experiments/tasks/train_forecast.py

#!/usr/bin/env python3
"""Train + evaluate frame-level future verb_fine forecasting.

Outputs per-horizon top-1 frame accuracy on the test set, saved to
results.json under <output_dir>.
"""
from __future__ import annotations
import argparse
import json
import os
import random
import sys
import time
from pathlib import Path

import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader

THIS = Path(__file__).resolve()
sys.path.insert(0, str(THIS.parent))
sys.path.insert(0, str(THIS.parents[1]))
try:
    from experiments.dataset_forecast import (
        ForecastDataset, collate_forecast, build_train_test,
        IDLE_LABEL, NUM_FORECAST_CLASSES,
    )
    from experiments.models_forecast import build_forecast_model
except ModuleNotFoundError:
    from dataset_forecast import (
        ForecastDataset, collate_forecast, build_train_test,
        IDLE_LABEL, NUM_FORECAST_CLASSES,
    )
    from models_forecast import build_forecast_model


def set_seed(seed: int):
    random.seed(seed); np.random.seed(seed)
    torch.manual_seed(seed); torch.cuda.manual_seed_all(seed)


def train_epoch(model, loader, optimizer, criterion, device):
    model.train()
    total, n_frames, correct = 0.0, 0, 0
    for x, y, _ in loader:
        x = {m: v.to(device) for m, v in x.items()}
        y = y.to(device)            # (B, T_fut)
        optimizer.zero_grad()
        logits = model(x)            # (B, T_fut, C)
        loss = criterion(logits.reshape(-1, logits.size(-1)),
                         y.reshape(-1))
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        total += loss.item() * y.numel()
        n_frames += y.numel()
        correct += (logits.argmax(-1) == y).sum().item()
    return total / max(n_frames, 1), correct / max(n_frames, 1)


@torch.no_grad()
def evaluate(model, loader, device, t_fut: int):
    model.eval()
    # Per-horizon counts (overall, ignore-idle)
    per_h_correct = np.zeros(t_fut, dtype=np.int64)
    per_h_total   = np.zeros(t_fut, dtype=np.int64)
    per_h_correct_action = np.zeros(t_fut, dtype=np.int64)
    per_h_total_action   = np.zeros(t_fut, dtype=np.int64)

    for x, y, _ in loader:
        x = {m: v.to(device) for m, v in x.items()}
        y = y.to(device)            # (B, T_fut)
        logits = model(x)            # (B, T_fut, C)
        pred = logits.argmax(-1)     # (B, T_fut)
        for h in range(t_fut):
            yh = y[:, h]; ph = pred[:, h]
            per_h_correct[h] += (ph == yh).sum().item()
            per_h_total[h]   += yh.numel()
            mask = (yh != IDLE_LABEL)
            per_h_correct_action[h] += ((ph == yh) & mask).sum().item()
            per_h_total_action[h]   += mask.sum().item()

    return {
        "per_h_acc":         (per_h_correct / np.maximum(per_h_total, 1)).tolist(),
        "per_h_acc_action":  (per_h_correct_action / np.maximum(per_h_total_action, 1)).tolist(),
        "frame_acc":         float(per_h_correct.sum() / max(per_h_total.sum(), 1)),
        "frame_acc_action":  float(per_h_correct_action.sum() / max(per_h_total_action.sum(), 1)),
    }


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", type=str, required=True,
                    choices=["daf", "futr", "deepconvlstm", "rulstm", "avt"])
    ap.add_argument("--modalities", type=str, default="imu,emg,eyetrack,mocap,pressure",
                    help="Comma-separated modality list")
    ap.add_argument("--t_obs", type=float, default=1.5)
    ap.add_argument("--t_fut", type=float, default=0.5)
    ap.add_argument("--anchor_stride", type=float, default=0.25)
    ap.add_argument("--contact_only", action="store_true",
                    help="Only keep anchors whose past+future window has any "
                         "frame with pressure-sum > threshold (Plan B).")
    ap.add_argument("--contact_threshold_g", type=float, default=5.0)
    ap.add_argument("--epochs", type=int, default=15)
    ap.add_argument("--batch_size", type=int, default=64)
    ap.add_argument("--lr", type=float, default=3e-4)
    ap.add_argument("--weight_decay", type=float, default=1e-4)
    ap.add_argument("--d_model", type=int, default=128)
    ap.add_argument("--dropout", type=float, default=0.1)
    ap.add_argument("--label_smoothing", type=float, default=0.05)
    ap.add_argument("--num_workers", type=int, default=2)
    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--patience", type=int, default=5)
    ap.add_argument("--output_dir", type=str, required=True)
    args = ap.parse_args()

