- Update link and python command by @APatrickJ in #490
- Propagate tokenization config to max_t computation by @gabrieltseng in #492
- Refactor 2D position encoding to support rectangular grids by @robmarkcole in #484
- Add script to create subset of pre-training dataset by @favyen2 in #489
- Use optional dependencies instead of dependency groups. by @favyen2 in #498
- Add EuroCrops data by @favyen2 in #485
- Fixing eurosat drop with single bandset by @yawenzzzz in #494
- Map ablations by @gabrieltseng in #474
- Default norms 2std by @Hgherzog in #500
- Beaker config retries default by @Hgherzog in #503
- refactoring to have cleaner data containers by @Hgherzog in #486
- NDVI + time explicit masking by @gabrieltseng in #501
- Config fixes for loading checkpoints with TokenizationConfig by @gabrieltseng in #508
- Add configurable primary eval metric with namespace separation by @Hgherzog in #504
- Henryh/training speedups by @Hgherzog in #495
- Henryh/small bug fix by @Hgherzog in #509
- Henryh/pipe studio tasks into evals by @Hgherzog in #506
- Add an eval wrapper for the band dropout experiments by @gabrieltseng in #512
- Pjreddie/embedding output size by @pjreddie in #478
- eval fixes by @gabrieltseng in #517
- Update wandb script to reflect new changes to evals by @gabrieltseng in #513
- minor readability update to flexi_vit by @Farbum in #523
- Add embedding diagnostics eval mode by @Hgherzog in #529
- Fix one more edge case for finetuning with preemption by @yawenzzzz in #532
- fix geo evals primary metric by @gabrieltseng in #533
- Add vectorized masked-negatives patch discrimination loss by @Hgherzog in #542
- Fix WorldCereal data by using SINGLE_COMPOSITE and setting correct NODATA by @favyen2 in #507
- New eval task: 8 day burn-risk forecast by @Farbum in #544
- Fix copy_wandb_run script by @yawenzzzz in #535
- Get best LR from finetuning runs on wandb by @gabrieltseng in #545
- Fix band dropout being applied during fine-tuning bug that Gabi identified by @favyen2 in #548
- Single bandset ablations by @gabrieltseng in #519
- Update Beaker budget (ai2/es-platform -> ai2/atec-olmoearth) by @favyen2 in #550
- Support adding additional hidden layers to the initial projection by @favyen2 in #546
- Use nodata_value instead of nodata_vals. by @favyen2 in #552
- Fix worlcover config by @piperwolters in #553
- rslearn bump by @Farbum in #557
- Update st_model.py by @favyen2 in #547
- Nano hidden run by @gabrieltseng in #551
- Add pretrain auxiliary probes and unify low-label eval config by @Hgherzog in #561
- Update Tessera band order documentation and correct band sequence for compatibility with released checkpoints by @FrankFeng-23 in
#554 - Add s1, s1+s2 for srtm by @gabrieltseng in #565
- Add v1.1 scripts by @gabrieltseng in #566
- Mitigate memory leak in pre-training runs by @favyen2 in #563
- rslearn task ingestion: verbose and tag filtering by @Farbum in #538
- add skip task names arg to finetune sweep script by @piperwolters in #555
- Add dice loss option to finetune args by @piperwolters in #560
- UNet head for finetuning by @piperwolters in #559
- Snapshot the evals instead of reading from pretraining data by @gabrieltseng in #575
- Add lfmc and single probe for regression tasks by @Farbum in #569
- Add 50 cities dataset by @gabrieltseng in #570
- Mapbiomas eval (Brazil) by @Farbum in #578
- Add temporal RoPE model support by @Hgherzog in #572
- Update readme with v1.1 model weights by @gabrieltseng in #580
- GEO-Bench-2, UNet finetuning head, Finetuning dice loss by @piperwolters in #556
- Fix scripts to reflect our actual runs by @gabrieltseng in #581
New Contributors
- @APatrickJ made their first contribution in #490
- @robmarkcole made their first contribution in #484
- @Farbum made their first contribution in #523
- @piperwolters made their first contribution in #553
- @FrankFeng-23 made their first contribution in #554
Full Changelog: v0.1.0...v0.1.1