scripts/train.py

"""Train or fine-tune the SEEDS conditional diffusion model."""

from __future__ import annotations

import argparse
import math
import os

import numpy as np
import torch
import torch.distributed as dist
from torch import nn
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DistributedSampler

from common import SEEDS, build_model, choose_device, load_config, resolve_path, set_seed
from data_loader import SEEDSDataset, build_dataloader


def _is_distributed() -> bool:
    return int(os.environ.get("WORLD_SIZE", "1")) > 1


def _setup_distributed(requested_device: str) -> tuple[torch.device, int, int, bool]:
    distributed = _is_distributed()
    if not distributed:
        return choose_device(requested_device), 0, 1, False
    if not dist.is_initialized():
        backend = "nccl" if torch.cuda.is_available() else "gloo"
        dist.init_process_group(backend=backend, init_method="env://")
    rank = dist.get_rank()
    world_size = dist.get_world_size()
    local_rank = int(os.environ.get("LOCAL_RANK", rank))
    if torch.cuda.is_available():
        torch.cuda.set_device(local_rank)
        device = torch.device("cuda", local_rank)
    else:
        device = torch.device("cpu")
    return device, rank, world_size, True


def _denoising_loss(
    model: nn.Module,
    clean: torch.Tensor,
    seeds: torch.Tensor,
    climate: torch.Tensor,
) -> torch.Tensor:
    core_model = model.module if isinstance(model, DDP) else model
    diffusion_time = torch.rand(clean.shape[0], device=clean.device, dtype=clean.dtype)
    noise = torch.randn_like(clean)
    sigma = core_model.sigma(diffusion_time).view(-1, 1, 1, 1, 1)
    noisy = clean + sigma * noise
    model_input = noisy / torch.sqrt(1.0 + sigma.square())
    prediction = model(model_input, seeds, climate, diffusion_time)
    return ((prediction - noise) ** 2).flatten(1).mean()


def _run_validation(model: nn.Module, loader, device: torch.device, distributed: bool) -> float:
    model.eval()
    losses = []
    cuda_devices = [device.index] if device.type == "cuda" and device.index is not None else []
    with torch.random.fork_rng(devices=cuda_devices):
        torch.manual_seed(12345)
        with torch.no_grad():
            for batch in loader:
                loss = _denoising_loss(
                    model,
                    batch["targets"].to(device),
                    batch["seeds"].to(device),
                    batch["climate"].to(device),
                )
                losses.append(float(loss))
    if not losses:
        raise RuntimeError("validation loader produced no batches")
    value = torch.tensor(float(np.mean(losses)), device=device)
    if distributed:
        dist.all_reduce(value, op=dist.ReduceOp.SUM)
        value /= dist.get_world_size()
    return float(value.item())


