Use previous iteration checkpoint for 3DFlex reconstruct on (Jupyter)

Hello, I’m trying to use intermediate weights(checkpoints) of 3DFlex in

I could use intermediate weights in CryoSPARC webpage, but couldn’t find how to use those weights in Jupyter-notebook


Welcome to the forum @minkyujeon .
Are you aware that the job.load_output() has a version= parameter?

Hello @wtempel, thanks for your response!
I tried to use it, but what the input should be for loading intermediate checkpoints?


You mentioned that

Please can you show a screenshot of the upstream jobs Outputs section and the downstream jobs Inputs section as you connected them in the UI. Please also indicate the job precise types ob both jobs.

Hi @wtempel,
Here’s what I attempted in Jupyter Lab (as shown in the screenshot): I aimed to load latent space of intermediate checkpoints (Not final checkpoints) of 3DFlex and apply K-means clustering to generate new latents. Then, I created an external job to upload the new latents, which will be utilized in 3DFlex Generator.

@minkyujeon My colleague explained to me what is going on:
The version= parameter was planned to correspond to the iteration number you see in the UI
Except: There currently is a bug such that version= should instead correspond to the 0-based index of the desired iteration.
Thus, to load components from the 200th iteration, you should
job.load_output(name='particles', version=2)
There is a plan to correct this behavior in a future software release.
Does this help?