Open olibclarke opened 1 year ago
Looking at the run.log it looks like something went a little pear-shaped after ~it18 - even though the UMAP plots & volumes look fine, the KLD & loss values look not quite right. Possibly related? These particles are preprocessed with binning to Apix=5.84, is that too much downsampling?
EDIT: Not sure if this is the reason, as analyze_convergence.py still fails if I specify iteration 17...
(this was definitely caused by the nan
s and inf
s. A second round of train_vae
, after excluding a small (0.2%!) population of outlier particles, had no such stability issues, and analyze_convergence
worked fine)
Hi,
I ran
analyze_convergence.py
on one run oftrain_vae
and it worked as expected.When I ran it on a second run, it crashed with the appended output.
The only differences I can see between the run that worked and the one that failed are that in the one that worked, train_vae had run for 18 iterations with 128px preprocessed particles, while for the one that failed it had run for 31 iterations with 64px preprocessed particles.
cryodrgn analyze
run on the final iteration works fine, generates UMAP plots and cluster volumes etc.