config dict | model_version string | parameters int64 | split dict | provenance_warning string | history list | analysis dict | ri_head dict | dvorak_head dict |
|---|---|---|---|---|---|---|---|---|
{
"epochs": 16,
"size": 96,
"seq": 6,
"batch": 24,
"lr": 0.002,
"weight_decay": 0.0001,
"max_storms": 90,
"holdout_seasons": 4,
"threads": 8,
"seed": 0,
"out": "/Users/arya/Documents/College/projects/trinetra/data/models/trinetra.pt"
} | trinetra-0.4.1 | 585,485 | {
"protocol": "chronological holdout",
"test_seasons": [
2021,
2022,
2023,
2024
],
"train_fixes": 4890,
"test_fixes": 797,
"group_key": "sid"
} | The image channels are generated by this project's own forward model. The analysis metrics below therefore measure how well the network inverts that forward model and validate the pipeline. They are not a claim about satellite intensity estimation. The forecast table, which runs on real best-track predictors, is. | [
{
"epoch": 1,
"seconds": 262.7,
"train_loss": 3.2294,
"test_vmax_rmse": 16.697,
"test_centre_median_km": 60.01,
"test_category_macro_f1": 0.0918,
"test_ri_auc": 0.6233,
"losses": {
"vmax": 0.1906,
"pmin": 0.2087,
"category": 0.5514,
"detection": 0.2796,
... | {
"vmax_rmse_ci": {
"point": 1.4703,
"ci_lo": 1.1621,
"ci_hi": 1.7115,
"level": 0.95
},
"vmax": {
"n": 797,
"rmse": 1.4703018038004185,
"mae": 0.871214654843511,
"bias": -0.17624831109902492,
"by_intensity": [
{
"bin": "17-33 kt (D, DD)",
"n": 409,
... | {
"n": 677,
"base_rate": 0.056129985228951254,
"mean_forecast": 0.04830054006273678,
"brier": 0.04270414029803309,
"brier_skill_score": 0.193948360157433,
"roc_auc": 0.8915657688822997,
"reliability": [
{
"bin_lo": 0,
"bin_hi": 0.1,
"n": 557,
"mean_forecast": 0.0168,
"obs... | {
"status": "unvalidated",
"reason": "Scene labels are weak, derived from the IMD CI number rather than from the CIMSS ADT archive, which was not pulled. No hand-labelled seed set exists, so the head is trained and served but not scored."
} |
No dataset card yet
- Downloads last month
- 64