Restricting the validation area to GB

[IFenton]

No description provided.

[IFenton]

I think this is now ready for review. In this PR:

• Vaildation metrics are now only calculated for a narrow cutout of the UK
• This works for both the full size images and the smaller cropped images
• A box is added to the videos showing the outlines of this cutout area

Examples can be seen here: https://wandb.ai/manchester_prize/cropping?nw=nwuserifenton. This has four runs:

• Nocrop. Large image without mask. Original data, main branch
• Nocrop-cutout. Large image with mask. Original data, this branch
• Crop. Small image without mask. Cropped data, main branch
• Crop-cutout. Small image with mask. Cropped data, this branch

They seem to behave as I expect, though given they are only run on five batches I wouldn't place too much weight on the exact results.

[dfulu]

Hi @IFenton, looks good. Before we merge can we check that we get the same validation scores for these runs?

• Nocrop-cutout. Large image with mask. Original data, this branch
• Crop-cutout. Small image with mask. Cropped data, this branch

If i understand right, they should have the same scores since they are evaluated on the same GB crop.

I think it should be possible to do that by setting the numpy random seed so that the shuffle is reproducible. And also setting number of workers to zero in the validate function so that multiprocessing doesn't cause different samples to be delivered. If that works I'd be happy :)

[IFenton]

Hi @IFenton, looks good. Before we merge can we check that we get the same validation scores for these runs?

* Nocrop-cutout. Large image with mask. Original data, this branch

* Crop-cutout. Small image with mask. Cropped data, this branch

If i understand right, they should have the same scores since they are evaluated on the same GB crop.

I think it should be possible to do that by setting the numpy random seed so that the shuffle is reproducible. And also setting number of workers to zero in the validate function so that multiprocessing doesn't cause different samples to be delivered. If that works I'd be happy :)

I ran that test, and the results of this are in the same WandB link. As you expected they do give identical results

[dfulu]

Great, looks good!