Unoptimized Weights with bitAccurateSynapses

[Michaeljurado42]

Describe the bug

I was converting lava-dl networks to lava via netx and noticed that the weight values were inflated. From what I understand, loihi2 should optimize weight matrices.

The issue can be solved by calling optimize_weight_bits after truncate_weights.

To reproduce current behavior

Try running this test: https://github.com/lava-nc/lava-dl/files/15328518/small_test.zip
and you will see that the weights are scaled from 64 to 128 with weight_exp = -1.

Expected behavior
weights should be optimized to 1 with weight_exp = 6.

Additional context

This issue can also technically be fixed in lava-dl by pre-optimizing the weights in lava-dl like this.

Although, if we properly fix the issue in lava we can remove most of the logic in lava-dl's fix_optimize_weights which is redundant with lava anyway

[PhilippPlank]

Thank you for reporting this bug/potential improvement. We will take a look when we find some time.