scripts/train.py

import sys
from pathlib import Path

# 获取项目根目录（train.py上级的上级）
root_path = Path(__file__).parent.parent
sys.path.append(str(root_path))
import torch
import os
import numpy as np
import torch.distributed as dist
import logging
import time

from model.dgmr import DGMR
from model.dgmr_official.losses import loss_hinge_disc, loss_hinge_gen
from onescience.datapipes.climate import ERA5Datapipe
from onescience.utils.YParams import YParams

try:
    from apex import optimizers
    _FUSED_ADAM = True
except Exception:
    _FUSED_ADAM = False


def main():
    logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
    logger = logging.getLogger()

    ## Model config init
    config_file_path = os.path.join(current_path, "conf/config.yaml")
    cfg = YParams(config_file_path, "model")

    ## Distributed config init
    cfg.world_size = 1
    if "WORLD_SIZE" in os.environ:
        cfg.world_size = int(os.environ["WORLD_SIZE"])
    world_rank = 0
    local_rank = 0
    if cfg.world_size > 1 and torch.cuda.is_available():
        dist.init_process_group(backend="nccl", init_method="env://")
        local_rank = int(os.environ["LOCAL_RANK"])
        world_rank = dist.get_rank()
    device = f"cuda:{local_rank}" if torch.cuda.is_available() else "cpu"

    ## DataLoader init
    cfg_data = YParams(config_file_path, "datapipe")
    cfg['N_in_channels'] = len(cfg_data.dataset.channels)
    cfg['N_out_channels'] = len(cfg_data.dataset.channels)
    datapipe = ERA5Datapipe(
        dataset_dir=cfg_data.dataset.data_dir,
        used_variables=cfg_data.dataset.channels,
        used_years=cfg_data.dataset.train_time,
        distributed=dist.is_initialized(),
        input_steps=cfg.num_context,
        output_steps=cfg.forecast_steps,
        batch_size=cfg_data.dataloader.batch_size,
        num_workers=cfg_data.dataloader.num_workers,
    )
    train_dataloader, train_sampler = datapipe.get_dataloader("train")
    datapipe = ERA5Datapipe(
        dataset_dir=cfg_data.dataset.data_dir,
        used_variables=cfg_data.dataset.channels,
        used_years=cfg_data.dataset.val_time,
        distributed=dist.is_initialized(),
        input_steps=cfg.num_context,
        output_steps=cfg.forecast_steps,
        batch_size=cfg_data.dataloader.batch_size,
        num_workers=cfg_data.dataloader.num_workers,
    )
    val_dataloader, val_sampler = datapipe.get_dataloader("valid")

    # Model init
    model = DGMR(
        forecast_steps=cfg.forecast_steps,
        num_context=cfg.num_context,
        input_channels=cfg.input_channels,
        output_shape=cfg.output_shape,
        conv_type=cfg.conv_type,
        latent_channels=cfg.latent_channels,
        context_channels=cfg.context_channels,
        generation_steps=cfg.generation_steps,
        grid_lambda=cfg.grid_lambda,
        precip_weight_cap=cfg.precip_weight_cap,
    ).to(device)

    # 生成器与判别器使用独立的 Adam 优化器（论文 lr=1e-4）
    if _FUSED_ADAM:
        optimizer_g = optimizers.FusedAdam(model.generator.parameters(), lr=cfg.lr)
        optimizer_d = optimizers.FusedAdam(model.discriminator.parameters(), lr=cfg.lr_disc)
    else:
        optimizer_g = torch.optim.Adam(model.generator.parameters(), lr=cfg.lr)
        optimizer_d = torch.optim.Adam(model.discriminator.parameters(), lr=cfg.lr_disc)
    scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, factor=0.2, patience=5, mode='min')
    scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, factor=0.2, patience=5, mode='min')

    ## Train process init
    os.makedirs(cfg.checkpoint_dir, exist_ok=True)
    train_loss_file = f"{cfg.checkpoint_dir}/trloss.npy"
    valid_loss_file = f"{cfg.checkpoint_dir}/valoss.npy"
    best_valid_loss = float("inf")
    best_loss_epoch = 0
    train_losses = np.empty((0,), dtype=np.float32)
    valid_losses = np.empty((0,), dtype=np.float32)

    ## Get model params count
    if cfg.world_size == 1:
        total_params = sum(p.numel() for p in model.parameters())
        print("\n\n")
        print("-" * 50)
        print(f"📂 now params is {total_params}, {total_params / 1e6:.2f}M, {total_params / 1e9:.2f}B")
        print("-" * 50, "\n")

    ## Load model weight if there exist well-trained model
    if os.path.exists(f"{cfg.checkpoint_dir}/model_bak.pth"):
        if world_rank == 0:
            print("\n\n")
            print("-" * 50)
            print(f"✅ There has a model weight, load and continue training...")
            print(f'If you want to train a new model, ensure there is no *.pth file in {cfg.checkpoint_dir}')
            print("-" * 50, "\n")
        ckpt = torch.load(f"{cfg.checkpoint_dir}/model_bak.pth", map_location=device, weights_only=False)
        model.load_state_dict(ckpt["model_state_dict"])
        optimizer_g.load_state_dict(ckpt["optimizer_g_state_dict"])
        optimizer_d.load_state_dict(ckpt["optimizer_d_state_dict"])
        scheduler_g.load_state_dict(ckpt["scheduler_g_state_dict"])
        scheduler_d.load_state_dict(ckpt["scheduler_d_state_dict"])
        best_valid_loss = ckpt["best_valid_loss"]
        best_loss_epoch = ckpt["best_loss_epoch"]
        train_losses = np.load(train_loss_file)
        valid_losses = np.load(valid_loss_file)

