search

DIAGNijmegen / pathology-streaming-pipeline

Use streaming to train whole-slides images with single image-level labels, by reducing GPU memory requirements with 99%.
Apache License 2.0
75 stars 10 forks source link

Streaming for ResNet50 with FPN #5

Open rutgerfick opened 3 years ago

rutgerfick commented 3 years ago

Hello,

I've been reading your paper with interest and I was wondering how to apply the streaming technique to other networks.

In particular, I'm interested in implementing a streaming ResNet50 with Feature Pyramid Network on top of it.

Can you give me some hints on how to get started? Despite the documentation, the code is kind of a beast ;-)

Thank you, Rutger Fick

hanspinckaers commented 3 years ago

Hi Rutger,

A feature pyramid is gonna be quite difficult since we never have the full feature maps in memory. I think it would be possible, but would require some engineering 😊

It may be easier to start with the scnn.py code alone, outside of this whole pipeline. Something like the pseudo code here: https://github.com/DIAGNijmegen/StreamingCNN Also look at those notebooks, those should be relatively straightforward to follow.

Let me know if you have any questions!

Hans