fryabcfry's picture
Publish AMP-PD checkpoints, portable loaders and verified metrics (part 3)
04ae9bb verified
Raw History Blame Contribute Delete
3.58 kB
"""Protein corrections selected from training-patient cross-validation."""
from __future__ import annotations
import json
from pathlib import Path
import numpy as np
from ampd.current_protein import errors
from ampd.omics_boost import fit_pls, transform_pls
from ampd.protein_representation import RepresentationPredictor, encode, fit_representation
from ampd.protein_residual import correct
FAMILIES = {"pca": [2, 4, 8, 16], "pls": [1, 2, 4, 8, 16], "missing": [2, 4, 8, 16]}
ALPHAS = (1., 10., 100., 1000.)
BLENDS = sorted([(a, b) for a in [0., .25, .5, .75, 1.] for b in [0., .25, .5, .75, 1.]
if a+b <= 1.], key=lambda p: (sum(p), p))
def fit_features(family, dim, month, proteins, residual):
if family == "pls":
return fit_pls(proteins, residual, dim)
params, _ = fit_representation(month, proteins, "missing" if family == "missing" else "pca")
return params
def features(family, dim, proteins, params):
if family == "pls":
return transform_pls(proteins, params)
raw = encode("missing" if family == "missing" else "pca", proteins, params)
return ((raw-params["feature_mean"])/params["feature_scale"])[:, :dim].astype(np.float64)
def blend_predictions(base, pca, pls, weights):
result = base.copy().astype(float)
for target, (a, b) in enumerate(weights):
cols = slice(target*4, (target+1)*4)
result[:, cols] = np.clip(base[:, cols] + a*(pca[:, cols]-base[:, cols]) + b*(pls[:, cols]-base[:, cols]), 0, 86)
return result
def choose_blend(y, base, pca, pls):
result, candidates = [], []
for target in range(4):
cols = slice(target*4, (target+1)*4)
scores = []
for a, b in BLENDS:
prediction = np.clip(base[:, cols]+a*(pca[:, cols]-base[:, cols])+b*(pls[:, cols]-base[:, cols]), 0, 86)
scores.append(float(np.nanmean(errors(y[:, cols], prediction))))
index = next(i for i, score in enumerate(scores) if score <= min(scores)+1e-10)
result.append(BLENDS[index])
candidates.append({"target": target+1, "candidates": [{"pca": a, "pls": b, "smape": s}
for (a, b), s in zip(BLENDS, scores)]})
return result, candidates
class CVProteinPredictor:
def __init__(self, bundle):
bundle = Path(bundle)
self.spec = json.loads((bundle / "model.json").read_text())
self.baseline = RepresentationPredictor(bundle / "baseline")
self.models = {}
for family in FAMILIES:
with np.load(bundle / f"{family}.npz", allow_pickle=False) as archive:
self.models[family] = dict(archive)
with np.load(bundle / "shared_pls.npz", allow_pickle=False) as archive:
self.shared = dict(archive)
def predict(self, events):
base = self.baseline.predict(events)
proteins = np.asarray([e["proteins"] for e in events], dtype=float)
result = {"none": base, "shared_pls": correct(base, transform_pls(proteins, self.shared) @ self.shared["coef"].T,
self.shared["weights"])}
for family, model in self.models.items():
dim = self.spec["selections"][family]["dim"]
x = features(family, dim, proteins, model)
result[f"cv_{family}"] = correct(base, x @ model["coef"].T, model["weights"])
result["equal_blend"] = .5*result["cv_pca"]+.5*result["cv_pls"]
result["cv_blend"] = blend_predictions(base, result["cv_pca"], result["cv_pls"], self.spec["blend_weights"])
return result