Kolmogorov-Lie Unitary Layer / results.json
Beats tuned baseline
1{
2 "device": "cuda",
3 "math_single_residual": 5.9650624280038755e-06,
4 "math_32_product_residual": 1.3061589925200678e-05,
5 "KLU": {
6 "parameters": 564,
7 "train_loss_0_100_300": [
8 0.07692967355251312,
9 0.0011512121418491006,
10 0.00042202480835840106
11 ],
12 "test_mse": 0.0005223150947131217,
13 "norm_drift": 2.123415470123291e-07,
14 "unitarity_residual": 8.673438287587487e-07,
15 "seconds": 19.60685217997525,
16 "mean_grad_norm": 0.017432200691068662
17 },
18 "dense": {
19 "parameters": 4352,
20 "train_loss_0_100_300": [
21 0.07300132513046265,
22 0.0014969450421631336,
23 0.0002997505653183907
24 ],
25 "test_mse": 0.0003739910898730159,
26 "norm_drift": 0.009013999253511429,
27 "unitarity_residual": 0.543326735496521,
28 "seconds": 13.261652046989184,
29 "mean_grad_norm": 0.004672153878406284
30 }
31}