Non-deterministic relion_autopick on GPU

[bforsbe]

Originally reported by: Dimitry Tegunov (Bitbucket: DTegunov, GitHub: DTegunov)

---

Running

#!bash
`which relion_autopick` --i micrographs_mini.star --ref templates.star --odir average/ --pickname autopick --invert  --ctf  --ang 5 --shrink 1 --lowpass 20 --angpix 0.8 --angpix_ref 0.8 --particle_diameter 200  --threshold 0.3 --min_distance 100 --max_stddev_noise 1.1 --gpu

4 times yields 4 different results with 154, 159, 160, or 161 particles. micrographs_mini.star contains one 2954x3056 px micrograph; templates.star contains 7 280x280 px templates.

System:
Ubuntu 14.04,
Titan X cards,
352.79 drivers,
7.5 SDK,
Kernels compiled with sm_35

---

• Bitbucket: https://bitbucket.org/tcblab/relion2-beta/issue/5

[bforsbe]

Original comment by Bjoern Forsberg (Bitbucket: bforsbe, GitHub: bforsbe):

---

Thanks for providing such a detailed description. Is it possible to attach or specify the mrc and star files as well? I will try to reproduce with some data we have, but fairly recent tests of very similar runs have not shown these effects, so it might be observable only in conjuction with this specific data.

Did you evaluate via cmd-line, or did you start the run through the GUI?

[bforsbe]

Original comment by Sjors Scheres (Bitbucket: scheres, GitHub: scheres):

---

Perhaps adding --random_seed 0 would help? THere is some random noise generation involved at some point.

[bforsbe]

Original comment by Dimitry Tegunov (Bitbucket: DTegunov, GitHub: DTegunov):

---

Ah yes, fixing the seed did the trick. This is extra embarrassing as I had re-implemented the algorithm for tomograms at some point and should have remembered noise was involved. Sorry!

From a usability perspective, maybe the seed should be fixed by default? Users will typically run it multiple times to adjust the threshold, and seeing different results can be confusing in that scenario.