nasa-cisto-data-science-group/satvision-pretrain-small

Satvision Pretraining Dataset - Small

• Developed by: NASA GSFC CISTO Data Science Group
• Model type: Pre-trained visual transformer model
• License: Apache license 2.0

This dataset repository houses the pretraining data for the Satvision pretrained transformers. This dataset was constructed using webdatasets to limit the number of inodes used in HPC systems with limited shared storage. Each file has 100000 tiles, with pairs of image input and annotation. The data has been further compressed to ease the download from HuggingFace.

SatelliteVision-Base (SatVis-B) is a pre-trained vision transformer based on the SwinV2 mode architecture. The model is pre-trained on global MODIS surface reflectance data from which 1.99 million image chips were used. SatVis-B is pre-trained using the masked-image-modeling (MIM) contrastive pre-training strategy. The MIM pre-training approach utilizes random masking of the input geospatial image chip, using a linear layer to regress the raw pixel values of the masked area with an l1 loss serving as the loss function.

Resolution of the pre-training MODIS chips was 128x128 with a window size of 16x16. SatViz-B was pre-trained for 800 epochs on 8x A100 GPUs and 12x V100 GPUs.

SatVision Transformer

Pre-trained models pre-trained on MODIS-Small dataset

name	pre-train epochs	pre-train resolution	#params	pre-trained model
SatVision-Base	800	128x128	84.5m	checkpoint/config

Getting Started with SatVision-Base

• Training repository: https://github.com/nasa-nccs-hpda/pytorch-caney
• Pre-training dataset repository: https://huggingface.co/datasets/nasa-cisto-data-science-group/satvision-pretrain-small

Installation

If you have singularity installed

$ git clone git@github.com:nasa-nccs-hpda/pytorch-caney.git
$ singularity build --sandbox pytorch-caney.sif docker://nasanccs/pytorch-caney:latest
# To shell into the container
$ singularity shell --nv -B <mounts> pytorch-caney.sif

Anaconda installation

$ git clone git@github.com:nasa-nccs-hpda/pytorch-caney.git
$ conda create -n satvision-env python==3.9

Fine-tuning Satvision-Base

• Create config file example config
• Download checkpoint from this HF model repo
• $ git clone git@github.com:nasa-nccs-hpda/pytorch-caney.git
• Add a new pytorch dataset in pytorch-caney/pytorch_caney/data/datasets/
• Add new pytorch dataset to dict in pytorch-caney/pytorch_caney/data/datamodules/finetune_datamodule.py

torchrun --nproc_per_node <NGPUS> pytorch-caney/pytorch_caney/pipelines/finetuning/finetune.py --cfg <config-file> --pretrained <path-to-pretrained> --dataset <dataset-name (key for new dataset)> --data-paths <path-to-data-dir> --batch-size <batch-size> --output <output-dir> --enable-amp

Pre-training with pytorch-caney

Pre-training with SatVision-Base with Masked Image Modeling and pytorch-caney

To pre-train the swinv2 base model with masked image modeling pre-training, run:

torchrun --nproc_per_node <NGPUS> pytorch-caney/pytorch_caney/pipelines/pretraining/mim.py --cfg <config-file> --dataset <dataset-name> --data-paths <path-to-data-subfolder-1> --batch-size <batch-size> --output <output-dir> --enable-amp

For example to run on a compute node with 4 GPUs and a batch size of 128 on the MODIS SatVision pre-training dataset with a base swinv2 model, run:

singularity shell --nv -B <mounts> /path/to/container/pytorch-caney-container
Singularity> export PYTHONPATH=$PWD:$PWD/pytorch-caney
Singularity> torchrun --nproc_per_node 4 pytorch-caney/pytorch_caney/pipelines/pretraining/mim.py --cfg pytorch-caney/examples/satvision/mim_pretrain_swinv2_satvision_base_192_window12_800ep.yaml --dataset MODIS --data-paths /explore/nobackup/projects/ilab/data/satvision/pretraining/training_* --batch-size 128 --output . --enable-amp

SatVision-Base Pre-Training Datasets

name	bands	resolution	#chips	meters-per-pixel
MODIS-Small	7	128x128	1,994,131	500m

Citing SatVision-Base

If this model helped your research, please cite satvision-base in your publications.

@misc{satvision-base,
    author       = {Carroll, Mark and Li, Jian and Spradlin, Caleb and Caraballo-Vega, Jordan},
    doi          = {10.57967/hf/1017},
    month        = aug,
    title        = {{satvision-base}},
    url          = {https://huggingface.co/nasa-cisto-data-science-group/satvision-base},
    repository-code = {https://github.com/nasa-nccs-hpda/pytorch-caney}
    year         = {2023}
}