Start Training YOLOv3 Using Darknet to Detect Custom Objects || YOLOv3 Series 5

In this video we'll modify the cfg file, put all the images and bounding box labels in the right folders, and start training YOLOv3! P.S. check out the description for all the links!)

I really encourage you to ask questions, if something's not clear or you just want to, happy to help!)

classes= 2
backup = backup/

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
It would mean the world to me, if you decided to support me and the channel =)

►You may consider watching ads that show up on the videos

Making these videos takes a lot of time and effort, so If you decide to support me, please don't hesitate get in touch with me as I'd like to thank you personally!
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Thanks for watching!

[IvanGoncharovAI]

Hey! Thanks for watching! Here are some links to the previous parts of the series:

[leminhtantai]

Thank you so much for your videos. I'm waiting for your new videos.
YOLOV3 Python Darknet ...

[TheAkhilvsrp]

Im working on a problem and this solves my doubts, Thank you

[acityowl]

Hi Ivan, firstly thank you so much for the videos. You have inspired me greatly!
I have a question regarding testing. Should I be using ./darknet detector test ? If I do this, must I uncomment the testing part of the cfg file and comment out the training? Or should I use map instead? Currently, I'm using map for my validation but not sure what to do for testing. Thank you!

[yaggeroriginal]

Awesome tutorials! The most consistent and detailed tutorial I was able to find so far. I am new to CNNs, so every detail matters.
Mind answering my newbie question? Why do we always use pre-trained weights, even though we are training to recognise completely custom objects, even though we don't care about the pre-trained objects.
In less words, why do we custom-train on top of pre-trained weights?

[Burke]

Super helpful video man! Thanks for taking the time to put this together. I'm trying to use this to make a cat door with a webcam that locks one cat away from a feeder because she needs to go on a diet haha. I'm hoping I can use Yolo to detect what cat is in the frame.


Thanks again!

[DataScienceGarage]

Hi Ivan,
thank you for really useful tutorial. That's impressive. I have one question. On 10:33 you makes changes in batch=64 and subdivisions=8 on lines no. 3 and 4, but didn't uncommented those lines. That is how it should be? No need to remove # at the beginning?

[ghasemabdi]

Hello Ivan: I tried to estimate the anchors by this command "darknet.exe detector calc_anchors data/obj.data -num_of_clusters 6 -width 416 -height 416" as you described. The command was run successfully but without any computation and information as well. on the other hand, I trained the yolov3-tiny on my custom dataset until my avg loss reached to 0.001 and surprisingly found that the trained net could not detect anything. Could you please advise me within these questions? Best, Ghasem

[Mpardis]

Отличные видео! :) Есть пару вопросиков: config файл остается неизменным после обучения сети? Чтобы запустить свою обученную сеть надо взять weights из папки backup(после обучения), старый config файл и запустить их как в 1 уроке? И если нет, то откуда брать эти файлы? Спасибо заранее, продолжай в том же духе ;)

[densaface]

Иван, спасибо большое за видео. Тренирую модель на 30 картинках более суток, но что-то мне подсказывает 5: 6529.853516, 6531.959473 avg loss, rate, 13507.561000 seconds, 320 images
Loaded: 0.009000 seconds - что это плохие результаты и надо что-то менять. Модель взята на базе скринов из 3d игры, там цели подсвечены зеленой рамкой. Или для yolo такое слишком сложно?

[RnanBerbel]

Hi Ivan, thanks for the great video. It helped me a lot. I finally managed to start making a custom object detector. I have a question about this temporary file. Is it generated after every 100 interactions with images or batches?

[mohdanaskhan]

Hi Ivan,
your videos are super helpful. Could you please make a video on like how we can stop training, when should we stop training and how we can resume training?

[omarahmed-gsop]

Hello Ivan, i followed your videos and i think i didn't make any mistake but i can't see the weights file in the backup folder, can you suggest what to do ? thanks in advance!


Learning Rate: 0.001, Momentum: 0.9, Decay: 0.0005
If error occurs - run training with flag: -dont_show
Resizing
608 x 608
Loaded: seconds
Used AVX
Used FMA & AVX2
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 16 Avg (IOU: 0.381670, GIOU: 0.330054), Class: 0.616785, Obj: 0.556268, No Obj: 0.515605, .5R: 0.125000, .75R: count: 16
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 23 Avg (IOU: 0.363459, GIOU: 0.308566), Class: 0.497933, Obj: 0.528091, No Obj: 0.538773, .5R: 0.176471, .75R: count: 34
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 16 Avg (IOU: 0.382751, GIOU: 0.289021), Class: 0.573551, Obj: 0.468930, No Obj: 0.514701, .5R: 0.250000, .75R: 0.062500, count: 16
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 23 Avg (IOU: 0.358248, GIOU: 0.220704), Class: 0.444470, Obj: 0.504094, No Obj: 0.538594, .5R: 0.166667, .75R: count: 24

1: 822.111450, 822.111450 avg loss, rate, 182.774000 seconds, 64 images
Loaded: seconds
Cannot load image data/obj/BBBB.jpg
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 16 Avg (IOU: 0.416553, GIOU: 0.316236), Class: 0.572676, Obj: 0.582136, No Obj: 0.515543, .5R: 0.346154, .75R: 0.115385, count: 26
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 23 Avg (IOU: 0.395552, GIOU: 0.265847), Class: 0.474094, Obj: 0.553545, No Obj: 0.539156, .5R: 0.342105, .75R: 0.026316, count: 38
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 16 Avg (IOU: 0.341640, GIOU: 0.265565), Class: 0.579418, Obj: 0.521651, No Obj: 0.514760, .5R: 0.153846, .75R: count: 26
v3 (mse loss, Normalizer: (iou: 0.750000, cls: Region 23 Avg (IOU: 0.427411, GIOU: 0.339863), Class: 0.518983, Obj: 0.544532, No Obj: 0.538632, .5R: 0.434783, .75R: count: 23

2: 823.878174, 822.288147 avg loss, rate, 370.658000 seconds, 128 images

[Donz]

Hi Ivan, I followed your YOLOv3 series and managed to retrained yolo for custom dataset (love your series btw!). Is there any way or code/command to filter or display the time taken for YOLO to detect objects in real-time input? Using real-time detection, we can only obtain the FPS, class detected, and scores which is also difficult due to it constant updating (with the speed of FPS).

[neliquidmusic]

Иван подскажи пожалуйста каким путем ты шел, чтобы вникнуть во все это? Очень сложно в это погружаться. Ранее дело с нейронными сетями не имел. Может литературу какую..

[Raducki]

hello. really awsome tutorial. boosted up my comprehension and ability to deploy this.

a question : at 15:22 you type 15 as the last param. But as i understand it the yolov3-tiny has 13 [convolutional] layers.
it has 2 in the second [yolo] layer
and 11 in the first [yolo] layer.

Shouldn't we use 13 instead of 15 ?
tx.

[disheensolanki]

heyy the video was really helpful..But i had one doubt ..Once i am done with training what exactly should i do to test the model and make sure whether the object detection is accurate or not

[touilsarra]

@Ivan Goncharov thank you for this tutorials
what's the difference between darknet53 weights and darknet15 weights

[sebastianquesada]

Hey Ivan! As always a great video. I just finish training yolo with my data, but a I have the dumb question, of how I see it working. Like, which will be the command on the cmd to try it out

[bibin]

Hey Ivan, Gr8 Vedio.. appreciate it.. How does the configuration file looks like if i wanted to add new classifier let say "number", also use the old classifier from pre-trained weight let say "Person"