Allow use of a patch generator about attention-sampling HOT 5 OPEN

Hi Daniel,

The code is supporting it under the current API but it is not implemented. The API is in https://github.com/idiap/attention-sampling/blob/master/ats/data/base.py and you can see the FromTensors implementation of this API.

So basically if you want to implement openslide support (which would be awesome) the todo list would be like the following:

1. Implement a MultiResolutionBatch that uses an openslide dataset
2. Adapt the modules in ats.core.ats_layer such that they support multiple MultiResolutionBatch implementations
3. Use the rest of the code as is.

Let me know if any of these makes sense and if you have any more questions I 'd be glad to answer them.

Angelos

from attention-sampling.

Hi Angelos,

sorry I didn't answer earlier, i have a bit of time now it's the weekend.

If I understand the structure correctly, I would need to modify the function patches in from_tensors to have it call the function patches from MultiResolutionBatch, then depending on the type of dataset we would need MultiResolutionBatchto either call the function that you have implemented or openslide or whatever else.

Do you agree or do you see a better way to do it?

from attention-sampling.

Hmm I am not sure I understand what you mean.

If everything goes well you would not have to modify any existing file in the ats.data module. You would just implement the MultiResolutionBatch interface for an openslide backed dataset.

The FromTensors implementation uses predefined tensors to extract the patches. I would create an openslide implementation that would use paths to openslide images or possibly handles to openslide files.

Let me know if this makes more sense.

from attention-sampling.

Maybe I'm misunderstanding something, but my understanding of the current structure is as follows:

• the ats_layer calls FromTensors.patches to get the patches
• FromTensors.patches calls the extract_patches function from ..ops.extract_patches (and not patches from MultiResolutionBatch as it is not implemented)

Now in an implementation where we want to do the same thing, I would just have FromTensors to call MultiResolutionBatch.patches and I would have MultiResolutionBatch.patches call extract_patches. But this requires to have x_high passed in as a tensor. My goal would be to pass x_high only as an openslide object and use the read_region function to extract just the patches, without having to read the whole high resolution image. I don't really understand how you propose to implement this above. If you want, we can have a skype call to discuss about it, maybe by voice it would be easier.

from attention-sampling.

I would also be interested in integrating OpenSlide with this method (the MultiResolutionBatch seems like a nice place to start). @DanielRobertNicoud did you implement it already?

from attention-sampling.