File size: 8,425 Bytes
6fa9282 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 | """Train the approved residual map and retain validation-selected checkpoints."""
from __future__ import annotations
import copy, json, time, platform
from pathlib import Path
from dataclasses import dataclass, asdict
import numpy as np
import torch
from pivot.models.pivot import PIVOT
from pivot.models.encoders import build_pert_tensors
from pivot.training.losses import compute_losses
from pivot.evaluation.rewards import rbf_mmd2
from pivot.utils.common import set_seed
@dataclass
class TrainConfig:
d_pert: int = 64
hidden: int = 512
depth: int = 4
epochs: int = 60
batch_size: int = 1024
lr: float = 1e-3
weight_decay: float = 1e-5
lam_tan: float = 1.0
lam_semi: float = 0.5
lam_reg: float = 1e-4
lam_dist: float = 0.0
n_dist_perts: int = 4
dist_n: int = 64
grad_clip: float = 5.0
match: str = "batch"
rep_mode: str = "gene_op"
seed: int = 0
device: str = "cpu"
threads: int = 4
endpoint_weight: float = 0.0
def make_model(data, cfg):
if cfg.rep_mode not in ("gene_op", "gene_only", "op_only", "random_id"):
raise ValueError("Supported encoders use gene/operation metadata only")
return PIVOT(
data.d,
len(data.genes_vocab),
len(data.op_vocab),
len(data.perturbations),
d_pert=cfg.d_pert,
hidden=cfg.hidden,
depth=cfg.depth,
rep_mode=cfg.rep_mode,
).to(cfg.device)
@torch.no_grad()
def validation_loss(model, data, cfg) -> float:
"""Mean condition-level endpoint MSE against validation centroids in PCA space."""
rng = np.random.default_rng(cfg.seed + 910)
ctr = data.indices("val", True)
vid = data.indices("val", False)
c0 = torch.as_tensor(
data.emb[rng.choice(ctr, min(128, len(ctr)), replace=False)], device=cfg.device
)
model.eval()
loss = []
for label in data.labels("val"):
ids = np.intersect1d(data.pert_to_idx[label], vid)
g, o, m, pid = build_pert_tensors(data, [label], cfg.device)
pred = model.endpoint_from_pert(c0, g, o, m, pid).mean(0)
truth = torch.as_tensor(data.emb[ids].mean(0), device=cfg.device)
loss.append((pred - truth).square().mean().item())
return float(np.mean(loss))
def train(data, cfg: TrainConfig, output: str, resume: str | None = None) -> dict:
"""Fit using training cells/controls; write best.pt, last.pt, config and history.
Checkpoints retain the vocabulary and cache fingerprint. Resume restores
optimizer, scheduler, NumPy RNG, and Torch RNG state before the next epoch.
"""
set_seed(cfg.seed)
torch.set_num_threads(cfg.threads)
if cfg.device.startswith("cuda") and not torch.cuda.is_available():
raise RuntimeError("Requested CUDA is unavailable")
out = Path(output)
out.mkdir(parents=True, exist_ok=True)
model = make_model(data, cfg)
opt = torch.optim.AdamW(
model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay
)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=cfg.epochs)
rng = np.random.default_rng(cfg.seed)
start = 0
history = []
best = float("inf")
if resume:
ck = torch.load(resume, map_location=cfg.device, weights_only=False)
if ck["data_meta"] != data.meta or ck["config"] != asdict(cfg):
raise ValueError("Resume configuration or data mismatch")
model.load_state_dict(ck["model"])
opt.load_state_dict(ck["optimizer"])
sched.load_state_dict(ck["scheduler"])
rng.bit_generator.state = ck["numpy_rng"]
torch.set_rng_state(ck["torch_rng"].cpu())
if ck.get("cuda_rng") is not None and torch.cuda.is_available():
torch.cuda.set_rng_state_all(ck["cuda_rng"])
start = ck["epoch"] + 1
history = ck["history"]
