How to Train YOLO v4 Tiny (Darknet) on a Custom Dataset

YOLOv4-tiny is smaller version of YOLO v4 that emphasizes speed in model predictions, which is perfect for limited compute environments (even CPUs) like mobile phones or embedded machine learning.
This YOLO v4 tiny tutorial breaks down what YOLOv4-tiny is, preparing labeled bounding box data for object detection, training a YOLO v4 tiny Darknet model with free resources on Google Colab on your own dataset, and using that model to perform inference. In this example, we train our example model on infrared thermal images.
✅ Subscribe: bit.ly/rf-yt-sub
Tools:
Roboflow - roboflow.com
Public datasets - public.roboflo...
Colab notebook - models.roboflo...
Blog post - blog.roboflow....
Create Labeled Data
CVAT - blog.roboflow....
Follow us on Twitter: / roboflowai

[@asdg199]

I know its already been mentioned below. But if you get this error: CUDA Error: no kernel image is available for execution on the device CUDA Error: no kernel image is available for execution on the device: File exists darknet: ./src/utils.c:325: error: Assertion `0' failed. you can try to change the -gencode arch and code. What I didn't realise was that one line above that comment the guy who wrote the program put this: # compute_30, sm_30 for Tesla K80 # compute_75, sm_75 for Tesla T4 so yeah it was a bit embarrassing that i didn't see that. Hope this helps you

[@ironpopcorn6471]

If you're running the "Train Custom YOLOv4 Detector" step and it doesn't work due to (CUDA Error: no kernel image is available for execution on the device CUDA Error: no kernel image is available for execution on the device: File exists darknet: ./src/utils.c:325: error: Assertion `0' failed.) The architecture settings are probably not compatible with google colabs K80 GPU. Just add ARCH= -gencode arch=compute_37, code=sm_37 to the make file and you're good to go. github.com/pjreddie/darknet/issues/98

[@GlentoranMark]

Could you post your line of code for this? I get the same error whatever I try. I've been given the Tesla T4 so tried (among many combinations): !sed -i 's/ARCH= -gencode arch=compute_75,code=sm_75/ARCH= -gencode arch=compute_75,code=sm_75/g' Makefile Same error.

[@noorfaleh82]

I have same error , I thought it’s depending on cuda on nvidia gpu I have amd !!!

[@hessaaleissa1460]

At “set up training file directions for custom dataset” there is error “No such file or directory ‘/content/darknet/‘/content” how can i fix that error?

[@samvandhapathak2167]

Thanks for the tutorial. I finally got the YOLO v4- tiny working with idea from your videos. Now with the saved weights I want to use my Raspberry Pi to detect the same in real time for a object detection project. Please guide me through. Any help will be appreciated. Thanks

[@mustafagoksu7661]

Have you been able to run these resulting weights in raspbery?

[@samvandhapathak2167]

@@mustafagoksu7661 yes I implemented this on raspberry pi 4 4gb using tensor flow rather than suggested tensor flow lite. However the fps is around 1.5 only

[@mustafagoksu7661]

@@samvandhapathak2167 very beautiful. Can you explain how to do it? What should I do with these weights? What should I do with these weights?

[@zeze3949]

Hi , can you post the tutorial you followed for that plz

[@daffaaminuddin8580]

@@mustafagoksu7661 where i can get the saved weights file after trained? drive?

[@you3667]

I solved "CUDA Error: no kernel image is available for execution on the device: File exists". 1. Check the GPU version. Enter the "!nvidia-smi", you can see the GPU version ex) | 0 Tesla K80 Off | 00000000:00:04.0 Off | 0 | | N/A 53C P8 31W / 149W | 0MiB / 11441MiB | 0% Default | you can see the GPU version K80. 2. Change the compute_capability. ex) if your version is K80, execute "%env compute_capability=30" else version is T4, execute "%env compute_capability=75" I don't know exactly. But sometime this kind of situation occurs. "Yesterday version was P100, but today version is K60." So make sure check the GPU version 3. Change the "!sed" part. ex) %cd darknet/ !sed -i 's/OPENCV=0/OPENCV=1/g' Makefile !sed -i 's/GPU=0/GPU=1/g' Makefile !sed -i 's/CUDNN=0/CUDNN=1/g' Makefile !sed -i "s/ARCH= -gencode arch=compute_60,code=sm_60/ARCH= -gencode arch=compute_(change this),code=sm_( change this)/g" Makefile !make ex) if your version is K80, chage this part ---> !sed -i "s/ARCH= -gencode arch=compute_60,code=sm_60/ARCH= -gencode arch=compute_30,code=sm_30/g" Makefile

[@mehmetkaanergul8641]

YOU ARE AN AMAZING PERSON THANK YOU VERY MUCH

[@you3667]

@@mehmetkaanergul8641 ^^

[@gaganagavesh3679]

Does the GPU version change time to time. With Tesla P4 I have above error and I still couldn't fix it...Please help.

