lit-ecology-classifier 2.0.2

Image Classifier optimised for ecology use-cases

Lit Ecology Classifier

Documentation: https://lit-ecology-classifier.readthedocs.io/en/latest/ Lit Ecology Classifier is a machine learning project designed for image classification tasks. It leverages PyTorch Lightning for streamlined training and evaluation processes.

Features

• Easy configuration and setup
• Utilizes PyTorch Lightning for robust training and evaluation
• Supports training on multiple GPUs
• Test Time Augmentation (TTA) for enhanced evaluation
• Integration with Weights and Biases for experiment tracking

Installation

To install Lit Ecology Classifier, use pip:

pip install lit-ecology-classifier

Usage

Training

To train the model, use the following command:

python -m lit_ecology_classifier.main --max_epochs 20 --dataset phyto --priority config/priority.json

Inference

To run inference on unlabelled data, use the following command:

python -m lit_ecology_classifier.predict --datapath /path/to/data.tar --model_path /path/to/model.ckpt --outpath ./predictions/

Configuration

The project uses an argument parser for configuration. Here are some of the key arguments:

Training Arguments

• --datapath: Path to the tar file containing the training data.
• --train_outpath: Output path for training artifacts.
• --main_param_path: Main directory where the training parameters are saved.
• --dataset: Name of the dataset.
• --use_wandb: Use Weights and Biases for logging.
• --priority_classes: Path to the JSON file with priority classes.
• --balance_classes: Balance the classes for training.
• --batch_size: Batch size for training.
• --max_epochs: Number of epochs to train.
• --lr: Learning rate for training.
• --lr_factor: Learning rate factor for training of full body.
• --no_gpu: Use no GPU for training.

Inference Arguments

• --outpath: Directory where predictions are saved.
• --model_path: Path to the model file.
• --datapath: Path to the tar file containing the data to classify.
• --no_gpu: Use no GPU for inference.
• --no_TTA: Disable test-time augmentation.

Documentation

Detailed documentation for this project is available at Read the Docs.

Example SLURM Job Submission Script

Here is an example SLURM job submission script for training on multiple GPUs:

#!/bin/bash
#SBATCH --account="em09"
#SBATCH --constraint='gpu'
#SBATCH --nodes=2
#SBATCH --ntasks-per-core=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=12
#SBATCH --partition=normal
#SBATCH --constraint=gpu
#SBATCH --hint=nomultithread
#SBATCH --output=slurm/slurm_%j.out
#SBATCH --error=slurm/slurm_%j.err
export OMP_NUM_THREADS=12 #$SLURM_CPUS_PER_TASK
cd ${SCRATCH}/lit_ecology_classifier
module purge
module load daint-gpu cray-python
source lit_ecology/bin/activate
python -m lit_ecology_classifier.main --max_epochs 2 --dataset phyto --priority config/priority.json