How to train a custom YOLOv5 model using CodeProject.AI Training for YoloV5 6.2 module
- Thread starter MikeLud1
- Start date
Then I suppose you could break each image into smaller fixed images. So maybe 18 images that are each 640 × 720. You'd have to get a batch program for cropping and first run it for a crop of (0,0) to (639,719). Then for (640,0) to (1279,719), etc. Each time you move over 640 pixels in x-direction. After six crops, move down 720 pixels and restart from the left at (0,720) to (639,1439). Do 6 more, then move down to do the last row of 6.
Edit: Ultimately, you'll need to run all these images through the ALPR module. Do you have a plan of how to do that automatically? And what are you doing with the results?
Last edited:
How was the license plate number of the vehicle shown in this photograph determined?
I've been thinking of trying to create some custom models for testing but I don't have any GPU cards (all integrated with the CPU). I don't play games or need anything that would need something high powered so that's the reason.
I was thinking of trying Google Colab to run it based on some of what I was reading and videos I watched. I looked and they use Nvidia Tesla T4 for as their base GPU option. There are options to pay more for more powerful GPUs but I'm not sure I understand their pricing for "compute units". I found one site that said it roughly equates to 1 units per hour.
I have someone local selling a T4 for $500. That's high for me but willing to sacrifice if it pays for itself over time.
Is the T4 for good for this and worth purchasing? I'm just wondering if cheaper in the long run to buy the T4 than use Google Colab. Plus I don't have to worry about losing my session with Colab. I know the T4 is old, not the fastest, and focuses on efficiency (like a Honda vs a Corvette) but I'm OK with that based on my funds. Trying to find something more powerful with higher VRAM like a RTX 3090 at least doubles the price if not 3x to 4x more and even those are hard to find now a days.
My understanding too was that the more VRAM the better like the 3090 vs other 30 series GPUs which have less VRAM.
Bottom line / TLDR:
Is the T4 for good for this and worth purchasing for $500? I know it needs a custom bracket for a fan too since it's meant for a data center rack but those are like $20 I think.
FYI - my most powerful PC (sadly) is only an i7-9700K with 64 GB of RAM. I'm assuming that is OK since the GPU is most important but let me know if not.
Thanks.
EDIT: forgot to mention that the reason for looking at Colab over CPAI in first place was lack of GPU. If getting the T4 then I can use CPAI and/or other options as well.
I was thinking of trying Google Colab to run it based on some of what I was reading and videos I watched. I looked and they use Nvidia Tesla T4 for as their base GPU option. There are options to pay more for more powerful GPUs but I'm not sure I understand their pricing for "compute units". I found one site that said it roughly equates to 1 units per hour.
I have someone local selling a T4 for $500. That's high for me but willing to sacrifice if it pays for itself over time.
Is the T4 for good for this and worth purchasing? I'm just wondering if cheaper in the long run to buy the T4 than use Google Colab. Plus I don't have to worry about losing my session with Colab. I know the T4 is old, not the fastest, and focuses on efficiency (like a Honda vs a Corvette) but I'm OK with that based on my funds. Trying to find something more powerful with higher VRAM like a RTX 3090 at least doubles the price if not 3x to 4x more and even those are hard to find now a days.
My understanding too was that the more VRAM the better like the 3090 vs other 30 series GPUs which have less VRAM.
Bottom line / TLDR:
Is the T4 for good for this and worth purchasing for $500? I know it needs a custom bracket for a fan too since it's meant for a data center rack but those are like $20 I think.
FYI - my most powerful PC (sadly) is only an i7-9700K with 64 GB of RAM. I'm assuming that is OK since the GPU is most important but let me know if not.
Thanks.
EDIT: forgot to mention that the reason for looking at Colab over CPAI in first place was lack of GPU. If getting the T4 then I can use CPAI and/or other options as well.
Last edited:
That's a training data set, meaning someone had to go through each image and tag objects that were important to their training objective. In this case, the person is putting rectangles around cars, license plates, etc. but ignoring street signs, lamp posts, buildings, etc. There are some tools that can help automate the process (such as using a prior model already trained on similar objects), but generally it's a largely manual process. Once you have sufficient images in your dataset, then you can train a model to hopefully do the same thing on other images automatically. Also, in that dataset, it's not being trained to recognize the contents of the license plate, just where one is in the picture. That's the first step to ALPR, but it misses the actual OCR to read the characters.
Personally, I would try using Collab on a test project and see how far you can take it. I think it's better to first determine if you can make a training data set and get some results using someone else's hardware first. Or if you have a low end GPU (for example I started with a 1050Ti), see what you can do. You may find for your model sizes and the amount of data, you can easily start something at the end of the day and come back in the morning and have a trained model.
Is it possible to use already trained system on these photos in "Training for YoloV5 6.2"? If yes, how to do it?
I'm talking about Open Images Dataset V7
What is it you are trying to accomplish? That's a set of data that is ready to be used for training. But it has more than just vehicle license plates so you'd have to go through and throw out objects that you didn't want. The result would be a trained model that could detect when an image contains a possible plate. But you can already use the existing license-plate.pt model that @MikeLud1 created which was trained on both day plates and night plates. What is it you hope to accomplish by training your own custom model on this data?
I'm a newbie, I have stupid questions for now, sorry.
What's wrong with Object Detection detecting other/extra images? As far as I understand, License Plate Reader will only use its own detect class.
Nothing is detected for me right now. I had an idea to take a model from Open Images Dataset V7.
Thanks. I may still end up doing that.
Also, by "custom" I meant taking the working datasets from others and combining only the classes I wanted.
If you are having issues getting objects detected then this is not the correct thread for that. You should probably create a new thread.
In the new thread state what you are trying to do exactly and then what part is not working.
For example:
Are these live video cameras feeding CPAI or are you trying to run existing pictures from a video feed and run it through it CPAI?
Are you trying to find out if there was just a plate or are you trying to read the plate as well?
If live video then I will tell you right now your camera is too far away for CPAI to read plates as is. If you cannot move the camera then you will have to do it manually or come up with a custom solution as VideoDad said.
No one can help you if you don't state the basics. And again it probably should be in a new thread.