[@gaganagavesh3679]

Fixed it, Thank you very much for the solution.....😍😍😍

[@noorfaleh82]

@@gaganagavesh3679 can you tell me how fix it ?

[@eswar9255]

Now there is no Darknet executable file. what can I do? from where can I get that file?

[@__________________________6910]

I love how you explain and your talking style is very fine love this

[@musabahmed567]

I'm getting this error: "Makefile:162: recipe for target 'obj/convolutional_layer.o' failed make: *** [obj/convolutional_layer.o] Error 1". Can anybody please help?

[@Roboflow]

Hi! Could you post a bug report with all the details on our notebooks repository: github.com/roboflow-ai/notebooks

[@strizerz3514]

I am getting a new error with darknet - "/bin/bash: ./darknet: No such file or directory". I have been stuck on this for hours, Do you have any solutions? Thank you!

[@rifqyfachrizi7891]

are you solve this problem?

[@strizerz3514]

@@rifqyfachrizi7891 no

[@ballajaisheel1120]

@@strizerz3514 same error...if you got it just reply me

[@samvandhapathak2167]

At the end of the video you switch to a model that is already trained and it can be clearly seen that the colab page you were on also shows a CUDA error. Please suggest ways to correct that. Great video and also love roboflow for data prep

[@Roboflow]

Thanks for your kind words! Yes, the CUDA configuration will depend on what GPU you're using in Colab. The notebook assumes a P100 (and this is the one I switch to at the end). The prior is on Colab free and uses a K80. The CUDA config file in the notebook allows this to be modified to your hardware.

[@samvandhapathak2167]

@@Roboflow while I am using the free version with K80, what changes should i make to run the training. Tried a few things didn't work☹️

[@junaidkhan-uc2eb]

Were you able to resolve this issue?

[@noorfaleh82]

@@junaidkhan-uc2eb I have the same error

[@ArunMPEdison]

Great video! Thank you for sharing, curiously watching for which tools you have used for custom training from scratch.

[@noorfaleh82]

How can download the complete file with weights and cfg file after training it on google colab to use on GPU kit ?

[@vaperon]

Hello, i tried the make same example work but couldnt managed to get your example work. Getting this error: CUDA status Error: file: ./src/blas_kernels.cu : () : line: 841 : build time: Aug 10 2020 - 14:41:39 CUDA Error: no kernel image is available for execution on the device CUDA Error: no kernel image is available for execution on the device: File exists darknet: ./src/utils.c:325: error: Assertion `0' failed. Any solution for this?

[@samaylakhani7959]

Check the !sed command on top of the notebook. It’s built for P100 GPUs. Uncomment the K80 or T4 if you’re on Google Colab. If you’re using another GPU check the other comments.

[@mehmetkaanergul8641]

I solved "CUDA Error: no kernel image is available for execution on the device: File exists". 1. Check the GPU version. Enter the "!nvidia-smi", you can see the GPU version ex) | 0 Tesla K80 Off | 00000000:00:04.0 Off | 0 | | N/A 53C P8 31W / 149W | 0MiB / 11441MiB | 0% Default | you can see the GPU version K80. 2. Change the compute_capability. ex) if your version is K80, execute "%env compute_capability=30" else version is T4, execute "%env compute_capability=75" I don't know exactly. But sometime this kind of situation occurs. "Yesterday version was P100, but today version is K60." So make sure check the GPU version 3. Change the "!sed" part. ex) %cd darknet/ !sed -i 's/OPENCV=0/OPENCV=1/g' Makefile !sed -i 's/GPU=0/GPU=1/g' Makefile !sed -i 's/CUDNN=0/CUDNN=1/g' Makefile !sed -i "s/ARCH= -gencode arch=compute_60,code=sm_60/ARCH= -gencode arch=compute_(change this),code=sm_( change this)/g" Makefile !make ex) if your version is K80, chage this part ---> !sed -i "s/ARCH= -gencode arch=compute_60,code=sm_60/ARCH= -gencode arch=compute_30,code=sm_30/g" Makefile

[@victoravila7305]

@@mehmetkaanergul8641 It worked! Thank you, bro!

[@tiwill345]

@@mehmetkaanergul8641 thx ! this helped me !

[@noorfaleh82]

@@mehmetkaanergul8641 I check not work the same error appear 😢

[@sureshkumar0707]

How to plot the learning graph for this Project?

[@TheBradleydwyer]

I don't know if the energy level one-uppance is sustainable. But I'm excited to find out...

[@Roboflow]

Stay tuned!

[@brili2065]

I finished these steps. So, how to transplant the training model in colab to jetson nano platform?

