This was a fun little side project to learn a bit of Keras (coming from a pytorch background) and train a boat load of models.
The task is the classic image single-label classification scenario where we have 9 sets of weed variants and need to predict which, if any, are present in a picture. A couple of challenges with this dataset were 1. image quality is quite poor and has bad leakage with background objects / camera rotation etc, and 2. there is a big imbalance towards the negative class (class with no weeds). Have a look at the notebooks folder for the light EDA workings.
The data can be accessed via tensorflow datasets - https://www.tensorflow.org/datasets/catalog/deep_weeds
Please note that we use git lfs for training artifacts (anything under outputs/ or multirun/), so before running make sure you have git-lfs installed, and the artifacts cloned locally.
git lfs fetch --all
git lfs pull- Install
uvas this is how we will manage the project dependencies
brew install uvor
curl -LsSf https://astral.sh/uv/install.sh | sh- Create a local environment
uv sync --frozen --compile-bytecode-
Manually run
notebooks/01_data_exploration.ipynbto generate our local dataset -
Now you're ready to run experiments, try changing some model or training parameters in
conf/config.yamlthen start a new training job
uv run python src/deepweeds/train.py-
You can check the training artifacts such as exact input configuration, stdout log, etc. by finding the date/time folder under
outputs/corresponding to when you started a run -
Evaluate results through tensorboard if you'd like to compare against past runs
uv run tensorboard --logdir=outputsor if you want to look at one of the grid search runs
uv run tensorboard --logdir=multirunThe code is not guaranteed in any way to run as intended under this configuration due to various issues like MacOS 13 not supporting a few required tensor ops on the metal gpu, tensorflow not having a cuda enabled distribution for windows etc.
So try at your own risk...