SatlasPretrain: add new dataset #2248
Conversation
github-actions
bot
added
documentation
|
|
||
| Reference implementation: | ||
|
|
||
| * https://github.com/allenai/satlas/blob/main/satlas/model/dataset.py |
There was a problem hiding this comment.
Not sure if it's worth adding this in the docstring or just leaving it in the comments only. Happy to add pointers to the official codebase/data loaders for Satlas which may be preferred by some users.
torchgeo/datasets/satlas.py
Outdated
| 'metadata': (), | ||
| } | ||
|
|
||
| # NOTE: 'tci' is RGB (b04-02), not BGR (b02-04) |
There was a problem hiding this comment.
Could optionally add a bands parameter that allows users to specify the order of spectral bands returned by the model, but so far I don't think we need this feature.
torchgeo/datasets/satlas.py
Outdated
| channels.append(torch.tensor(np.array(img, dtype=np.float32))) | ||
| return torch.cat(channels) | ||
|
|
||
| def _load_label(self, label: str, col: int, row: int) -> Tensor: |
There was a problem hiding this comment.
TODO: add support for vector labels.
torchgeo/datasets/satlas.py
Outdated
| sample: dict[str, Tensor] = {} | ||
|
|
||
| for image in self.images: | ||
| sample[f'image_{image}'] = self._load_image(image, col, row) |
There was a problem hiding this comment.
The design decision here is to use image_{landsat,naip,sentinel} as the key so we can retain support for kornia.augmentation.AugmentationSequential auto-detecting the type of key/value pairs. Not sure how important this actually is since many augmentations like Normalize will be unique to each image, but it could be useful for augmentations like RandomCrop.
|
I'm curious if anyone has ever managed to successfully download https://github.com/allenai/satlas/blob/main/satlaspretrain_urls.txt because I have been trying for weeks and the download always dies in the middle. |
|
Have you tried the aws-cli? |
|
How do I convert these URLs to s3 equivalents? |
|
You can also do: From working with Maxar Open Data on S3 and similar on Azure -- using |
|
+1 to Caleb's suggestion. Best to run in a screen / tmux - mine took a long time, and unzipping s2 took a day or so (on slower disks)!
|
adamjstewart
force-pushed
the
datasets/satlas
branch
from
5713bda
to
8df93ec
Compare
This comment was marked as outdated.
This comment was marked as outdated.
adamjstewart
force-pushed
the
datasets/satlas
branch
from
0103711
to
796e544
Compare
This PR adds a data loader for the SatlasPretrain dataset.
This is a work in progress:
References:
@favyen2 @piperwolters can you review this PR as time permits? I'm still in the process of downloading the entire dataset, so it's going to be a bit before I can actually test it myself, but wanted to open a WIP PR anyway. I have a ton of questions for you that I'll leave in-line and we can resolve as we finalize the PR. The first draft will likely have significantly limited functionality compared to your reference implementation, but that's fine for our use case. We can always expand it in the future.
@ando-shah and I are planning on heavily using your dataset for our next paper. Our specific use case requires us to only sample from tiles where all low-resolution products (S1, S2, L) are available. My current plan is to create a custom
metadata/train_lowres_matching.jsonfile containing the paired-down list of tiles. We can host this in our repo, or you're also welcome to include this in yourmetadata.tar.gzfile once it's complete.(yes, the original image is that washed out, the TCI product is not great)