How to Train YOLOv4 on a Custom Dataset in Darknet

A video of how to train YOLO v4 to recognize custom objects in Google Colab in the Darknet framework. In this video we will take the following steps to train our custom detector:

1) Gather and process our dataset
2) Load dataset into Google Colab
3) Build Darknet framework in Google Colab
4) Write custom YOLO v4 training configuration
5) Train custom YOLO v4 detector
6) Use trained YOLO v4 detector for inference
7) Export YOLO v4 weights

Label your images:

Corresponding training blog post:

Colab Notebook:

What is MaP?

[kaibe]

I have to admit, when I first started watching this I thought to myself "ugh, another AWS overpriced solution". But no, this is actually awesome. You've saved me a ton of hours in getting the data ready for our own models - new subscriber AND customer!

[lutzludtke]

Thanks a lot for allowing so many to get their feet wet with visuals.
Just starting of with integrating it into our existing solutions.
QUESTION:
How do I change/reduce the file size of the generated .weigths file?
Which factores do control it and where can I change them?

[jfhbfmm]

Hi I have some problems I use custom data on roboflow when I'm training it have errors path directory not have, how i can fix this problem

[blueming]

now the colab has changed, the code inside the colab notebook provided unable to run completely already. Is there anyone is currently doing yolov4 with this colab notebook but can run completely ?

[BrunoMartins-qvvo]

i got this error "error: undeclared (first use in this function); did you mean when i run "!make". Someone knows how to solve this?

[dribletdrop]

i have taked erorr: "/bin/bash: ./darknet: No such file or directory"

[farahkhan]

can you do age prediction by facial feature recognition using yolo v4

[sevenscapes]

Hi there,
Can Yolo model v4 be trained on mac with m2 chips?

[yxdzicp]

Thanks you share the procedure, .
I've the question is that could I use the weights on video?

[RyanRuddell]

Thanks for the great videos.
you Rock!!!

[rhalppalanca]

very helpful sir for my project. upload more cool videos

[alastrareymart]

Goof afternoon sir...sir I would like to ask on how you installed your pip requirements using google colab, , because I followed your tutorial but I have encounter some error in converting yolo to tflite model

[bintangrahardjoputra]

Hi, thanks for the great explanation. Is there any chance that we can see the accuracy of the model (YOLO v4) after the training process done? thank you

[pallavi]

How to upload the data directly without using your Roboflow app. Because in Roboflow, I cannot split my train, Val, and test sets category-wise if the data is imbalanced. Not sure how your build custom model did not take that into consideration.

[lukaswyon]

@Robotflow Great video! How can i get the coordinates of the predicted boundboxes?

[iposipos]

Thank you for your easy to follow videos:). Please I have a question. Why do we have to pre-structure the train and validation images before running the training? Can't we implement k-fold cross validation on yolo during the training?
Most of the tutorials that i have found in yolo whether v3, v4 or v5 has the train and validation folders already setup. Thanks for your response. Am really curious about this.

[angelinakancheva]

Once we've got our (trained) best weights, how can we use them on unseen data of the same kind just for testing? And can inference be tested on multiple images simultaneously?

[ZuesNgcobo]

I am getting the below error when I have to perform the prediction

Unable to init server: Could not connect: Connection refused

[darksideofthemoon]

Hi! Thank you for this tutorial. Can you tell me how can I view/download my annotation labels in txt format? I used roboflow to generate my dataset and it created txt files for every image in the test directory but they remain empty even after testing my detector. I need the generated darknet format labels for my test set.

[Mahmoudkentut]

Hi! Thank you for the explanation. I have a few questions. When you download the pre-trained weights, is this called a transfer learning? What's the difference between transfer learning and fine-tuning? Is there a way to show the mAP or loss in graph format? Sorry I'm still new to this..