LITHOS Official Benchmark Repository
LITHOS is the largest publicly available dataset and benchmark for automated petrography, containing over 211,000 high-resolution polarized light image patches and 105,000 expert-annotated mineral grains across 25 categories. This repository provides the dataset, code, and pretrained models, including a dual-encoder transformer baseline that leverages multiple polarization modalities for improved mineral classification. LITHOS aims to accelerate research and reproducibility in automated mineralogical analysis from thin section images.
To create the conda environment for LITHOS, run the following commands:
conda env create -f environment.yaml
conda activate LITHOSThe LITHOS Dataset is available via a private (for now) Kaggle URL. To download:
- Register for a Kaggle account if you don't have one.
- Access the private Kaggle dataset link.
- Follow the instructions on Kaggle to download the data (we recommend using curl for efficient downloading).
You can find our pretrained weights here!
Usage: main.py arguments
-h, --help show this help message and exit
--lr LR Initial learning rate (default: 1e-3)
--epochs EPOCHS Maximum number of epochs (default: 10)
--batch BATCH Batch size (default: 64)
--fold FOLD Fold to train (default: 1)
--test Only test the model
--ft Fine-tune a model
--resume Resume training a model
--model_type MODEL_TYPE
Model architecture (default: PolarViT)
--name NAME Name of the experiment (default: PolarViT_exp)
--load_model LOAD_MODEL
Weights to load (default: best_acc)
--n_classes N_CLASSES
Number of classes (default: 2)
--root_dataset ROOT_DATASET
Root path to the dataset (default: LITHOS_DATASET_P/LITHOS_DATASET)
--balance_dataloader Balance the dataloader
--balance_loss Balance loss
--use_alpha Use learnable α-gate fusion (default)
--use_linear Use nn.Linear on concatenated
features
--use_xpl Use XPL data
--optimizer {Adam,SGD}
Optimizer type (default: Adam)
--momentum MOMENTUM Momentum for SGD (default: 0.9)
--scheduler {StepLR,CosineAnnealingLR,None}
LR scheduler type (default: StepLR)
--step_size STEP_SIZE
Step size for StepLR scheduler (default: 10)
--gamma GAMMA Gamma for StepLR scheduler (default: 0.1)
--T_max T_MAX T_max for CosineAnnealingLR scheduler (default: 50)