For partinet denoise, do I need to use a weight model?

[Henry-946]

Dear Developer,

I am new to PartiNet, and have several questions to ask:

1, I noticed there are two weight files:
denoised_micrographs_v2.pt raw_micrographs.pt,

for micrographs denosie, I noticed the tutorial showed as:
partinet denoise
--source /data/my_project/motion_corrected
--project /data/my_project

so, do I need to use one of the model files above? also I noticed the help showed I can use the "--num_workers", what does this mean, is this mean CPU or GPU, e.g., if my node have 192 physical (384 logical) cores and 4 L40S GPU, what is the recommended number?

2, for the detect, the tutorial showed:
partinet detect
--weight /path/to/model_weights.pt
--source /data/partinet_picking/denoised
--device 0,1,2,3
--project /data/partinet_picking

for the --weight /path/to/model_weights.pt, which model file should I point to?

Many thanks!

Best,

Fengwei

[MihinP]

Hi Fengwei,

Thanks for testing out PartiNet.

1. partinet denoise does not need a weight file as it's a heuristic denoiser instead of an AI denoiser. The only arguments you need for denoising are --source = a directory containing the motion-corrected micrographs, and --project = the output directory where you want the denoised outputs (partinet will automatically create a folder in this path called "denoised").

For --num-workers you can leave this empty as most optimal, PartiNet will see the number of cores available and use half of them for denoising. Our testing suggested that for a single node, you need about half the cores for denoising, and the other half for I/O operations as there is a lot of file reading/writing in this step (as well as other overhead operations). For your node with 192 physical cores: if this argument is left empty, denoising will use 96 cores for denoising, and 96 for writing and reading micrographs (most optimal).

2. If you use PartiNet denoiser, you must use the denoised_micrographs_v2.pt, this weight file has been trained on micrographs that were denoised with PartiNet denoise.

Let me know if you have more questions/issues.

Cheers,
Mihin

[Henry-946]

Hi Mihin,

Thanks for your kind and prompt reply! I am little confused, so basicly you mean, I also can skip the 1st step - denoise, and in the 2nd step, I also need to designate the weight file, and point to raw_micrographs.pt, is my understanding correct?

Many thanks!

Fengwei

[MihinP]

Hi Fengwei,

Yes it is perfectly possible to skip the 1st step as you said. You designate the weight file as raw_micrographs.pt and run it on your motioncorrected micrographs. There were some initial bugs with this but I have (hopefully) cleaned up those problems with Issue #45.

If you do intent to do this please note we have a new version of raw_micrographs.pt available at https://huggingface.co/MihinP/PartiNet. You will have to dowload the weights again.

Lastly, in general we see degraded performance when not using the denoiser and running PartiNet with raw micrographs. Whilst you will still be able to pick particles properly, we found with our testing that denoising gave best performance.

Cheers,
Mihin

[Henry-946]

Hi Mihin,

Got it, thank you so much!

Have a wonderful day!

Best,
Fengwei