Train YOLOv9 Object Detection on a Custom Dataset | Step by step guide

hello everyone my name is Alan from icomia and today we'll see how you can train a YOLO v9 model on your custom data set first we will quickly take a look at the official GitHub repository and then you will see the step-by-step process on fine-tuning your model using a collab notebook let's check out the official YOLO v9 repository if we go down a little bit we can see the performance comparison on the MS Coco object detection data set so we see the two new release models which are YOLO v9 and Gan and what stands out is that Yol 9 is a current stateof the object detection model and if we compare with yolo V8 it's interesting to see that with similar performance the Euro v9 model is much smaller and let's check out the different sizes which are the s m c and e in this tutorial I will use the euro v9 C version which is a good compromise between the Precision and the model size now let's get into the fun part by jumping straight into the training notebook first let's check out our run time make sure that we are using using GPU so by changing run time and this is correctly selected you can validate that we are using T4 GPU by running the following commands and then the only dependence that you have to install will be ecoma we keep install ecoma so I will run it now here we can see that we have 15 gab of GPU during the installation process we can check out the data set we'll be using it will be the basketball player data set looking at an image we have the labels for the players the ball referees number of points the period and times going back to the collab so it looks like it's installed we just have to restart the session so the next step is to download the data set so directly from roof flow if we refresh we can see that we have the three f folders test train and the validation data set now that we have our data set we're ready to create our training workflow so the first step is to initialize it then we load the data set with the data set Coco algorithm and we add the trend Yow v9 algorithm to the workflow we can set different parameters that like the model name the number of epo batch size frame and test image sizes and also the split ratio between the train data set and the test data set let's run this workflow here I'm missing a comma it will download all the necessary algorithm like data set Coco and train Yow v9 and install all the necessary packages so it might take a few minutes for the first run so the installation took about 5 minutes and now the training is starting with the first EPO so we will be back at the end of the training all right so the model has done training after 45 minutes and the model weights are saved into this folder you can see the metric for the different classes its close are above 0.9 for the different classes which is pretty good for more information about the metrics over time we can check the graphs so they will be under the save directory and results so we have the losses for the trend and the validation also we have the matric for the Precision and recall we can see that we are almost reaching a plateau we could maybe have better function our model with a with a bit more aox but this is definitely good enough for this example now now we can test our model so by running inference first so we have to set the output folder name we copy this one then we create the inference workflow so by initializing adding the inferior L9 task and setting the parameters so by default the algorithm inferi v9 will use thee train Coco model and we will test that first and then we'll use our custom train model so let's run the workflow so here we just added the task and now we have to run on the image let's run this cell as well and here we have the results of the pre-train Coco model with only the class person which is detected so now let's try out our function model and we will oncommand the parameters and let's on the workflow again and here is the output of our custom model you can see that it detected the player the referee the hoop and the scoreboard maybe you can't see the scoreboard and here this is better so this is the end of this tutorial I hope you enjoy this video and I will see you in the next one

Channel: Ikomia

Views: 592

Keywords: YOLOv9, Ikomia, Computer Vision, Yolov7, YOLOv5, Deep Learning

Id: 3g-mw49ZkZM

Length: 4min 52sec (292 seconds)

Published: Fri Mar 01 2024