

@AlainMindana

I tried this on yolov4 and it worked, i'm pretty sure its the same for u r requirments
Get the .weight file then use this repo https://github.com/hunglc007/tensorflow-yolov4-tflite (which has both a conversion and an android project) to:
1- convert .weights -> .tflite (with the flag "--quantize_mode float16")
2- replace the tflite file in android project with your generated tflite file
3- go to download_model.gradle and remove the code which download a predefined tflite model