Issue with Neural Network-based particle picking

[doub1esheep]

Dear nextpyp Development Team,

I am encountering an issue when using the neural network-based picking in nextpyp. I am working with the EMPIAR-10987 dataset and selected ~50 particles in a tomogram for training, everything seemed to go well. However, when I used the trained model (model_last_contrastive.pth) to perform particle picking, no particles were picked.

I am wondering if there might be a particular reason for this outcome. Additionally, I would appreciate access to the trained model used in the tutorial video, so I can perform a comparison test.

[abartesaghi]

You probably want to spread your ~50 particles among 2-3 tomograms (ideally with different defocus). When training, make sure you set Bounding box = 12, Threshold for soft/hard positives = 0.5, and Max number of particles = 600.

Note: This assumes your tomograms were reconstructed using Binning factor for reconstruction = 12 as indicated in the practical.

The trained model is included in the tarball used for the workshop which can be downloaded using:

wget https://nextpyp.app/files/data/nextpyp_workshop_data.tbz

After decompressing the file, the trained model can be found in: 10987/model_last_contrastive.pth

[doub1esheep]

Thanks a lot for your suggestions! I spread the ~50 particles among 3 tomograms, this made the training model work well, both for EMPIAR-10987 dataset and my own datasets.
I’ve got a quick question: do the parameters in the "training/evaluation (pyp)" panel only apply during the training stage, or do they also affect the evaluation stage? Just trying to figure out if tweaking them during evaluation makes any difference.

[abartesaghi]

There are two parameters that affect evaluation: Max number of particles and Threshold for soft/hard positives. Increasing the former or decreasing the latter should give more particles.