How to Download the VHR Dataset

[osgeokr]

I'm attempting to practice Object Detection using only the Positive image set from the VHR-10 dataset.
Following the guidelines on torchgeo.datasets, I've installed additional libraries pycocotools and rarfile. My torchgeo version is 0.5.2.

%pip install -q -U torchgeo
%pip install -q -U pycocotools
%pip install -q -U rarfile

However, when I try to download the VHR10 dataset with the code below, I encounter an error message stating NotRarFile: Not a RAR file. Could you let me know what I might be missing? Thank you in advance for your response.

from torchgeo.datasets import VHR10
ds = VHR10(root='data/VHR10/', split='positive', download=True)

[adamjstewart]

@ashnair1 can you confirm that the download still works properly? @osgeokr can you add checksum=True to verify if you're getting the expected checksum?

P.S. You can also install all optional dataset dependencies using pip install torchgeo[datasets]

[ashnair1]

It's working on my end.

@osgeokr Can you post the full stack trace?

[osgeokr]

@adamjstewart @ashnair1 During the attempt of pip install torchgeo[datasets] as you suggested, I encountered an error due to the zipfile-deflate64 package not supporting Python 3.11. Consequently, I downgraded to Python 3.10 and proceeded with the reinstallation (pip install torchgeo[datasets]).

Afterwards, when I executed the following code, I encountered a ModuleNotFoundError: No module named 'gdown' error:

from torchgeo.datasets import VHR10
ds = VHR10(root='data/VHR10/', download=True, checksum=True)

I successfully downloaded the NWPU VHR-10 dataset.rar after installing the gdown package. However, during the decompression process, I encountered the RarCannotExec: Cannot find working tool error, and the compression did not unzip properly. For your information, I am using Windows 10 operating system. This is suspected to be an error with the rarfile package.

After installing the aspose-zip package additionally, I successfully decompressed the file using the following code:

import aspose.zip as az

with az.rar.RarArchive("data/VHR10/NWPU VHR-10 dataset.rar") as archive:
    archive.extract_to_directory("data/VHR10/")

Up to this point, these are the observations made in the local environment. I'm curious if there's a learning example of downloading the VHR10 dataset in a Colab environment. Thank you. :)

[adamjstewart]

Surprised zipfile-deflate64 doesn't have newer wheels yet, I commented on brianhelba/zipfile-deflate64#41

The missing gdown is the result of pytorch/vision#8237, which replaced requests with gdown but didn't add an explicit dependency on gdown. I wonder if we can convince them to add this as a required dependency...

It sounds like the root of the issue is that you don't have unrar installed. This is related to markokr/rarfile#72. Basically, rarfile requires unrar, but doesn't install it automatically because it isn't a Python library and the license prevents them from bundling it. So you'll have to install unrar yourself using a package manager like choco or pacman or a binary download like https://gnuwin32.sourceforge.net/packages/unrar.htm.

[osgeokr]

I downloaded the UnRAR for Windows and extracted the unrarw64.exe file.
After that, the following code worked successfully!

from torchgeo.datasets import VHR10
ds = VHR10(root='data/VHR10/', split='positive', download=True)

Here are the extra steps I confirmed in the Windows environment:

• Installing torchgeo via pip install torchgeo[datasets]
• Installing the gdown package
• Downloading and extracting the unrarw64.exe file

I think it would be helpful to add related information to the VHR-10 Note.

Thank you for your advice! :)