[rewrite] rewrite all repo to properly use DVC, MLFlow and data engineering practices

[Chouffe]

Description of the PR:

• Rewrite from scratch of the code of the repo following best practices and making model training reproducible.
• Add random hyperparameter search and instructions in the README file
• Move all data assets to DVC in an S3 bucket - added instructions in the README file
• Only the baselines and the best models are stored with DVC and are a dvc pull away

How to test:

Follow the README file to test the repo and dvc setup.

I just copied and pasted all code from the repo developed here: https://github.com/earthtoolsmaker/pyronear-mlops

[MateoLostanlen]

Hi @Chouffe,

Thanks a lot Arthur for this amazing work !

I left you a comment on something that's buggy, but the rest is fine with me.

Just one thing, I'm not sure about the relevance of having all three stages:

• train_yolov8_baseline_small_dataset
• train_yolov8_baseline_full_dataset
• train_yolov8_best

For me, the first two are simple tests. You can explain how to launch them in the README, but I don't think they should be there. I suggest keeping train_yolov8_best and renaming it train_yolov8. What do you think?

[MateoLostanlen]

Got an error here :

(.venv) mateo@mateo:~/pyronear/vision/mlops/pyro-mlops$ make yolov8_benchmark
python ./scripts/model/yolov8/benchmark.py \
  --input-dir ./data/04_models/yolov8/ \
  --output-dir ./data/06_reporting/yolov8/ \
  --loglevel "info"
INFO:root:{'input_dir': PosixPath('data/04_models/yolov8'), 'output_dir': PosixPath('data/06_reporting/yolov8'), 'loglevel': 'info'}
INFO:root:Loading files data/04_models/yolov8/best/args.yaml and data/04_models/yolov8/best/results.csv
/home/mateo/pyronear/vision/mlops/pyro-mlops/./scripts/model/yolov8/benchmark.py:77: PerformanceWarning: DataFrame is highly fragmented.  This is usually the result of calling `frame.insert` many times, which has poor performance.  Consider joining all columns at once using pd.concat(axis=1) instead. To get a de-fragmented frame, use `newframe = frame.copy()`
  df[k] = str(args[k])
/home/mateo/pyronear/vision/mlops/pyro-mlops/./scripts/model/yolov8/benchmark.py:77: PerformanceWarning: DataFrame is highly fragmented.  This is usually the result of calling `frame.insert` many times, which has poor performance.  Consider joining all columns at once using pd.concat(axis=1) instead. To get a de-fragmented frame, use `newframe = frame.copy()`
  df[k] = str(args[k])

[Chouffe]

Is it only log warning?
If that's the case I would not mind TBH. Or we could spend some time optimizing the dataframe operations to comply with the warning?

[Chouffe]

Hi @Chouffe,

Thanks a lot Arthur for this amazing work !

I left you a comment on something that's buggy, but the rest is fine with me.

Just one thing, I'm not sure about the relevance of having all three stages:

• train_yolov8_baseline_small_dataset
• train_yolov8_baseline_full_dataset
• train_yolov8_best

For me, the first two are simple tests. You can explain how to launch them in the README, but I don't think they should be there. I suggest keeping train_yolov8_best and renaming it train_yolov8. What do you think?

Thanks for taking the time to review this @MateoLostanlen.
I left a comment on the log warning issue you mentioned.

The small stages were used to tighten the feedback loop to training yolo models with a smaller dataset. That usually saves me a lot of time instead of burning GPU compute for something that is not sure to work well.

I think we can get rid of them if you think that's confusing or the other options would be to add some comments?
Let me know what you prefer and I can take care of it.

[Chouffe]

@MateoLostanlen I have been working directly on my github repo here to add hyperparameter search for yolov9 and yolo10.
Happy to open a PR once we get this one merged in.