fryabcfry's picture
Publish AMP-PD checkpoints, portable loaders and verified metrics (part 3)
04ae9bb verified
Raw History Blame Contribute Delete
5.27 kB
"""Config loading, defaults, validation, and saving."""
from __future__ import annotations
from copy import deepcopy
from pathlib import Path
from typing import Any, Dict
from .utils import deep_update, load_yaml, save_yaml
DEFAULTS: Dict[str, Any] = {
"experiment": {
"name": "ampd_exp",
"seed": 42,
"device": "auto",
"output_dir": "runs",
"tags": [],
"notes": "",
},
"data": {
"root": "amp_data",
"base_visits": "protein",
"targets": ["updrs_1", "updrs_2", "updrs_3", "updrs_4"],
"horizons": [0, 6, 12, 24],
"include_supplemental": False,
"has_protein_only": False,
"impute": "median",
"standardize": True,
"target_transform": "none",
"process": {
"enabled": True,
"visit_source": "clinical",
"months": [0, 6, 12, 18, 24, 36, 48, 60, 72, 84],
"include_control": True,
"include_gap": True,
"include_prior_count": True,
"include_visit_month": True,
},
"proteins": {
"enabled": False,
"log1p": True,
"pca": 0,
},
"peptides": {
"enabled": False,
"mode": "protein_mean",
"top_k": 0,
"log1p": True,
},
},
"cv": {
"scheme": "patient_kfold",
"n_splits": 5,
"seed": 42,
"stratify_by": "cohort",
"cohort_scheme": "rule",
"holdout_cohorts": None,
"save_folds": True,
"single_val_fraction": 0.2,
},
"model": {
"name": "mlp",
"output_type": "regression",
"params": {
"hidden_dims": [64, 64],
"dropout": 0.1,
"activation": "relu",
"output_activation": "softplus",
"n_classes": 87,
},
},
"loss": {
"name": "masked_mae",
"params": {},
},
"optimizer": {
"name": "adamw",
"lr": 0.001,
"weight_decay": 0.0001,
},
"scheduler": {
"name": "cosine",
"interval": "epoch",
"params": {"eta_min": 1e-5},
},
"trainer": {
"max_epochs": 200,
"batch_size": 64,
"num_workers": 0,
"amp": True,
"grad_clip": 1.0,
"accumulate_grad_batches": 1,
"max_minutes": 0,
"stop_on_nan": True,
"log_every": 10,
"early_stopping": {
"enabled": True,
"monitor": "val_smape",
"mode": "min",
"patience": 20,
"min_delta": 0.0001,
},
"checkpoint": {
"monitor": "val_smape",
"mode": "min",
"save_top_k": 3,
"save_last": True,
"load_best_at_end": True,
},
"resume": "",
},
"evaluation": {
"metrics": ["smape_plus_1", "mae"],
"postprocess": {"updrs_4": "none"},
"per_target": True,
"per_horizon": True,
"per_cohort": True,
"bootstrap": 200,
"save_predictions": True,
},
"test": {
"enabled": False,
"test_csv": "",
"output": "submission.csv",
},
}
def default_config() -> Dict[str, Any]:
return deepcopy(DEFAULTS)
def load_config(path: str | Path, overrides=None) -> Dict[str, Any]:
user = load_yaml(path) or {}
cfg = deep_update(default_config(), user)
if overrides:
from .utils import apply_overrides
cfg = apply_overrides(cfg, overrides)
validate_config(cfg)
return cfg
def validate_config(cfg: Dict[str, Any]) -> None:
if cfg["data"]["base_visits"] not in {"all", "protein"}:
raise ValueError("data.base_visits must be 'all' or 'protein'")
if cfg["model"]["output_type"] not in {"regression", "distribution"}:
raise ValueError("model.output_type must be 'regression' or 'distribution'")
if cfg["cv"]["scheme"] not in {"patient_kfold", "cohort_heldout", "single"}:
raise ValueError("cv.scheme must be patient_kfold, cohort_heldout, or single")
if cfg["data"].get("process", {}).get("visit_source", "clinical") not in {"clinical", "sample"}:
raise ValueError("data.process.visit_source must be clinical or sample")
if cfg["model"]["output_type"] == "distribution" and cfg["loss"]["name"] != "masked_cross_entropy":
raise ValueError("distribution output requires loss.name=masked_cross_entropy")
if cfg["model"]["output_type"] == "regression" and cfg["loss"]["name"] == "masked_cross_entropy":
raise ValueError("regression output cannot use masked_cross_entropy")
if len(cfg["data"]["targets"]) * len(cfg["data"]["horizons"]) <= 0:
raise ValueError("targets x horizons must be non-empty")
if cfg["data"].get("target_transform") == "standardize" and cfg["model"]["output_type"] == "regression":
act = cfg["model"].get("params", {}).get("output_activation", "softplus")
if act == "softplus":
raise ValueError("target_transform=standardize cannot be combined with output_activation=softplus; use output_activation=none")
def save_config(cfg: Dict[str, Any], path: str | Path) -> None:
save_yaml(cfg, path)