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--config", default="conf/config.yaml")
    parser.add_argument("--max-steps", type=int, default=None)
    parser.add_argument("--max-batches", type=int, default=None)
    parser.add_argument("--device", default=None)
    parser.add_argument("--finetune", action="store_true")
    args = parser.parse_args()
    config = load_config(args.config)
    device, rank, world_size, distributed = _setup_distributed(args.device or config["training"]["device"])
    set_seed(config["project"]["seed"] + rank)
    data, paths, training = config["data"], config["paths"], config["training"]
    root = args.config
    train_path, val_path = resolve_path(paths["train_data"], root), resolve_path(paths["val_data"], root)
    train_dataset = SEEDSDataset(train_path, len(data["variables"]), data["faces"], data["height"], data["width"], data["seed_count"])
    val_dataset = SEEDSDataset(val_path, len(data["variables"]), data["faces"], data["height"], data["width"], data["seed_count"])
    train_sampler = DistributedSampler(train_dataset, num_replicas=world_size, rank=rank, shuffle=True) if distributed else None
    val_sampler = DistributedSampler(val_dataset, num_replicas=world_size, rank=rank, shuffle=False) if distributed else None
    train_loader = build_dataloader(train_path, len(data["variables"]), data["faces"], data["height"], data["width"], data["seed_count"], training["batch_size"], True, training["num_workers"], args.max_batches, train_sampler, train_dataset)
    val_loader = build_dataloader(val_path, len(data["variables"]), data["faces"], data["height"], data["width"], data["seed_count"], training["batch_size"], False, training["num_workers"], args.max_batches, val_sampler, val_dataset)
    model = build_model(config).to(device)
    checkpoint_path = resolve_path(paths["checkpoint"], root)
    if args.finetune or training.get("resume"):
        source = resolve_path(training.get("resume") or paths["checkpoint"], root)
        if not source.exists():
            raise FileNotFoundError(f"checkpoint does not exist: {source}")
        state = torch.load(source, map_location=device, weights_only=False)
        model.load_state_dict(state["model"] if "model" in state else state)
    if distributed:
        model = DDP(model, device_ids=[device.index] if device.type == "cuda" else None)
    optimizer = torch.optim.AdamW(model.parameters(), lr=training["learning_rate"], weight_decay=training["weight_decay"])
    train_losses, val_losses = [], []
    completed_steps = 0
    accumulation_steps = training.get("gradient_accumulation_steps", 1)
    if accumulation_steps < 1:
        raise ValueError("gradient_accumulation_steps must be positive")
    loss_ema = None
    for epoch in range(training["epochs"]):
        if train_sampler is not None:
            train_sampler.set_epoch(epoch)
        model.train()
        epoch_losses = []
        optimizer.zero_grad(set_to_none=True)
        accumulated_batches = 0
        for batch in train_loader:
            loss = _denoising_loss(
                model,
                batch["targets"].to(device),
                batch["seeds"].to(device),
                batch["climate"].to(device),
            )
            if not torch.isfinite(loss):
                raise FloatingPointError(f"non-finite training loss at epoch {epoch + 1}: {loss.item()}")
            (loss / accumulation_steps).backward()
            epoch_losses.append(float(loss.detach()))
            completed_steps += 1
            accumulated_batches += 1
            if accumulated_batches == accumulation_steps:
                torch.nn.utils.clip_grad_norm_(model.parameters(), training["grad_clip_norm"])
                optimizer.step()
                optimizer.zero_grad(set_to_none=True)
                accumulated_batches = 0
            if args.max_steps is not None and completed_steps >= args.max_steps:
                break
        if accumulated_batches:
            torch.nn.utils.clip_grad_norm_(model.parameters(), training["grad_clip_norm"])
            optimizer.step()
            optimizer.zero_grad(set_to_none=True)
        if not epoch_losses:
            raise RuntimeError("training loader produced no batches")
        epoch_loss_tensor = torch.tensor([sum(epoch_losses), len(epoch_losses)], device=device, dtype=torch.float64)
        if distributed:
            dist.all_reduce(epoch_loss_tensor, op=dist.ReduceOp.SUM)
        epoch_loss = float((epoch_loss_tensor[0] / epoch_loss_tensor[1]).item())
        train_losses.append(epoch_loss)
        loss_ema = epoch_loss if loss_ema is None else 0.9 * loss_ema + 0.1 * epoch_loss
        validation = _run_validation(model, val_loader, device, distributed)
        val_losses.append(validation)
        if rank == 0:
            print(
                f"epoch={epoch + 1}/{training['epochs']} "
                f"train_loss={epoch_loss:.6f} train_ema={loss_ema:.6f} val_loss={validation:.6f}"
            )
        if args.max_steps is not None and completed_steps >= args.max_steps:
            break
    checkpoint_path.parent.mkdir(parents=True, exist_ok=True)
    if rank == 0:
        state_dict = model.module.state_dict() if distributed else model.state_dict()
        torch.save({"model": state_dict, "config": config, "epoch": len(train_losses), "step": completed_steps}, checkpoint_path)
        np.save(checkpoint_path.parent / "train_loss.npy", np.asarray(train_losses, dtype=np.float32))
        np.save(checkpoint_path.parent / "val_loss.npy", np.asarray(val_losses, dtype=np.float32))
    if not math.isfinite(train_losses[-1]):
        raise FloatingPointError("final training loss is not finite")
    if rank == 0:
        print(f"saved checkpoint: {checkpoint_path}")
    if distributed:
        dist.destroy_process_group()


if __name__ == "__main__":
    main()