Code Model release v0.0.1#
This changelog is a summary of the changes to the source code of the Clay model. Released on 2024/01/12.
For release notes for the trained model, see Pretrained Model release v0.0.1
đ« Highlights#
đ First release of Clay Foundation Model đ
Implement MAE with support for position, time, latlon & channel embeddings (#47)
đ Model#
Generate embeddings from CLAYModule trained with latlon/time encodings (#96)
Add multigpu support & UI to test embeddings (#109)
Refactor model for multi-device usage and easier disabling of masking (#95)
Rename embeddings file to include MGRS code and store GeoTIFF source_url (#86)
Callback function to log Masked Autoencoder reconstructions to WandB (#88)
Adapt model to load 512x512 images from s3 bucket (#85)
Save embeddings with spatiotemporal metadata to GeoParquet (#73)
Generate embeddings via prediction loop (#56)
Initial Vision Transformer architecture with MAE decoder (#37)
đïž Data Pipeline#
Adapted sampling strategy (#81)
Allow ClayDataModule to load GeoTIFF files directly from s3 (#92)
Let ClayDataModule return same spatiotemporal fields as GeoTIFFDataModule (#91)
Improve date handling for data pipeline (#76)
Let LightningDataModule return spatiotemporal metadata (#66)
check for no data on a tile level in sentinel 1 vv and vh, sentinel 2 and DEM (#60)
Batch setup (#54)
LightningDataModule to load GeoTIFF files (#52)
Ready for batch (#44)
Tiler module (#41)
Landcover based sampling strategy (#29)
Datacube (#27)
đ Documentation#
Document how the benchmark dataset labels were prepared (#100)
Document how to finetune pretrained model on downstream task (#99)
Document how to generate vector embeddings (#98)
Document how to run the datacube pipeline with a batch job (#97)
Initialize Jupyter Book documentation (#89)
Setting the model license to OpenRail-M (#63)
Create CODE_OF_CONDUCT.md (#53)
𧰠Maintenance#
đ§âđ€âđ§ Contributors#
Full Changelog: https://github.com/Clay-foundation/model/compare/v0.0.0âŠv0.0.1