Measure-Valued Forecast Martingale Regularizer / results.json
Mechanism failed
1{
2 "toy_check": {
3 "unbiased_moment_norm": 8.772099840083613e-05,
4 "biased_moment_norm": 0.03824778968300311,
5 "ratio_biased_to_unbiased": 436.01635161779626,
6 "passed": true
7 },
8 "runs": {
9 "0.0": {
10 "final_nll": 0.29930374026298523,
11 "final_accuracy": 0.89453125,
12 "all_step_nll": 0.5140613913536072,
13 "final_ece": 0.04871402928545539,
14 "step_ece": [
15 0.057396860123822405,
16 0.036185531412019425,
17 0.05958419666168996,
18 0.06957862667328078,
19 0.05579519336511572,
20 0.04871402928545539
21 ],
22 "heldout_moment_norm_mean": 0.019760587625205515,
23 "heldout_moment_norm_last_transition": 0.008101353421807289
24 },
25 "0.1": {
26 "final_nll": 0.29929980635643005,
27 "final_accuracy": 0.89453125,
28 "all_step_nll": 0.5140594840049744,
29 "final_ece": 0.048712579374168265,
30 "step_ece": [
31 0.05739706746652234,
32 0.03619036146749489,
33 0.059585514939611894,
34 0.06957796747819048,
35 0.05579668340897716,
36 0.048712579374168265
37 ],
38 "heldout_moment_norm_mean": 0.0197584617882967,
39 "heldout_moment_norm_last_transition": 0.00810413621366024
40 },
41 "1.0": {
42 "final_nll": 0.2992631196975708,
43 "final_accuracy": 0.89453125,
44 "all_step_nll": 0.5140413045883179,
45 "final_ece": 0.04870111735940631,
46 "step_ece": [
47 0.05815269974651401,
48 0.03703450499680155,
49 0.05959957834400827,
50 0.07002872681997613,
51 0.05580956063328311,
52 0.04870111735940631
53 ],
54 "heldout_moment_norm_mean": 0.019740549474954606,
55 "heldout_moment_norm_last_transition": 0.008131156675517559
56 }
57 },
58 "comparison": {
59 "baseline": {
60 "final_nll": 0.29930374026298523,
61 "final_accuracy": 0.89453125,
62 "all_step_nll": 0.5140613913536072,
63 "final_ece": 0.04871402928545539,
64 "step_ece": [
65 0.057396860123822405,
66 0.036185531412019425,
67 0.05958419666168996,
68 0.06957862667328078,
69 0.05579519336511572,
70 0.04871402928545539
71 ],
72 "heldout_moment_norm_mean": 0.019760587625205515,
73 "heldout_moment_norm_last_transition": 0.008101353421807289
74 },
75 "idea_lambda_0.1": {
76 "final_nll": 0.29929980635643005,
77 "final_accuracy": 0.89453125,
78 "all_step_nll": 0.5140594840049744,
79 "final_ece": 0.048712579374168265,
80 "step_ece": [
81 0.05739706746652234,
82 0.03619036146749489,
83 0.059585514939611894,
84 0.06957796747819048,
85 0.05579668340897716,
86 0.048712579374168265
87 ],
88 "heldout_moment_norm_mean": 0.0197584617882967,
89 "heldout_moment_norm_last_transition": 0.00810413621366024
90 },
91 "idea_lambda_1.0": {
92 "final_nll": 0.2992631196975708,
93 "final_accuracy": 0.89453125,
94 "all_step_nll": 0.5140413045883179,
95 "final_ece": 0.04870111735940631,
96 "step_ece": [
97 0.05815269974651401,
98 0.03703450499680155,
99 0.05959957834400827,
100 0.07002872681997613,
101 0.05580956063328311,
102 0.04870111735940631
103 ],
104 "heldout_moment_norm_mean": 0.019740549474954606,
105 "heldout_moment_norm_last_transition": 0.008131156675517559
106 }
107 }
108}