Export deep train model


I have used Deep Train to get a model and am confused with the export option. Basically I want to export this model to another project which shares similar target protein.

After I exported the model and imported it as a group to the new project, the Deep Inference job only picked 30 particles from over 3000 micrographs. I tested the model in old project and it was able to pick good amount of particles. I then tried to export the entire deep train job and import it as a job into the new project. This error message came up

[CPU: 819.1 MB]  Traceback (most recent call last):
  File "/cryosparc/cryosparc_master/cryosparc_compute/jobs/runcommon.py", line 1790, in run_with_except_hook
    run_old(*args, **kw)
  File "/cryosparc/cryosparc_master/cryosparc_command/command_core/__init__.py", line 162, in run
  File "/cryosparc/cryosparc_master/cryosparc_command/command_core/__init__.py", line 4552, in import_job_run
    rc.import_job(project_uid, new_uid, abs_path_export_job_dir, logger=lambda msg: job_send_streamlog(project_uid, new_uid, msg), notification_id=notification_id)
  File "/cryosparc/cryosparc_master/cryosparc_compute/jobs/runcommon.py", line 1298, in import_job
    assert False, "[IMPORT_JOB] : Unable to find data referenced by cs files. Aborting import of %s %s from %s"%(project_uid, job_uid, abs_path_export_job_dir)
AssertionError: [IMPORT_JOB] : Unable to find data referenced by cs files. Aborting import of P11 J41 from /PATH_OF_IMPORT_JOB

Even with this error message, the original Deep Train job showed up in my new project workspace.

My question is as following: what is the best approach to export a deep train model for use at a different project and/or workspace? Thanks!



Hi @jzhu, the individual result group export you tried first was the correct strategy. The full job import you tried would have yielded the same 30 particles result if it had worked because the exported model is exactly the same in both cases. The AssertionError you got was likely as due to broken links when copying the exported job into the new project; Iā€™m checking to see if there are any bugs in cryoSPARC that would cause this.

The reason for the small amount of picks is likely because model depends not only on the type of particle, but also attributes on the original captured data such as pixel size, CTF estimation, and others. If any of these differ between the two datasets you tried it on, the model will yield poor results. Your best option here is to train a new model with picks from the new dataset.

Thanks for the explanation!

1 Like