Datasets problem

[towardshch]

Excuse me, when I tred to train your model I found the datasets.md was empty.
So how can I prepare my datasets or have the datasets.md finished?
Thanks!

[hukkelas]

Hi! I'll take a look at this today. Due to the messy license of the images in the COCO dataset, I will only release the annotations from CSE.
Along with the annotations, I'll publish a script to parse the coco dataset/CSE annotations, and generate the final dataset (with images & annotations).

[towardshch]

Thanks a lot!

[hukkelas]

The docs/DATASETS.md is now updated. Please notify me if you have any issues with downloading the COCO-cse dataset.

[towardshch]

Thanks for your update!
But I found some mistakes in the docs/DATASETS.md

• Somewhere the datasets use Coco2017 and somewhere is Coco2014.
• In the introduce of Coco_cse structure, there appears three times of train.

I followed your train code and find some mistakes which made it stop at the dataloader's collate_fn. Maybe it should add a resize function before the dataloader.

[hukkelas]

Thanks for notifying me! I fixed the typos now :)

Regarding the error in train.py, it's due to a bug in the coco-cse downloading script. This commit fixes the issue. You need to delete the dataset and regenerate it again.