Augumentation doesn't sync box and image

[wassname]

In my testing, augmentation doesn't flip box and image together. Here is a notebook showing/plotting it https://gist.github.com/wassname/970c914745d49bd5999d9896e2933db2.

The problem is in keras, and my keras PR from months ago fixes it (I've tested this). But it's waiting on a review so if you want this fixed head over and give it a review. All I did is add random.seed(seed) where needed.

[hgaiser]

Thanks for looking into this. I'll have to look at this later so that I can give it some more attention. All I can say at the moment is that the RNG part in Keras is bit of a mess, but I think you've noticed that too ;)

I'll hopefully have some time to check this somewhere this week.

[wassname]

Yeah the RNG part is doing my head in. No hurry, there's more important things than data augmentation.

[de-vri-es]

This could be fixed together with #68. I had some initial work on it in the cheap-augmentation branch but I have other priorities right now so I don't know when I'll continue on it. If someone wants to continue from that branch, feel free.

[wassname]

Nice I'll check out that branch, not sure if I'll continue it (depends on how it fits into my current project), but good to know

[wassname]

Not sure if you've seen torchvision, but their transforms.py is tidy and nice and uses PIL. It's similar to your approach of making modular transforms. Here's an example of a custom transform of masks and images together

[hgaiser]

Actually you had a small bug in the plotting code, the fixed version is here. The order of the boxes is (x1, y1, x2, y2), but you used it as (y1, x1, y2, x2).

[wassname]

Thanks for taking a look.

That might be an error I made, but it can't be the only cause. If that was the cause then my annotations would be consistently flipped in the x and y axis, but they aren't.

Still this shows that augmentation works for you, which means you can't reproduce it. I'll try it again with the latest code/keras to check it's still a problem. If you have time could you send me the output of pip freeze so I can compare our python environments please?

[awilliamson]

wassname, it might also be worth choosing different bbox coords than 0, 0, 2, 2 just so any artefacts in the flipping can pop out. Using boundary numbers isn't always a good idea for minimal working examples.

[wassname]

I tried again with keras 2.1.2 and the latest retinanet and it all worked fine. Sorry about this everyone.

I think it might have been a keras 2.0.9 vs 2.1.2 issue since this appears to have fixed some augmentation repeatability issues.

(also choose non symmetric boxes)