scripts/train.py

#!/usr/bin/env python3
"""DeepGenopix Training Script — Phase 1: Biological Validation.

Hyperparameter sweep script for TE family classification.
Supports baseline_v1, stride2_v1, stride8_v1, latent128_v1, latent768_v1,
layers4_v1, stem5_v1, stem7_v1 presets.

Usage:
    python scripts/train.py --preset baseline_v1 --data_dir /path/to/data
"""

import argparse
import json
import os
import sys
from pathlib import Path

import torch
import trackio
from torch.utils.data import DataLoader

# Add src to path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))

from deepgenopix.config import (
    BP_PER_TOKEN,
    DIM_FEEDFORWARD,
    D_MODEL,
    DROPOUT,
    NHEAD,
    NUM_LAYERS,
    STEM_KERNEL_SIZE,
    COMPRESSOR_STRIDE,
)
from deepgenopix.dataset import DeepGenopixDataset, dynamic_collate_fn
from deepgenopix.io_utils import load_json
from deepgenopix.model import DeepGenopixClassifier, count_parameters
from deepgenopix.trainer import train_model


# ── Preset Definitions ────────────────────────────────────────────────────────
PRESETS = {
    "baseline_v1": {
        "description": "Stride 4 (48bp/token), 256d latent, stem kernel 3, 2 layers",
        "compressor_stride": 4,
        "d_model": 256,
        "stem_kernel": 3,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
    "stride2_v1": {
        "description": "Stride 2 (24bp/token). High fidelity compression test.",
        "compressor_stride": 2,
        "d_model": 256,
        "stem_kernel": 3,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
    "stride8_v1": {
        "description": "Stride 8 (96bp/token). Extreme abstraction test.",
        "compressor_stride": 8,
        "d_model": 256,
        "stem_kernel": 3,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
    "latent128_v1": {
        "description": "128d latent vector. Compressed semantics test.",
        "compressor_stride": 4,
        "d_model": 128,
        "stem_kernel": 3,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 512,
    },
    "latent768_v1": {
        "description": "768d latent vector. LLM-scale embedding test.",
        "compressor_stride": 4,
        "d_model": 768,
        "stem_kernel": 3,
        "num_layers": 2,
        "nhead": 8,
        "dim_feedforward": 3072,
    },
    "layers4_v1": {
        "description": "4-layer Transformer Encoder. Deep grammar test.",
        "compressor_stride": 4,
        "d_model": 256,
        "stem_kernel": 3,
        "num_layers": 4,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
    "stem5_v1": {
        "description": "Conv1d stem kernel 5. Wider retina (60bp field).",
        "compressor_stride": 4,
        "d_model": 256,
        "stem_kernel": 5,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
    "stem7_v1": {
        "description": "Conv1d stem kernel 7. Widest retina (84bp field).",
        "compressor_stride": 4,
        "d_model": 256,
        "stem_kernel": 7,
        "num_layers": 2,
        "nhead": 4,
        "dim_feedforward": 1024,
    },
}


def parse_args():
    parser = argparse.ArgumentParser(description="DeepGenopix Training")
    parser.add_argument(
        "--preset",
        type=str,
        default="baseline_v1",
        choices=list(PRESETS.keys()),
        help="Hyperparameter preset",
    )
    parser.add_argument(
        "--data_dir",
        type=str,
        default="/app/data/processed",
        help="Directory with registry.csv, classes.json, te_visuals/",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="/app/data/output",
        help="Output directory for checkpoints",
    )
    parser.add_argument(
        "--epochs",
        type=int,
        default=50,
        help="Maximum training epochs",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=32,
        help="Batch size",
    )
    parser.add_argument(
        "--lr",
        type=float,
        default=1e-4,
        help="Learning rate",
    )
    parser.add_argument(
        "--weight_decay",
        type=float,
        default=1e-4,
        help="Weight decay",
    )
    parser.add_argument(
        "--max_samples",
        type=int,
        default=None,
        help="Cap samples for dry run (None = all)",
    )
    parser.add_argument(
        "--num_workers",
        type=int,
        default=2,
        help="DataLoader workers",
    )
    parser.add_argument(
        "--trackio_project",
        type=str,
        default="deepgenopix-v1",
        help="Trackio project name",
    )
    return parser.parse_args()


def main():
    args = parse_args()
    preset = PRESETS[args.preset]

    print("=" * 70)
    print(f"DeepGenopix Training — Preset: {args.preset}")
    print(f"  {preset['description']}")
    print(f"  Data: {args.data_dir}")
    print(f"  Output: {args.output_dir}")
    print("=" * 70)

    data_dir = Path(args.data_dir)
    output_dir = Path(args.output_dir) / args.preset
    os.makedirs(output_dir, exist_ok=True)

