Datasets: Checksums for oxford_iiit_pet changed, and the dataset cannot be loaded.

[greentec]

Short description
Checksums for oxford_iiit_pet dataset is changed.

Environment information

• Operating System: Google Colab
• Python version: 3.6.9 (default, Nov 7 2019, 10:44:02) \n[GCC 8.3.0]
• tensorflow-datasets/tfds-nightly version: 2.1.0
• tensorflow/tensorflow-gpu/tf-nightly/tf-nightly-gpu version: 2.2.0-rc3

Reproduction instructions

import tensorflow_datasets as tfds
# dataset, info = tfds.load('oxford_iiit_pet:3.0.0', with_info=True)
dataset, info = tfds.load('oxford_iiit_pet:3.*.*', with_info=True)

Link to logs
https://gist.github.com/greentec/f36b550c9b8c51b559ae3bb80588ae53

Expected behavior
Load the Oxford Pet Dataset normally.

[Eshan-Agarwal]

@greentec Thanks for reporting

[vijayphoenix]

@Conchylicultor @Eshan-Agarwal
I think we need to generate checksums of all the datasets with base URL robots.ox.ac.uk/~vgg/.

oxford_flowers102
dtd
vgg_face2
oxford_iiit_pet
visual_domain_decathlon
voxceleb

Edit: vgg_face2 and voxceleb are manual dataset

[Eshan-Agarwal]

@vijayphoenix Yes it seems they have some internal change

[mch55]

Hi folks, fantastic you are over this. I am on a course and it uses the oxford_flowers102 dataset so all of us are currently stuck and have fingers crossed it is an easy fix?
I am not the strongest programmer and have spent the last 48 hours trying to import the dataset other ways / generate the check_sum file for you but no luck so far.

If there is something we can do to help let us know?

[Eshan-Agarwal]

@mch55 Have a look on PR #1980, checksums for oxford_flowers102 are updated. Also try with pip install tfds-nightly it's recently merged. And if you are using oxford_iiit_pet dataset it's updated in PR #1979

[greentec]

@Eshan-Agarwal I tried pip install tfds-nightly and loaded oxford_iiit_pet dataset again, but failed again.

https://gist.github.com/greentec/c962f27fc67bf6899fdf87ca928bc034

[vijayphoenix]

After PR #1979 is merged, the changes will be reflected in tfds-nightly within few days

[greentec]

Okay. Thank you for checking.

[logiccodes]

In Colab this has wrong checksum. error seems to exist even if the tensorflow_datasets is updated to the nighty version.

!pip install git+https://github.com/tensorflow/datasets.git
import tensorflow_datasets as tfds
# dataset, info = tfds.load('oxford_iiit_pet:3.0.0', with_info=True)
dataset, info = tfds.load('oxford_iiit_pet:3.*.*', with_info=True)

[greentec]

@logiccodes I am also waiting for the tfds-nightly version update.
If you are in a hurry, you can download the dataset manually with the following command and then load it to execute normally.

!python -m tensorflow_datasets.scripts.download_and_prepare --register_checksums --datasets=oxford_iiit_pet:3.1.0

import tensorflow_datasets as tfds
dataset, info = tfds.load('oxford_iiit_pet:3.*.*', with_info=True)

[Conchylicultor]

The PR was rolled back as it was creating issue with TF colab. I'll try to re-merge it sometimes today.

[Eshan-Agarwal]

What exact problem is in colab ? as I successfully run it here is colab

[Conchylicultor]

@Eshan-Agarwal yesterday we removed the change, so it break tfds-nightly. I resubmitted it in #1995, so it should be fixed (again) in the next tfds-nightly. Sorry for the inconvinience.