Add object detection training docs #1435
Conversation
| ```bash | ||
| git clone https://github.com/ultralytics/yolov5.git | ||
| git clone https://github.com/airockchip/yolov5.git airockchip-yolov5 | ||
| wget https://gist.githubusercontent.com/Alex-idk/9a512ca7bd263892ff6991a856f1a458/raw/e8c0c9d8d5a1a60a2bbe72c065e04a261300baac/onnx2rknn.py # This is the onnx to rknn convertion script |
There was a problem hiding this comment.
Script looks reasonable, seems like the same thing most online resources have been hinting at.
Do we wanna pull this into the photonvision repo proper?
There was a problem hiding this comment.
I think this script should be owned by the PV organization. Once we start supporting more platforms it can be updated to go from a single .onnx file to all of our supported platform.
There was a problem hiding this comment.
As part of this PR, commit this file to our scripts folder or something in the monorepo.
There was a problem hiding this comment.
It is inside the scripts folder, however I did update the script but haven't tested it yet.
|
|
||
| #### Training Command | ||
|
|
||
| Please research what each of these parameters do and adjust them to fit your dataset and training needs. Make sure to change the number of classes in the `models/yolov5s.yaml` file to how many classes are in your dataset, otherwise you will run into problems with class labeling. Currently as of `September 2024` only YOLOv5s models have been tested. |
There was a problem hiding this comment.
"you must know what you are doing" lolz.
Is there anything major you could recommend an initial user look at, just to get them started? Things like "Be careful touching these parameters" or "These parameters are inter-linked" or "Here's the values we started with for the note detector" or similar?
There was a problem hiding this comment.
The main parameters for training are dependent on your dataset and hardware. I can add a link to a docs page that explains what they do. But I'm also assuming that the people who are making a custom model already have an idea of what they are doing. The special thing about doing it for PV is the conversion to RKNN.
| Please research what each of these parameters do and adjust them to fit your dataset and training needs. Make sure to change the number of classes in the `models/yolov5s.yaml` file to how many classes are in your dataset, otherwise you will run into problems with class labeling. Currently as of `September 2024` only YOLOv5s models have been tested. | ||
|
|
||
| ```bash | ||
| python train.py --img 640 --batch 16 --epochs 10 --data path/to/dataset/data.yaml --cfg 'models/yolov5s.yaml' --weights '' --cache |
There was a problem hiding this comment.
Might still need a concrete example of what data.yaml is
There was a problem hiding this comment.
It might just be best to point people towards some other resources for preparing yolov5 datasets.
| #### Export Command | ||
|
|
||
| ```bash | ||
| cd /path/to/airockchip-yolov5 && python export.py --weights '/path/to/best.pt' --rknpu --include 'onnx' |
There was a problem hiding this comment.
If the instructions above were used, the path should just be airockchip-yolov5 , right?
There was a problem hiding this comment.
Rather than including this as a separate step could it be merged into the current conversion command?
My thought is people who want to share models (say the PV devs, or community members) could benefit from a single script that goes from the weights to all of our supported platforms.
| python onnx2rknn.py /path/to/best.onnx /path/to/export/best.rknn /path/to/imagePaths.txt | ||
| ``` | ||
|
|
||
| If you have any questions about this process feel free to mention `alex_idk` in the PhotonVision Discord server. |
There was a problem hiding this comment.
A wise man once told me to never put my name on anything :)
|
I'm going to try following these instructions and see if I can get it to work 👍 |
|
I will boot up a vm later and try this process from scratch and look for anything I missed. |
| Please research what each of these parameters do and adjust them to fit your dataset and training needs. Make sure to change the number of classes in the `models/yolov5s.yaml` file to how many classes are in your dataset, otherwise you will run into problems with class labeling. Currently as of `September 2024` only YOLOv5s models have been tested. | ||
|
|
||
| ```bash | ||
| python train.py --img 640 --batch 16 --epochs 10 --data path/to/dataset/data.yaml --cfg 'models/yolov5s.yaml' --weights '' --cache |
There was a problem hiding this comment.
It might just be best to point people towards some other resources for preparing yolov5 datasets.
| #### Export Command | ||
|
|
||
| ```bash | ||
| cd /path/to/airockchip-yolov5 && python export.py --weights '/path/to/best.pt' --rknpu --include 'onnx' |
There was a problem hiding this comment.
Rather than including this as a separate step could it be merged into the current conversion command?
My thought is people who want to share models (say the PV devs, or community members) could benefit from a single script that goes from the weights to all of our supported platforms.
|
|
||
| #### Conversion Command | ||
|
|
||
| Run the script, passing in the ONNX model and a text file containing paths to images from your dataset: |
There was a problem hiding this comment.
and a text file containing paths to images from your dataset:
So we've got to create this file ourselves?
There was a problem hiding this comment.
Yeah, I can probably make a revision of the conversion script to just take a path and make the file itself.
There was a problem hiding this comment.
find /path -type f -name "*.jpg" > images.txt
|
I was able to follow the instructions and got the model running on an Orange PI. I'll verify tonight if the model is working. As for the state of this PR - I think this is a huge help for the PV developers. I'm no longer concerned about the bus factor regarding training object detection models. However, I don't think these instructions would be enough for the average team. I wonder if we can just change the name of this content from "training a custom model" to "converting trained model to RKNN" for now. I think a proper "training custom model" section would require teaching how to set up an environment, installing all of the dependencies, how to label data, export to YOLO formats, training parameters, and only then conversion and uploading. |
|
Yeah broadly agree with @Alextopher -- it's a great set of docs for developers, and as long as we call that out i'd like to get this shipped before it bitrots :) |
|
Sorry for the late response, I've been busy with school and theater. Here is a revised version of the docs that removes the model training part, I will work on a update for the conversion script to just take in a directory instead of a file with paths and clean it up a bit. Is the scripts folder a good place for the conversion script to go in the PV repo? |
|
Sounds legit |
|
Any chance we can get this shoved in for beta (maybe tentatively this weekend?) |
|
Is this ready? |
|
Adds documentation for training a custom model.