Distillation using a detector specific dataset

[carbonscott]

Distillation csv (soft label): /sdf/data/lcls/ds/prj/prjcwang31/results/proj-peaknet/distill

[carbonscott]

Lessons:

• It does really well on epix.
• BiFPN can be shallower (2 BiFPN blocks).

[carbonscott]

Rayonix is the next one to try.

[carbonscott]

One way to check if this is a good result is to run inference and perform indexing.

[carbonscott]

Hmm... the rayonix distillation is not done yet, but it seems there's a benefit of training on a mix of data. It seems during the distillation, the model learns to find peaks that are not in the non-crystalline diffraction zone.

.., but keep in mind that the distillation is not done yet.

[carbonscott]

After another day of training, the model (atto-2bifpn512) still struggles to find peaks inside the water ring area (non-crystalline scattering). I suspect small peaks inside the ring don't provide enough signals for the model to learn. This may suggest that mixed detector training could be more advantageous.

Student, atto-2bifpn512

Teacher, huge-4bifpn512