How to replace blush with other denoising methods?

[lizrzr]

Hello, we are attempting to replace the existing Blush network framework with our own denoising algorithm. However, we did not find any Python source code for designing the network structure in the project. Could anyone please advise us on how to proceed?

[biochem-fan]

The Blush source code is in a separate repository. Please look at https://github.com/3dem/relion-blush/blob/main/model/model.py.

P.S. Please use a more descriptive title.

[yukijojo1]

Hello, we would like to replace the noise2noise framework in relion_blush with our denoising model. However, we are unclear about how the network framework of blush is integrated into the 3D auto-refine iterations in RELION. We noticed that Blush has its own source code and successfully reproduced it. In fact, our question is that we are not quite sure how Blush is integrated into RELION. Could you please provide us with some guidance?

[biochem-fan]

relion_refine_mpi writes unfiltered half maps and calls Blush.

See https://github.com/3dem/relion/blob/master/src/backprojector.cpp#L1206 (RELION side) and https://github.com/3dem/relion-blush/blob/main/relion_blush/command_line.py (Blush side).

[biochem-fan]

with our denoising model

By the way, please be really cautious at the risk of overfitting. This is discussed in our Blush paper and our earlier theoretical paper. A naive implementation easily leads to overfitting and false inflation of the resolution.