DynaMight tensor mismatch - The size of tensor a (2) must match the size of tensor b (10000) at non-singleton dimension 2

[Poko18]

Hi,
I've been testing DynaMight on a set of particles from empiar (10028).
I downsampled the particles from box size 360px to 128px with relion_image_handler, and prepared a refined model of size 128x128px.

DynaMight command:

dynamight optimize-deformations --refinement-star-file data/Refine3D/job010/run_data.star
--initial-model data/Refine3D/job010/run_class001.mrc 
--output-directory dynamight_run_128/ 
--gpu-id 1 --n-gaussians 10000 --initial-threshold 0.08 --regularization-factor 1 --n-threads 4

Error:

Training on GPU 1
Initializing the particle dataset
Assigning a diameter of 462 angstrom
Number of particles: 105247
Initialized data loaders for half sets of size 47361  and  47362
consensus updates are done every  0  epochs.
box size: 128 pixel_size: 3.76875 virtual pixel_size: 0.007751937984496124  dimension of latent space:  5
Number of used gaussians: 10000
Optimizing scale only
Initializing gaussian positions from reference
100%|##########| 50/50 [00:02<00:00, 17.78it/s]
Final error: 9.401748684467748e-05
Optimizing scale only
Initializing gaussian positions from reference
100%|##########| 50/50 [00:02<00:00, 17.24it/s]
Final error: 9.401141142006963e-05
consensus gaussian models initialized
consensus model  initialization finished
mean distance in graph for half 1: 4.976639747619629 Angstrom ;This distance is also used to construct the initial graph 
mean distance in graph for half 2: 4.976908206939697 Angstrom ;This distance is also used to construct the initial graph 
Computing half-set indices
100%|##########| 83/83 [00:03<00:00, 25.17it/s]
100%|##########| 83/83 [00:06<00:00, 12.16it/s]
new regularization parameter for half 1 is  0
new regularization parameter for half 2 is  0
100%|#########9| 370/371 [00:47<00:00,  7.82it/s]
The size of tensor a (2) must match the size of tensor b (10000) at non-singleton dimension 2
WARNING! batch size too large for gpu, trying with new batch size of: 64

I'm getting a tensor mismatch, and it probably has something to do with the n_gaussians.
I was wondering if anyone else has encountered a similar problem?

[schwabjohannes]

Hi Poko,

I think it is the same bug, that the remainder of particle images and batch size is 1 for one of the dataloaders. I will fix this. As a workaround you can change the batch-size to --batch-size 120

I hope this works,

Johannes

[Poko18]

That worked, thanks!