dataload and dataset issues

[mustansarfiaz]

Hi,

I find your work very interesting. Therefore, I have downloaded the dataset you provided on Huggingface.

Query about the dataset:
It has csv files for train/val for iSoundNat and iSatNat. I cannot find the images and sound/sat images. Similarly for TaxaBench, I can find only test.csv.
All these datasets have no json files.

Query about the dataloader,
For example, at ./TaxaBind/SoundBind/dataloader.py#L45 and ./TaxaBind/SoundBind/dataloader.py#L46
I cannot find the images and sound_mp3 directory.
My question is do, I need to download images and sound separately?
and similarly from where I can get TaxaBind train.json and val.json?
here is the example
/TaxaBind/SatBind/config.py#L7
/TaxaBind/SatBind/config.py#L8

Regards
Mustansar

[Vishu26]

Hi

Thank you for raising this issue. The entire datasets are not published due to licensing. However, I will enlist the steps needed to properly setup the datasets below:

1. The csv files for iSatNat contain the urls of satellite images which are queried through WMS . You can use wget to download the entire dataset. The csv files also contain an id column. This id corresponds to the id present in the iNat-2021 dataset. You can download the jsons from here.
2. The csv files for iSoundNat again contain the urls of audio and ground-level images. These are not part of the Inat-2021 dataset. Hence, you will need to download the ground-level images as well.
3. Please use the same steps to setup the TaxaBench-8k dataset.

The train.json and val.json come from the iNat-2021 repository. Please download them here

Thanks