Multilabel classification NN solution for Amazon forests satellite image. I made accent on "industrial quality" code with next technologies:
- pytorch_lightning
- timm
- ClearML
- linters (black, isort, nbstripout, flake8)
- types with pydantic
- DVC for local usage
Disclaimers:
- the project was originally crated and maintained in GitLab local instance, some repo functionality may be unenviable
- the project was created by me and me only as part of the CVRocket professional development course
- here are a short trained version of NNs (about 15 epochs each)
- this project is my first "industry grade" NN, for more advanced code and features please see car-plate projects
Include 40479 tiles of satellite image in jpg and 17 image types. More information about data you can find in notebook.
To download dataset from kaggle into dataset folder:
make download_dataset-
Create and activate python environment
python3 -m venv venv . venv/bin/activate -
Install libraries
make install
-
Run linters
make lint
-
Tune config.yaml
-
Train
make train- Inference example in notebook
- Best experiment in ClearML
- History of experiments