Topaz precision fall on test dataset

Hi all, I tried to train a topaz model to pick on a dataset, but the precision is very low and fall. The Extract job picked wrong particles.
I tried 9000, 20900, or 137377 particle to train a model, but all of them failed.
I used templet picker to generate the training set, and 2D classes seem good.


my parameters:

It’s very strange. Can anyone help me?

my parameters:

Hi @cloudme-mi,

Its likely that downsampling 16x is too much. TOPAZ works best when the downsampled pixel size is between 4-6Å. I typically downsample to 4Å and never have an issue (eg. for a dataset with a pixel size of 0.85Å, I would downsample 5x). Also, if you supply a particle diameter in the auto mode, but specify a downsample factor in the manual mode, the manual mode will override the auto mode.

Additionally, what was the full workflow to generate training data for TOPAZ?

Best,
Kye