yellow-rust-disease

This is the 2nd-place solution for the Beyond Visible Spectrum: AI for Agriculture 2023 challenge hosted in Kaggle.

Main ideas:

• Squeeze the 125 bands to only 3 bands using a "pre-Conv2D" layer
• Augmentations: RandomHorizontalFlip, RandomVerticalFlip, GaussianBlur
• EarlyStopping scheduler
• Freeze high level features' layers

How to reproduce the results

1. Go to this link, download the data, and put in the data folder.

2. Create a conda environment from an environment file:

conda env create -f environment.yml

3. Activate the conda environment:

conda deactivate  # if "base" is already activated
conda activate crop-disease

4. Train the model:

python3 -u src/train_nn.py --bs=56 > logs/train.log

This will create a folder like 20230912-153034_resnet18_bs56_acc0_760714.

5. Do predictions, passing as dir the folder the training script just created:

python3 -u src/predict_nn.py --dir=20230912-153034_resnet18_bs56_acc0_760714 > logs/pred.log

Results

Stratified KFold with k=5

• Mean accuracy: 0.760714
• Std. accuracy: 0.02208

Resources

• CentOS 7
• GPU: Tesla V100 with 32GB (only used around 2GB)
• Training: ~40 minutes
• Inference: ~3 minutes

Didn't work...oops

• Randomly choose a subset of bands (0 to 60 or 65 to 125, for example)
• Trying to learn which bands should the model focus on using "learnable weights"