    # ── Load class mapping ─────────────────────────────────────────────────
    classes_path = data_dir / "classes.json"
    classes = load_json(classes_path)
    num_classes = len(classes)
    class_names = sorted(classes.keys(), key=lambda k: classes[k])
    print(f"\n[Main] {num_classes} classes loaded")

    # ── Create datasets ────────────────────────────────────────────────────
    registry_path = data_dir / "registry.csv"
    visuals_dir = data_dir / "te_visuals"

    train_dataset = DeepGenopixDataset(
        registry_path, visuals_dir, split="train", max_samples=args.max_samples,
    )
    val_dataset = DeepGenopixDataset(
        registry_path, visuals_dir, split="val", max_samples=args.max_samples,
    )

    print(f"[Main] Train: {len(train_dataset)} samples, Val: {len(val_dataset)} samples")

    # ── DataLoaders ────────────────────────────────────────────────────────
    train_loader = DataLoader(
        train_dataset,
        batch_size=args.batch_size,
        shuffle=True,
        collate_fn=dynamic_collate_fn,
        num_workers=args.num_workers,
        pin_memory=True,
    )
    val_loader = DataLoader(
        val_dataset,
        batch_size=args.batch_size,
        shuffle=False,
        collate_fn=dynamic_collate_fn,
        num_workers=args.num_workers,
        pin_memory=True,
    )

    # ── Build model ────────────────────────────────────────────────────────
    model = DeepGenopixClassifier(
        num_classes=num_classes,
        stem_kernel=preset["stem_kernel"],
        compressor_stride=preset["compressor_stride"],
        d_model=preset["d_model"],
        nhead=preset["nhead"],
        num_layers=preset["num_layers"],
        dim_feedforward=preset["dim_feedforward"],
        dropout=DROPOUT,
    )

    params = count_parameters(model)
    bp_per_token = 12 * preset["compressor_stride"]

    print(f"\n[Main] Model: {params['trainable']:,} params")
    print(f"[Main] BP/token: {bp_per_token}")

    # ── Trackio init ───────────────────────────────────────────────────────
    trackio.init(
        project=args.trackio_project,
        run_name=args.preset,
    )
    trackio.log_params({
        "preset": args.preset,
        "description": preset["description"],
        "compressor_stride": preset["compressor_stride"],
        "d_model": preset["d_model"],
        "stem_kernel": preset["stem_kernel"],
        "num_layers": preset["num_layers"],
        "nhead": preset["nhead"],
        "bp_per_token": bp_per_token,
        "num_classes": num_classes,
        "num_params": params["trainable"],
        "batch_size": args.batch_size,
        "lr": args.lr,
    })

    # ── Train ──────────────────────────────────────────────────────────────
    print(f"\n[Main] Starting training...")
    model, metrics = train_model(
        model=model,
        train_loader=train_loader,
        val_loader=val_loader,
        num_classes=num_classes,
        class_names=class_names,
        epochs=args.epochs,
        lr=args.lr,
        weight_decay=args.weight_decay,
        output_dir=output_dir,
        trackio_project=args.trackio_project,
    )

    # ── Final summary ──────────────────────────────────────────────────────
    print(f"\n{'='*70}")
    print(f"Training Complete: {args.preset}")
    print(f"  Best Val Acc: {metrics.best_val_acc:.4f}")
    print(f"  Best Val F1:  {metrics.best_val_f1:.4f}")
    print(f"  Best Epoch:   {metrics.best_epoch}")
    print(f"{'='*70}")

    trackio.alert(
        f"Training Complete: {args.preset}",
        f"Acc={metrics.best_val_acc:.4f}, F1={metrics.best_val_f1:.4f}, Epoch={metrics.best_epoch}",
        level="success",
    )

    # ── Save final artifacts ───────────────────────────────────────────────
    summary = {
        "preset": args.preset,
        **preset,
        "best_val_acc": float(metrics.best_val_acc),
        "best_val_f1": float(metrics.best_val_f1),
        "best_epoch": metrics.best_epoch,
        "num_params": params["trainable"],
        "num_classes": num_classes,
        "bp_per_token": bp_per_token,
    }

    with open(output_dir / "summary.json", "w") as fh:
        json.dump(summary, fh, indent=2)

    print(f"[Main] Summary saved: {output_dir / 'summary.json'}")


if __name__ == "__main__":
    main()