    set_seed(args.seed)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"device={device} | seed={args.seed} | model={args.model} "
          f"modalities={args.modalities}")

    mods = args.modalities.split(",")
    train_ds, test_ds = build_train_test(
        modalities=mods,
        t_obs_sec=args.t_obs, t_fut_sec=args.t_fut,
        anchor_stride_sec=args.anchor_stride,
        contact_only=args.contact_only,
        contact_threshold_g=args.contact_threshold_g,
    )
    print(f"train={len(train_ds)} test={len(test_ds)} "
          f"T_obs={train_ds.T_obs} T_fut={train_ds.T_fut} "
          f"mod_dims={train_ds.modality_dims}")

    tr_loader = DataLoader(
        train_ds, batch_size=args.batch_size, shuffle=True,
        num_workers=args.num_workers, collate_fn=collate_forecast,
        drop_last=False,
    )
    te_loader = DataLoader(
        test_ds, batch_size=args.batch_size, shuffle=False,
        num_workers=args.num_workers, collate_fn=collate_forecast,
    )

    model = build_forecast_model(
        args.model, train_ds.modality_dims,
        num_classes=NUM_FORECAST_CLASSES,
        t_obs=train_ds.T_obs, t_fut=train_ds.T_fut,
        d_model=args.d_model, dropout=args.dropout,
    ).to(device)
    n_params = sum(p.numel() for p in model.parameters())
    print(f"params={n_params:,}")

    optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr,
                                  weight_decay=args.weight_decay)
    sched = torch.optim.lr_scheduler.CosineAnnealingLR(
        optimizer, T_max=args.epochs, eta_min=args.lr * 0.05
    )
    criterion = nn.CrossEntropyLoss(label_smoothing=args.label_smoothing)

    out_dir = Path(args.output_dir); out_dir.mkdir(parents=True, exist_ok=True)
    best = {"frame_acc_action": -1.0, "epoch": 0, "state_dict": None}

    for ep in range(1, args.epochs + 1):
        t0 = time.time()
        tr_loss, tr_acc = train_epoch(model, tr_loader, optimizer, criterion, device)
        ev = evaluate(model, te_loader, device, t_fut=train_ds.T_fut)
        sched.step()
        print(f"  E{ep:2d} | tr {tr_loss:.4f}/{tr_acc:.3f} "
              f"| te frame_acc {ev['frame_acc']:.3f} action {ev['frame_acc_action']:.3f} "
              f"| {time.time()-t0:.1f}s")
        if ev["frame_acc_action"] > best["frame_acc_action"]:
            best = {**ev, "epoch": ep, "state_dict": {k: v.cpu() for k, v in model.state_dict().items()}}
            torch.save(best["state_dict"], out_dir / "model_best.pt")

    # Final reporting from best epoch
    final = {k: v for k, v in best.items() if k != "state_dict"}
    out = {
        "method": args.model,
        "modalities": mods,
        "seed": args.seed,
        "n_params": n_params,
        "T_obs": train_ds.T_obs,
        "T_fut": train_ds.T_fut,
        "best_epoch": int(best["epoch"]),
        "frame_acc": float(best["frame_acc"]),
        "frame_acc_action": float(best["frame_acc_action"]),
        "per_h_acc": list(map(float, best["per_h_acc"])),
        "per_h_acc_action": list(map(float, best["per_h_acc_action"])),
        "args": vars(args),
    }
    with open(out_dir / "results.json", "w") as f:
        json.dump(out, f, indent=2)
    print(f"\n[done] best frame_acc_action {best['frame_acc_action']:.4f} (epoch {best['epoch']})")
    print(f"per_h_acc_action: {[f'{a:.3f}' for a in best['per_h_acc_action']]}")
    print(f"saved to {out_dir}/results.json")


if __name__ == "__main__":
    main()