best = ck["best_val"]
train_ids = data.indices("train", False)
ctrl = data.indices("train", True)
labels = data.obs.perturbation.to_numpy()
z = torch.as_tensor(data.emb, device=cfg.device)
groups = {p: train_ids[labels[train_ids] == p] for p in data.labels("train")}
lam = {"map": 1.0, "tan": cfg.lam_tan, "semi": cfg.lam_semi, "reg": cfg.lam_reg}
t0 = time.perf_counter()
for epoch in range(start, cfg.epochs):
model.train()
terms = []
for ids in np.array_split(
rng.permutation(train_ids), int(np.ceil(len(train_ids) / cfg.batch_size))
):
ci = data.sample_controls(ids, cfg.match, rng, ctrl)
g, o, m, pid = build_pert_tensors(data, labels[ids], cfg.device)
e = model.encode(g, o, m, pid)
total, parts = compute_losses(model.flow, e, z[ci], z[ids], lam)
if cfg.endpoint_weight:
le = (model.flow.endpoint(z[ci], e) - z[ids]).square().sum(-1).mean()
total = total + cfg.endpoint_weight * le
parts["endpoint"] = le.item()
if cfg.lam_dist:
ds = []
for p in rng.choice(
list(groups), min(cfg.n_dist_perts, len(groups)), replace=False
):
yi = rng.choice(
groups[p], min(cfg.dist_n, len(groups[p])), replace=False
)
xi = data.sample_controls(yi, cfg.match, rng, ctrl)
gd, od, md, pd = build_pert_tensors(data, [p], cfg.device)
yp = model.endpoint_from_pert(z[xi], gd, od, md, pd)
ds.append(rbf_mmd2(yp, z[yi], data.meta["mmd_gamma"]))
ld = torch.stack(ds).mean()
total = total + cfg.lam_dist * ld
parts["dist"] = ld.item()
opt.zero_grad()
total.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), cfg.grad_clip)
opt.step()
parts["total"] = total.item()
terms.append(parts)
sched.step()
val = validation_loss(model, data, cfg)
record = {
"epoch": epoch,
"validation_mse": val,
**{k: float(np.mean([t[k] for t in terms])) for k in terms[0]},
}
history.append(record)
improved = val < best
best = min(best, val)
ck = {
"protocol": "split-first-v1",
"model": model.state_dict(),
"optimizer": opt.state_dict(),
"scheduler": sched.state_dict(),
"epoch": epoch,
"best_val": best,
"config": asdict(cfg),
"data_meta": data.meta,
"gene_vocab": data.genes_vocab,
"perturbations": data.perturbations,
"numpy_rng": rng.bit_generator.state,
"torch_rng": torch.get_rng_state(),
"cuda_rng": (
torch.cuda.get_rng_state_all() if torch.cuda.is_available() else None
),
"history": history,
}
torch.save(ck, out / "last.pt")
if improved:
torch.save(ck, out / "best.pt")
print(
f"epoch {epoch+1}/{cfg.epochs} train={record['total']:.4f} val={val:.4f}",
flush=True,
)
info = {
"protocol": "split-first-v1",
"config": asdict(cfg),
"history": history,
"duration_s": time.perf_counter() - t0,
"n_train_cells": len(train_ids),
"n_parameters": sum(p.numel() for p in model.parameters()),
"software": {
"python": platform.python_version(),
"torch": torch.__version__,
"numpy": np.__version__,
},
}
(out / "training.json").write_text(json.dumps(info, indent=2))
return info
def load_checkpoint(path, data, device="cpu"):
"""Load trusted local checkpoint with an exact cache/vocabulary match."""
ck = torch.load(path, map_location=device, weights_only=False)
if ck.get("protocol") != "split-first-v1":
raise ValueError(
"Historical weights need their original preprocessing and archived loader"
)
if ck["data_meta"] != data.meta or ck["gene_vocab"] != data.genes_vocab:
raise ValueError("Checkpoint and cache do not match")
cfg = TrainConfig(**ck["config"])
cfg.device = device
torch.set_num_threads(cfg.threads)
model = make_model(data, cfg)
model.load_state_dict(ck["model"])
model.eval()
return model, cfg
|