Exact energy-preserving activation subsampling / results.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
 1{
 2  "seed": 155,
 3  "N": 16,
 4  "evaluations": 16,
 5  "device": "cuda",
 6  "math": {
 7    "matrix_frobenius_residual": 0.0,
 8    "max_relative_energy_error": 6.060764757252942e-16,
 9    "random_relative_rmse": 0.3231723090133205,
10    "random_mean_relative_std": 0.3222990582513704,
11    "float16_input_float32_accum_max_relative_error": 0.0007865975733926491
12  },
13  "gradient_variance": {
14    "exact_loss_mean": 1.0982712507247925,
15    "random_loss_mean": 1.1078505769371987,
16    "exact_mean_element_variance": 3.6805778756701444e-19,
17    "random_mean_element_variance": 4.0620361119181325e-07
18  },
19  "training": {
20    "exact": {
21      "final_loss": 0.001396093238145113,
22      "best_loss": 0.0013564005494117737,
23      "first_loss": 0.5757148265838623,
24      "time_sec": 0.9486046629899647
25    },
26    "random": {
27      "final_loss": 0.03257918870076537,
28      "best_loss": 0.005571001674979925,
29      "first_loss": 0.5806717872619629,
30      "time_sec": 0.9164062160125468
31    }
32  }
33}