[@nikhilshashidhar7195]

Can I use YOLOv4-Tiny generated weights and use those weights on a scaled v4 algorithm?

[@Huds-ux1xb]

Hey i cant use yolov4 with darknet , can you give me any tutorial

[@CoxopopYance]

Do you have a .py to use that custom pre trained dataset in my desktop??? I want it to use with my video camera

[@marcialzipaganjr.681]

@Roboflow how should I generate a tflite version of the model I just trained? Also, if I wanted to improve the model again what should I input with the config file? Example the model is initially trained on 2 classes and after sometime I wanted to try it out with 4 more new classes? Should I input 4 on the number of classes or 6?

[@Iamanandraj]

can you tell me which software you are using for screen recording and virtual webcam.

[@syedmuslim4532]

hey can you help me out i am stuck in your video in custom training config for yolo v4 the cfg file have you written it or have copied it from somewhere it'll a great help if you respond to make question!! waiting for your reply.. the timing around 19:20

[@triggerheaven]

Can you make video for Yolov7?

[@avirupdey7565]

At 20:15 you have got a CUDA error, exactly same as mine. Surprisingly, you just ignored that and headed to another notebook. How to fix that error? Pls reply.

[@shahrinnakkhatra2857]

what if i want to do augmentation for specific classes only? ( I want to use SMOTE to balance my dataset)

[@karlboyefio8121]

Hello, when at the ""Train Custom YOLOv4 Detector" stage, I get this response; "/bin/bash: ./darknet: No such file or directory" . What do I do to solve this?

[@Roboflow]

There is an open issue in our repository: github.com/roboflow/notebooks/issues/28 for now we don't have fix for that issue unfortunately :/

[@malikbilal6061]

If its possible to test Yolov4-Tiny Object detection Model on Arduino Nano .

[@preeyank8470]

I am getting cuda error everytime during training if I get K80 or T4 gpu. Can anyone help?

[@GlentoranMark]

The Colab Notebook is different from the video, can anyone explain what changes are needed? I found I was using the T4, is this code correct ? (I'm not too familiar with Python) #install environment from the Makefile %cd darknet/ # compute_30, sm_30 for Tesla K80 # compute_75, sm_75 for Tesla T4 !sed -i 's/ARCH= -gencode arch=compute_75,code=sm_75/g' Makefile !make

[@fixitmoditbreakit7982]

Got the same issue

[@GlentoranMark]

@@fixitmoditbreakit7982 I'm trying again on the K80 and nothing works, same errors no matter what I try.

[@fixitmoditbreakit7982]

GlentoranMark I’ve tried lots of different things. One thing I did notice is that the !sed-i’s line is red once uncommented!!

[@fixitmoditbreakit7982]

@@GlentoranMark hey i just set it up with out copying the notebook and it works!! the files get saved in darknet backup. i haven't tested yet but the tests it runs at the end worked fine . also you get the pc100 so dont have to edit that part

[@fixitmoditbreakit7982]

it works but when trying to train again i didnt get the pc100

[@anime_on_data7594]

how can we try to run on video. what command we should give for it?

[@nouhabenabdelljelil3271]

Hello there , I want to crop what is inside the bounding box and save it in a file in order to apply tesseract later(OCR).Can anyone help me plz?

[@juancamilocetinaberdugo4255]

hi, were you able to fix it?

[@akmalyafi1470]

hey your colab code is different from this video...

[@akashanil4540]

can we use inference on an vid by replacing img path with a video path

[@mustafagoksu7661]

How can we run these obtained weights in raspbery ?

[@elrayyes_leb8872]

darknet is no longer supported on colab

[@tarangini22]

How to inference using videos?

[@__________________________6910]

you are best

[@JoelPrabhod]

What are the tiny yolo weight sizes?

[@jamilahmarramduarte1575]

Is it okay to run this without gpu?

[@Spence1976]

Hello Bear :)

[@freecode.ai-]

I trained my own dataset and it works great! The weights are 22mb. Is there a way to reduce the size of the weights file? I can afford to reduce accuracy if I need to. Thanks

[@nabsmyfoot5308]

where do i download the trained weight file and the config file after training

[@Kaalkian]

is there python opencv script running tiny yolov4?

[@furqanguru7968]

You did not train it on custom dataset

[@_XY_]

Nice

[@selimtasdemir]

can you tell me pls how can i create our own dataset

[@Roboflow]

Yup! kzbin.info/www/bejne/hX7KgmWAr7R0jaM

[@TomattoTomatto]

/bin/bash: ./darknet: No such file or directory

[@ballajaisheel1120]

same error ...help me to go through it

[@tliu9896]

Not free

[@yoloyolo6310]

you need to pay money to export photos to google colab. it is trash.