Migrate from Radiant MLHub to Source Cooperative

[adamjstewart]

Summary

We currently have several datasets from the recently defunct Radiant MLHub that we need to switch to Source Cooperative if we want to be able to automatically download them.

• AgriFieldNet: AgriFieldNet: add download support #2180
• Benin Cashews: Benin Cashews: radiant mlhub -> source cooperative #2116
• Cloud Cover: Cloud Cover: radiant mlhub -> source cooperative #2117
• CV4A Kenya Crop Type: CV4A Kenya Crop Type: radiant mlhub -> source cooperative #2090
• Tropical Cyclone: Tropical Cyclone: radiant mlhub -> source cooperative #2068
• NASA Marine Debris: NASA Marine Debris: radiant mlhub -> source coop #2183
• Rwanda Field Boundary: Rwanda Field Boundary: radiant mlhub -> source cooperative #2118
• South Africa Crop Type: South Africa Crop Type: add download support #2181
• SpaceNet: SpaceNet: add SpaceNet 8, radiant mlhub -> aws #2203
• Western USA Live Fuel Moisture: WesternUSALiveFuelMoisture: radiant mlhub -> source coop #2206

Rationale

Downloads are currently broken, and many of these datasets have completely changed their file structure.

Implementation

See #2068 for an example that converts Tropical Cyclone. The resulting implementation is actually significantly simpler than the original code.

If you would like to volunteer to convert a particular dataset, please comment on this issue to say that you're working on this.

Alternatives

We can also rehost most datasets (depending on license) on Hugging Face.

[Haimantika]

Hi @adamjstewart I am interested to contribute to this, but I am fairly new and will need more guidance. What is a good place for me to start?

[adamjstewart]

Hi @Haimantika, thanks for volunteering!

Let's pick a single dataset, maybe torchgeo/datasets/benin_cashews.py, and try to convert it to the new syntax. Once that's working, we can repeat for the other 7 datasets and remove all mention of radiant-mlhub.

This is the new dataset website: https://beta.source.coop/technoserve/cashews-benin/

If you create an account, log in, and click generate credentials, you'll see that the Azure URI is https://radiantearth.blob.core.windows.net/mlhub/technoserve-cashew-benin

We'll add a new dependency on azure-storage-blob in pyproject.toml, requirements/datasets.txt, and requirements/min-reqs.old. I can help determine the minimum supported version.

We'll probably add something similar to download_radiant_mlhub_dataset but for Azure blobs in torchgeo/datasets/utils.py. This can then be imported in torchgeo/datasets/benin_cashews.py and used in _download.

Let me know if anything is unclear. The first dataset is going to be a bit of work, but once we have one working, the rest should be easy.

[Haimantika]

Hi @Haimantika, thanks for volunteering!

Let's pick a single dataset, maybe torchgeo/datasets/benin_cashews.py, and try to convert it to the new syntax. Once that's working, we can repeat for the other 7 datasets and remove all mention of radiant-mlhub.

This is the new dataset website: https://beta.source.coop/technoserve/cashews-benin/

If you create an account, log in, and click generate credentials, you'll see that the Azure URI is https://radiantearth.blob.core.windows.net/mlhub/technoserve-cashew-benin

We'll add a new dependency on azure-storage-blob in pyproject.toml, requirements/datasets.txt, and requirements/min-reqs.old. I can help determine the minimum supported version.

We'll probably add something similar to download_radiant_mlhub_dataset but for Azure blobs in torchgeo/datasets/utils.py. This can then be imported in torchgeo/datasets/benin_cashews.py and used in _download.

Let me know if anything is unclear. The first dataset is going to be a bit of work, but once we have one working, the rest should be easy.

This is very helpful. Thanks a lot. I will start working on it and get back with doubts, if any.

[Haimantika]

Hi @adamjstewart I finally got some time to work on it. I see a PR has been raised, is the issue solved already?

[adamjstewart]

I have not seen any PRs that implement download support for Source Cooperative. Which PR are you referring to?

[Haimantika]

I have not seen any PRs that implement download support for Source Cooperative. Which PR are you referring to?

My bad. This one just mentioned the issue.

[adamjstewart]

Yes, this is a 9th dataset that will benefit from your contribution.

P.S. I reached out to the folks at Source Cooperative. One thing to note is that azure-storage-blob will copy raw files/directories, not zip/tar files. So there won't be an easy way to checksum these. For now, let's just focus on downloading and ignore checksumming.

[Haimantika]

download_radiant_mlhub_dataset

Yes, this is a 9th dataset that will benefit from your contribution.

P.S. I reached out to the folks at Source Cooperative. One thing to note is that azure-storage-blob will copy raw files/directories, not zip/tar files. So there won't be an easy way to checksum these. For now, let's just focus on downloading and ignore checksumming.

Hi, I was doing a bit of research and the latest version of source cooperative that I could find was - beta.source.coop

Is that it? Or am I missing something? I have made the changes, can make a PR for you to take a look.

[adamjstewart]

Yes, that's the new website.

[Haimantika]

@adamjstewart I have raised a PR. There are chances that this is not the solution you are looking for. However I would like to give it one more try after your review and then unassign myself if it does not work to respect your time. :)

[darkblue-b]

review of MSFT azure-sdk-for-python that includes examples like this. Second view of the azcopy tool. python is preferred for torchGeo ; not clear how portable dependency management would work for azcopy .. Spack and conda have hooks but pip does not have good hooks for this kind of binary tool depends. Simply recommending azcopy and failing gracefully when it is not present was discussed briefly. not yet resolved

[adamjstewart]

We definitely don't need all of azure, azure-storage-blob would suffice.

[darkblue-b]

this file appears to implement basic functionality https://github.com/kartAI/kartAI/blob/master/azure/blobstorage.py

[adamjstewart]

@Haimantika @darkblue-b all preliminary work is now complete. If you want to claim 1 or more datasets from the above list and start working on them, #2068 will show you what is required to convert them. Note that most of the file changes in that PR are auto-generated by data.py. You really only need to change tests/data/foo/data.py to the new data structure and run it, and change torchgeo/datasets/foo.py to read and download the new data structure.

[Haimantika]

@Haimantika @darkblue-b all preliminary work is now complete. If you want to claim 1 or more datasets from the above list and start working on them, #2068 will show you what is required to convert them. Note that most of the file changes in that PR are auto-generated by data.py. You really only need to change tests/data/foo/data.py to the new data structure and run it, and change torchgeo/datasets/foo.py to read and download the new data structure.

Thanks Adam. I will take a look at the code and the dataset tonight and update you on which one I take up.

[adamjstewart]

Pinging the original dataset contributors:

• AgriFieldNet
• Benin Cashews
• Cloud Cover
• CV4A Kenya Crop Type
• Tropical Cyclone
• NASA Marine Debris
• Rwanda Field Boundary
• South Africa Crop Type
• SpaceNet: @ashnair1
• Western USA Live Fuel Moisture: @nilsleh

If any of you have time, would you be interested in revamping these datasets to download from Source Cooperative?

[Haimantika]

Hey @adamjstewart I will be taking up the NASA Marine Debris dataset. Will start working from this weekend.

[adamjstewart]

@ashnair1 it looks like SpaceNet is no longer hosted by Source Cooperative and is only on AWS, is that correct? We can use the same CLI function I wrote to update that dataset to its new download location. There's also a new SpaceNet8 released if you want to add it.