Replies: 1 comment
|
Thanks for reporting these. Both issues have been fixed in the release candidate 0.7.4. To upgrade:
|
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi,
I would like to raise awareness to two separate issues I noticed in the latest release (0.7.3) of nextPYP.
First, there seems to be an issue with the pixel size that arises if the Binning in the Tomography (from Raw Data) job is set as anything higher than 1. I binned my input data 4 times to speed up processing, and reconstructed my tomograms with an additional binning factor of 4, (total binning 16). However, looking at the scale in the tomograms and trying to measure the diameter of some prokaryotic ribosomes immediately indicates the pixel size must be 4 times off. The diameter of a prokaryotic ribosome is ~250 Å, but with the measurement tool, the value is ~60 Å. This error does not happen if I set the binning in my input data to 1 and reconstruct the tomogram with binning 16.
Second, pytom is having an issue in this new release of nextPYP, where it is failing to run simultaneously on more than on GPU. I have 3 GPUs on my workstation, and in version 0.7.2 pytom could lunch and run on all three GPUs simultaneously. Now, it seems to be running only on GPU 0. All jobs submitted to GPU 1 and GPU2 fail.
Here is the command that nextPYP executes:
export PATH=/opt/pixi/pytom/.pixi/envs/default/bin:$PATH; pytom_match_template.py -t pytom_template.mrc --mask template_mask.mrc -v L10_ts_002.rec -d pytom/ --particle-diameter 80.0 --voxel-size-angstrom 20.8 -a L10_ts_002.rawtlt --low-pass 20.0 --high-pass 500.0 --defocus 5.5633460938 --amplitude 0.07 --spherical 2.7 --voltage 300.0 -g 2 --log debug --half-precision
And here are is the error that pops up in the stream:
pytom_match_template.py: error: -g all gpu indices should be between 0 and {max_value-1}
Any insights on these issues would be appreciated!!
All reactions