Not an easy LoRa to work with, but it can work.<code>v2</code> is a slight improvement, but I still think I can do better. I think the training data needs higher quality.

Nina Warink

PICS: 131BLIP: 10 - 75<code>network_dim</code>: 64<code>network_alpha</code>: 32<code>noise_offset</code>: 0.05steps: 10% 220/2200 [05:17&lt;47:33, 1.44s/it, loss=0.0935]steps: 20% 440/2200 [10:42&lt;42:50, 1.46s/it, loss=0.0889]steps: 30% 660/2200 [16:07&lt;37:37, 1.47s/it, loss=0.0894]steps: 40% 880/2200 [21:32&lt;32:18, 1.47s/it, loss=0.0898]steps: 50% 1100/2200 [26:56&lt;26:56, 1.47s/it, loss=0.0894]steps: 60% 1320/2200 [32:21&lt;21:34, 1.47s/it, loss=0.0905]

Files with SHA256: 46de7b146b3c6357f6f6606425324109c2831e00a806991356fa1b4fc3d4aa98