How to Train YOLOv5 on a Custom Dataset

✅ Subscribe: bit.ly/rf-yt-sub
YOLOv5 is the latest evolution in the YOLO family of object detection models. It's the first YOLO implementation native to PyTorch (rather than Darknet) and emphasizes ease of use and quickness of training and inference. This YOLOv5 tutorial shows you how to train the model on your own dataset in Python.
Tools:
Roboflow - roboflow.com
Public datasets - public.roboflo...
Colab notebook - models.roboflo...
GitHub repo (for additional details, e.g. passing a video file or webcam for inference): github.com/ult...
YOLOv5 Details:
YOLOv4 vs YOLOv5 - blog.roboflow....
How to Train YOLOv5 (blog) - blog.roboflow....
Labeling guides for creating bounding boxes:
CVAT - blog.roboflow....
LabelImg - blog.roboflow....
Explore popular templates to kickstart your project:
roboflow.com/t...
Follow us on Twitter: / roboflowai

[@milesgiehtbrock8510]

Dude thanks for speaking in a way that teaches me instead of just assumes I already know this stuff

[@Roboflow]

Certainly - thanks for tuning in.

[@luqmansen]

I'm about to follow a yolo3 tutorial, and then I found the yolo4 tutorial, and then I found yolo5 tutorial, omg Thanks roboflow.

[@Roboflow]

Thanks! Let us know what you want to see next.

[@A_tavv]

this was the best video in the detection area by YOLO, believe me I tried a bunch of videos, non of them worked like this Thanks truly.

[@raunakmitra7868]

Brilliantly done! You know how to keep the viewers glued to your video.

[@tylersnard]

This worked GREAT for me! The best part was that it showed my training images and their labels, and it made me realize my training set was mis-labeled :)

[@Roboflow]

Great to hear! If you're interested, we'd be happy to share your success on our blog. Write us in-app or here: docs.roboflow.ai/support

[@tylersnard]

@@Roboflow Great, I just did!

[@dianavelychko6703]

this video is so transparent and really easy to understand and follow

[@philippehuet8347]

Great tutorial, maybe the best I have seen on Yolo. Thanks for sharing the Colab notebook

[@kunlunwang2395]

Best YOLO tutorial I've seen on KZbin so far! Thank you!

[@Roboflow]

Thanks so much!

[@user-hc7tp1zs2w]

this guy is still a legend to this day

[@sainandhan1108]

Thanks a ton, I finished my project easily by referring this video

[@Roboflow]

We love to hear that! What was the project that you worked on? :)

[@sainandhan1108]

@@Roboflow UAV Detection in Real Time

[@Roboflow]

@@sainandhan1108 sweet! Is that something you can share? Seems very interesting

[@harivarma3187]

@Sai Nandhan. Bro koncham adhi ravadam ledhu bro, ninnu contact avacha?

[@rezzailham7682]

can u explain what the meaning of "head", "backbone", and "anchors" on the models of yolov5?

[@bbb-th1vu]

i love your facial expressions. thanks. i dont know anything yet but hopefully i am planning to go into this, this summer.

[@tarot-oo8ff]

I liked the shaving part and selfie part ..by the way ..this was so fun learning that hard skill in a funniest way..thankur so much sir..

[@russf]

Apart from the nits, noted below, THANKS for a great video, delivered in such an upbeat way!

[@Roboflow]

Thanks for the feedback!

[@tarot-oo8ff]

Hats off..what a teacher

[@tatianarice2259]

wow, great! thanks for sharing!

[@nadirjackson5755]

When I try to run the train.py script I get the following issue: Traceback (most recent call last): File "train.py", line 492, in train(hyp, opt, device, tb_writer, wandb) File "train.py", line 91, in train model = Model(opt.cfg, ch=3, nc=nc).to(device) # create File "/content/yolov5/models/yolo.py", line 95, in __init__ self._initialize_biases() # only run once File "/content/yolov5/models/yolo.py", line 150, in _initialize_biases b[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) RuntimeError: a view of a leaf Variable that requires grad is being used in an in-place operation. Does anyone know How to Fix This?

[@hackercop]

Sir you are amazing thanks, you make we want to get roboflow Pro! Keep it up!

[@dinarakhaydarova4898]

perrrfectttt

[@nunobartolo2908]

how many trainable parameters does yolo5 have? does your training start with a pre trained model that used the coco dataset? if you use existing weights won't you need to re code the network head to change the classes to your own?

