Topaz Train: Plot of ave. precision Vs epoch shows Straight line

Hi,
I have picked some particles using blob picker and filtered in 2D classificaiton. Now, I am running topaz train and it shows straight line on the ave. precisoin vs epoch plot.
Is it normal? How to interpret this?

Also, when I am running Topaz Extract I am getting an error “AssertionError: No particles extracted. Possible causes include ‘Particle threshold’ parameter being too high or a version discrepancy between the Topaz or PyTorch instance that trained the model and the one that is being used for extraction.”

@roney.tucker Please can you post

  • the plots you mentioned in your post.
  • outputs of this command for both the Topaz training and extraction jobs
    cryosparcm cli "get_job('P99', 'J199', 'job_type', 'params_spec', 'version', 'started_at', 'completed_at')"
    

where you replace P99 and J199 with the appropriate project and job IDs, respectively.

Hi,

I have a similar problem. I am trying to train Topaz on a small particle (~50*80 Å). The training set is 118 micrographs and 1800 particles. When running the Topaz train job , it finishes, but the test average is flat and the training precision is empty (see attached images). The output of your command is:

{'_id': '672a7119984f45e8540de5ab', 'completed_at': 'Tue, 05 Nov 2024 20:24:03 GMT', 'job_type': 'topaz_train', 'params_spec': {'exec_path': {'value': '/home/xx/opt/wrappers/topaz'}, 'num_epochs': {'value': 30}, 'num_particles': {'value': 500}, 'par_diam': {'value': 80}}, 'project_uid': 'P367', 'started_at': 'Tue, 05 Nov 2024 19:44:59 GMT', 'uid': 'J101', 'version': 'v4.5.3'}

I am trying to re-run the job with no changes and can post an update if that works.


Are you sure the parameter setting

is appropriate given

  1. the tooltip for this parameter: “Expected number of particles per micrograph.”

The grid seems a bit overcrowded if anything. The original picking was not well centered, resulting in the majority of 2D classes aligning to 2 or 3 particles, and I only included well centered picks in the training set, hence the very low particle number.
After giving it a second thought, this Topaz training job would probably not have yielded good results even if it finished properly, I will take a step back and optimize template picking.