Dataset available?

[calebrob6]

Hey all, is the Clay dataset available? We'd like to pre-train a location encoder on it if possible.

cc @konstantinklemmer

[brunosan]

You mean the input dataset? The script to create it is this one. Not sure it's valuable as a ready-made dataset, but if you want access to the bucket, happy to give you (and others) access.

If you mean the embeddings for the input data, they are here https://beta.source.coop/repositories/clay/clay-model-v0-embeddings/description/

[konstantinklemmer]

Actually that would be great - I suppose the dataset / dataloaders in here will work with the data stored in the bucket?
https://github.com/Clay-foundation/model/blob/main/src/datamodule.py

[lauracchen]

Hi @konstantinklemmer happy to share access to the S3 bucket of training data for Clay v0. Can you shoot me an email to coordinate access - [email protected]? Thanks!

[konstantinklemmer]

Hi @lauracchen! Thanks, I already got the access details from @calebrob6.

[konstantinklemmer]

Hi @brunosan! Quick question for my understanding. I run the following code to get a Clay batch:

loader = GeoTIFFDataPipeModule(data_dir="s3://clay-tiles-02", batch_size=32, num_workers=8)
loader.setup(stage="fit")

#Get first batch of data from train dataloader
for batch in loader.train_dataloader():
    print(batch)
    break

When I get the shape of the training images batch['image'].shape I get torch.Size([32, 13, 512, 512]).

Why do I get only 13 channels? Should this not be a stack of Sentinel-1, Sentinel-2 and the DEM?

Sorry if this is a stupid question, but I couldn't find the answer myself.

[lauracchen]

Hi @konstantinklemmer thanks for checking in! 13 channels is correct! This is a stack of 10 bands of Sentinel-2, 2 bands of Sentinel-1, and 1 band of DEM data. You can find more on the model release notes here.

[konstantinklemmer]

Got it, thank you!

[konstantinklemmer]

Another quick question @lauracchen: You have made the embeddings for all training data available very conveniently here: https://beta.source.coop/repositories/clay/clay-model-v0-embeddings/description/.

Are you also hosting the val and test set embeddings on the same platform or somewhere else? Thanks for all your help throughout!

[weiji14]

Those embeddings are generated from all of the GeoTIFF files we had (which includes both training + validation, there is no test set). Do you need separate train/validation embeddings for some reason? If so, we would need to capture the indexes from the code here

model/src/datamodule.py

Lines 304 to 309 in 3f4210f

	if stage == "fit": # training/validation loop
	# Step 2 - Split GeoTIFF chips into train/val sets (80%/20%)
	# https://pytorch.org/data/0.7/generated/torchdata.datapipes.iter.RandomSplitter.html
	dp_train, dp_val = self.dp_paths.random_split(
	weights={"train": 0.8, "validation": 0.2}, total_length=423, seed=42
	)

[konstantinklemmer]

Got it! No, that makes sense. I was just confused by the wording on the website the embeddings are stored on.