[@delibrain2216]

Hi Joseph .A great tutorial. The dataset you used was heavily unbalanced. The augmentation is performed on all images combined and not the ones that were misrepresented(WBC AND platelets). Don't you thing that the trained model will be biased towards RBC? Could you give some suggestions how to cope with unbalanced dataset? or should I perform augmentation on the least represented classes?

[@snehitvaddi]

Will *weights file* and *config file* generated by Roboflow be acceptable by "dnn" module of OpenCV ie: cv2.dnn.readNetFromDarknet ??

[@yashrunwal5111]

How do i test the model for a video? What should we change in the line 'detect.py'?

[@Roboflow]

Note per the Ultralytics repo, when calling detect.py, source should be updated to "0" github.com/ultralytics/yolov5#inference

[@azerioauditore511]

He is funny af. And obvio the tutorial is awesome, thats very universal of him.

[@ammarfkhan]

from utils.utils import plot_results; plot_results() # plot results.txt as results.png Image(filename='./results.png', width=1000) # view results.png This is not working, I just copied your code, giving me an error "ModuleNotFoundError: No module named 'utils.utils'". Kindly anybody help me asap. Thanks.

[@pedroh.fonseca9753]

Best tutorial ever!! Thankss Joseph!! ;D

[@Roboflow]

Thanks!

[@kostisole]

Amazing work! Congrats! keep up

[@yurikleb]

min 18:20 is bit confusing, the command downloads the files into /content, but you open the /content/yolov5 folder? where do the files should go?

[@tahamansoor7396]

exactly did find the solution

[@francescopaganelli177]

Is it possible detect or recognize different person like different classes? Thanks great job!

[@mrpoeky2527]

Thank you! for the clear explanation :)

[@vendetta1273]

Can you tell how to change the model architecture? Like the layers. I would appreciate that. If you have already made a made for that please leave a link to that. Thanks.

[@JackJGX]

Roboflow is pretty awesome

[@Roboflow]

Thank you!

[@amaltp4893]

Its very nice. How can I download this model to my PC ? Is the weights are sufficient ?? I need the architecture also right ?? I want to make an android app for object detection , any tutorial available on that ? Pardon for any mistake in the question , I am a beginner.

[@21_jtd3f_sarahputritaulada5]

hi, may I ask how to use a "Try this model drop an image" box in a library that already exist? cuz when I drop the image it cant be detected an just shown this command { "predictions": [] }

[@sorrybird1]

Thanks for your awesome sharing. Helps us a lot!

[@hypercloud9689]

In very confused, once you mount the weights to your drive, how can you use it in your code????

[@moahaimen]

i need a full explanation of tensor board measures and understand them please

[@adityasareen6593]

how can i use download .cfg and weights file to use it locally on my system using python and open cv code ??? plzz help

[@sanabatool5573]

Please tell me how these saved weights can be used for prediction

[@alperenkuzhan1407]

Hi, I'm working on a project and I need help with something. How do I find the correct epoch size for my project? Different tutorial videos on your channel have different epoch sizes. How do you determine this?

[@abinayakalyanaraman5982]

Can you please tell me about how you labelled your dataset?

[@abdullateefagbaje1032]

This is a great video. I have a question, what if I want to use TensorFlow instead of PyTorch for YOLOv5 or 4?

[@Readwithaami]

I am a student .. I have used free plan roboflow for annotating data.but I couldn't download the dataset .. Export was not there ..so what should I do

[@JuneLetsPlays]

Trying to do this but getting the following error: Traceback (most recent call last): File "train.py", line 531, in train(hyp, opt, device, tb_writer, wandb) File "train.py", line 191, in train assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1) AssertionError: Label class 1 exceeds nc=1 in ../data.yaml. Possible class labels are 0-0

[@user-mn5ic3zl5h]

!!! awesome! going to test out now

[@mdtokitahmid2970]

Hey! One qus. I have my own images and xml files given and downloaded in my pc. How should I import those to collab?

[@pitiwatkittiwimonchai4656]

" define input image size " 19:47 I don't know what it mean ? can anyone help me explain what is number 416

[@xuxiang9529]

I following all steps you provided, but there is no predictions on the test dataset. I don't know why!!!!

[@iCore7Gaming]

I have my own GPU i can train on, how would i go about doing that? don't want to use external websites with the datasets i have, as they are very limited.

[@DucNguyen-cq5iw]

Hello, thanks for the amazing tutorial. I'm just wondering if we apply resizing for training data, but at inference, we use the original images. Does this have an impact on return?

[@nestorbao2108]

according to glenn jocher, you should use the same size at inference, but i think you can use the resized size for both

[@andrewp494]

How to change the bounding box thickness and label ?I mean if you have multible object you can't see the details in the img

[@PP-fh2mk]

At 18:35 1) If we want to combine our custom set in default which yolov5 given sets is it possible? and how 2) or can we delete some classes that we doesn't needs? I would like to make custom sets with person classes, But if I made custom sets in your videos' method Maybe i think i have to make person dataset again . ( I'd like to add my customs to the existing default datasets. )

[@mayamehta2220]

Thank you Roboflow , thank you Joseph for this Excellent tutorial and colab notebook. I want to create custom dataset for my final year project. Can you please guide what should be image size and how to create dataset? my dataset contains different types of invoices

[@itsshort6121]

I having issue with my dataset. my dataset is video based and it anootation file for bounding box is also given how can I convert it for yolo v5

[@hinainam5037]

Hello sir can you please tell me how to find accuracy???

[@sanabatool5573]

How can i use this saved weights for future prediction

[@faishalichal8934]

Should we use CUDA for training??

[@theminertom11551]

Great content! Thank you for this.

[@namangupta1249]

Hello sir, I have followed the same steps you have Given to train the model, but my MAP is decreasing per epoch. Please let me know how to solve this. Thanks in Advance

[@ChafiaIbtissamkl]

Dude thanks this video was really helpful. Just asking when i wanted to use roboflow to create my own data i didn't found the place where i have to use to generate data is it the new version that's why. Btw thanks again.

[@gguy156]

I honestly don't understand why this model is working in Colab and does not working on my PC. It's always displaying nan loss. But pytorch setup is working well with other network architectures. What can be a problem with it?

[@srushti9694]

Please upload video of auto annotation for custom dataset using roboflow

[@hemanthreddy8991]

Hey can any one say how to pass an image which in my desktop directory to detect and I have one class only

[@Teedies-oh7gi]

Thanks for the video, would like to try this out since the previous model can’t really detect small objects. My plan is to make an app to be able to detect ball and hoop and count field goal, so would like to ask is it possible to transfer all this into an app?

[@medcherifbraham1132]

thanks for the tutorial, but when i train my data it takes up to 5mins to complete one epochs on google colab, any help please ?

[@Team-fh2rr]

Good day! We have some issues with object detection on custom datasets and would be appreciated any advice. Is it possible to have a 1-hours consultation?

[@nicolassy1927]

Is it possible to just augment to the objects that yolov5 can detect already? Like for example I have a dataset of streetview images and I want to detect trees aside from cars and motorcycles. Do I have to retrain yolov5 again to detect cars and motorcycles, when it is already included in the original coco128.yaml file?

[@lukaszsiadaczka]

Great video! Do you plan to upload video with opencv gpu c++ and yolov5?

[@truptimahendrakar3486]

@Roboflow "RuntimeError: non-positive stride is not supported" I get this error when I run the training command. My architecture is the same as the video. Does anyone else get this?

[@victor22227]

Hey Nice video! Is there a way to like export the weights and such to use in a custom python project? I already trained a custom dataset, now I need to combine it with my python code that counts cars as they pass by. There seems to be no videos explaining how to use yolov5 other than on google collab

[@nitishagrawal145]

When I trained with the same, my model with the extension ".pt" is in compressed archive format, how do I get that back to a normal ".pt" file so that i can use it anywhere else? When I try to use the compressed archive directly, it gives an error of core dumped

[@toetoeag]

Great job 👏

[@Clash_Royale_Iranian]

Hi Joseph Is it possible to save the trained model, instead of weights? like TensorFlow "model.save" ??

[@abderrahmenmelliti5164]

Hi, thanks for the tutorial, actually I did the same steps and I used the model on a real-time video but I got only 4 FPS which is too slow , do you have an idea about that ?

[@andywong5193]

How canI save the model and use it? So I don’t need to train again，and then use the model

[@martinolmos1557]

Thanks for sharing, como dijeran en mi México lindo, eres la v3rga!!!!

[@jonnathanwilson5300]

Thank you very much for this great and detailed tutorial. I would like to track the size and position of the object and use it in unity (like handtracking). Is there a tutorial or something alike for the next steps? Thank you already for you help :)

[@sstainba]

I've been trying to use Yolov5 with PyTorch for the past few days... but there are some massive memory leaks during training. I'm constantly having to reboot.

[@linear-inteligencia2390]

amazing video !

[@TC-td7wk]

Hi, my folder 'weights' is not able to get the 'last_yolov5s_results.pt', the unique file contained after running all the code is 'download_weights.sh'. Does anyone know how to fix this issue? Thanks so much in advance.

[@simai2568]

with this method does the model is entirely learn end to end with our dataset or does it use pre-trained models with pre-configure weight ? Also thank you for this clear tutorial

[@dianavelychko6703]

I believe that it is supposed to be training from scratch. He explains it on the moment 20:21, you can do both but the default is to train from null

[@prakulsingla7112]

i have a json file and a zip file of yolo format. it contains txt files of all the images. which one should I use

[@tahaakgungor1026]

thank you for video and ı want to ask how can I find this trained model on mine local gpu on pycharm or jupyter

[@vialomur__vialomur5682]

Thank you so much!

[@QuickBrain]

Could you please share how to run yolov5 trained using python on iOS and Android using coreml or pytorch mobile?

[@hamzehshehab6167]

How can it be used to detect a person running in a stadium i mean how can i use yolov5 in activity recognition?

[@salmashahid7612]

Can anyone guide me on how can I extract the predicted boundary box image and save that image in a folder? I used the same Yolo v5 custom dataset training strategy.

[@huyle44346]

How to connect plc with roboflow software?

[@fakrulislam3140]

# this is the YAML file Roboflow wrote for us that we're loading into this notebook with our data %cat data.yaml I do not see the file and get the "error cat: data.yaml: No such file or directory"

[@tranmani]

same problem

[@anthonynguyen6293]

Thanks for making this video, helps a lot! If I wanted to train the model on the 2d object detection KITTI dataset, how can I get the annotations in the right format? Also, if the model expects the image to be square, does roboflow adjust the bounding box accordingly?

[@Roboflow]

Have you struggled to upload your data? Write us an email (see our docs for how). We do support 2D object detection bounding boxes, like those found here in the YOLO format: github.com/keshik6/KITTI-2d-object-detection Yes, Roboflow updates all bounding boxes correctly for any operation you perform.

[@jagannathbehera9483]

Thanks Roboflow and Joseph for wonderful tutorial, but im struggling to convert best.pt to .onnx file as all tutorials using .pth extension and torch.load() failing for .pt file, please help me on this

[@RicardoAntonello]

Amazing! Congratulations!!!

[@Roboflow]

Thanks!

[@dieweltentdecker5878]

How can I use the weights?

[@Roboflow]

You can download the weights (as our notebook does at its conclusion) and reuse them elsewhere.

[@dieweltentdecker5878]

Roboflow How can I convert it to weights?

[@diegoG21]

sadly the collab notebook doesn't work anymore :c

[@backdoorcynet1344]

Hi, it is a great project. Can yolov5 work on mobile devices or run live webcam on google colab?

[@nurhanisahazmi9636]

I encountered some problem while executing this line: %%time %cd /content/yolov5/ !python train.py --img 416 --batch 16 --epochs 100 --data '../data.yaml' --cfg ./models/custom_yolov5s.yaml --weights '' --name yolov5s_results --nosave --cache and it comes out like this: /content/yolov5 Traceback (most recent call last): File "train.py", line 481, in opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp) # check files File "/content/yolov5/utils/general.py", line 70, in check_file assert len(files), 'File Not Found: %s' % file # assert file was found AssertionError: File Not Found: ./models/custom_yolov5s.yaml CPU times: user 8.98 ms, sys: 6.35 ms, total: 15.3 ms Wall time: 1.92 s I have no idea what went wrong, can you help me :(

[@tahamansoor7396]

same

[@tahamansoor7396]

@Sulastri Tambunan M1141447 yes its just a file path issue.

[@rozhinataghdimi335]

@@tahamansoor7396 how did you fix it?

[@tahamansoor7396]

check the model path.

[@hareshlad6067]

Thanks for making this video,Can you make a dataset on Garbage detection?

[@kayalalakshmiswetha9527]

I am not getting test train val folders in colab after entering the key in colab.. Can anyone help me with this...and I am also unable to get testvalues file for generating ground images...I need help

[@Roboflow]

Hi 👋🏻 could you create new topic here: github.com/roboflow/notebooks/discussions/categories/q-a Describe your problem. I’ll try to help you.

[@dyaakhaled9219]

Many thanks for this 🙏 But how to deploy it to my own website?

[@rakeshswain1880]

First of all, really a great and fun video thanks for sharing. I was able to train with one of my custom datasets. Very simple and easy to use notebook too!! I was just wondering what format of annotation does the Pytorch V5 model accept for bounding boxes? Is it still center coordinate, width and height?

[@Roboflow]

The YOLO v5 PyTorch implementation uses .txt annotations in-line with the YOLO standard and adds a YAML config file to define train and valid.

[@rakeshswain1880]

@@Roboflow - Thankyou.