    ## Distributed model
    if dist.is_initialized():
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[local_rank], output_device=local_rank)
    world_rank == 0 and logger.info(f"start training ...")

    for epoch in range(cfg.max_epoch):
        if dist.is_initialized():
            train_sampler.set_epoch(epoch)
            val_sampler.set_epoch(epoch)
        model.train()
        train_loss = 0
        start_time = time.time()
        for j, data in enumerate(train_dataloader):
            invar = data[0].to(device, dtype=torch.float32)      # [B, num_context, C, H, W]
            outvar = data[1].to(device, dtype=torch.float32)     # [B, forecast_steps, C, H, W]
            full_real = torch.cat([invar, outvar], dim=1)        # 上下文 + 真实未来帧

            # --- 判别器（hinge loss，生成图像 detach 不传梯度） ---
            gen_images = model.generator(invar)                  # [B, forecast_steps, C, H, W]
            full_fake = torch.cat([invar, gen_images], dim=1)
            score_real = model.discriminator(full_real)
            score_generated = model.discriminator(full_fake.detach())
            disc_loss = loss_hinge_disc(score_generated, score_real)
            optimizer_d.zero_grad()
            disc_loss.backward()
            optimizer_d.step()

            # --- 生成器（hinge + 网格单元正则器，MC 采样估计期望） ---
            score_generated = model.discriminator(full_fake)
            gen_loss = loss_hinge_gen(score_generated)
            gen_samples = torch.stack(
                [model.generator(invar) for _ in range(cfg.generation_steps)], dim=0
            ).mean(dim=0)                                        # 取 MC 均值作为生成均值图像
            grid_loss = model.grid_regularizer(gen_samples, outvar)
            gen_loss = gen_loss + cfg.grid_lambda * grid_loss
            optimizer_g.zero_grad()
            gen_loss.backward()
            optimizer_g.step()

            loss = gen_loss.item() + disc_loss.item()
            train_loss += loss
            if world_rank == 0:
                logger.info(f'Train: Epoch {epoch}-{j+1}/{len(train_dataloader)} '
                            f'[cost {int((time.time()-start_time) // 60):02}:{int((time.time()-start_time) % 60):02}] '
                            f'[{(time.time()-start_time)/(j+1): .02f}s/{cfg_data.dataloader.batch_size}batch] '
                            f'gen_loss:{gen_loss.item(): .04f} disc_loss:{disc_loss.item(): .04f} '
                            f'loss:{train_loss / (j+1): .04f}')

        train_loss /= len(train_dataloader)

        # 验证只使用网格单元正则器（确定性，单个生成样本）
        model.eval()
        valid_loss = 0
        with torch.no_grad():
            start_time = time.time()
            for j, data in enumerate(val_dataloader):
                invar = data[0].to(device, dtype=torch.float32)
                outvar = data[1].to(device, dtype=torch.float32)
                gen_images = model.generator(invar)
                grid_loss = model.grid_regularizer(gen_images, outvar)
                loss = grid_loss

                if dist.is_initialized():
                    loss_tensor = loss.detach().to(device)
                    dist.all_reduce(loss_tensor)
                    loss = loss_tensor.item() / cfg.world_size
                    valid_loss += loss
                else:
                    valid_loss += loss.item()
                if world_rank == 0:
                    logger.info(f'Valid: Epoch {epoch}-{j+1}/{len(val_dataloader)} '
                            f'[cost {int((time.time()-start_time) // 60):02}:{int((time.time()-start_time) % 60):02}] '
                            f'loss:{valid_loss / (j+1): .04f}')

        valid_loss /= len(val_dataloader)
        is_save_ckp = False
        if valid_loss < best_valid_loss:
            best_valid_loss = valid_loss
            best_loss_epoch = epoch
            world_rank == 0 and save_checkpoint(model, optimizer_g, optimizer_d, scheduler_g, scheduler_d, best_valid_loss, best_loss_epoch, cfg.checkpoint_dir)
            is_save_ckp = True
        scheduler_g.step(valid_loss)
        scheduler_d.step(valid_loss)

        if world_rank == 0:
            logger.info(f"Epoch [{epoch + 1}/{cfg.max_epoch}], "
                        f"Train Loss: {train_loss:.4f}, "
                        f"Valid Loss: {valid_loss:.4f}, "
                        f"Best loss at Epoch: {best_loss_epoch + 1}"
                        + (", saving checkpoint" if is_save_ckp else "")
                        )
            train_losses = np.append(train_losses, train_loss)
            valid_losses = np.append(valid_losses, valid_loss)
            np.save(train_loss_file, train_losses)
            np.save(valid_loss_file, valid_losses)

        if epoch - best_loss_epoch > cfg.patience:
            print(f"Loss has not decrease in {cfg.patience} epochs, stopping training...")
            exit()


def save_checkpoint(model, optimizer_g, optimizer_d, scheduler_g, scheduler_d, best_valid_loss, best_loss_epoch, model_path):
    model_to_save = model.module if hasattr(model, "module") else model
    state = {"model_state_dict": model_to_save.state_dict(),
             "optimizer_g_state_dict": optimizer_g.state_dict(),
             "optimizer_d_state_dict": optimizer_d.state_dict(),
             "scheduler_g_state_dict": scheduler_g.state_dict(),
             "scheduler_d_state_dict": scheduler_d.state_dict(),
             "best_valid_loss": best_valid_loss,
             "best_loss_epoch": best_loss_epoch,
            }
    torch.save(state, f"{model_path}/model.pth")
    ### the weight file saving may interrupted due to DCU queue limit, get a backup to ensure there at least has one model
    os.system(f"mv {model_path}/model.pth {model_path}/model_bak.pth")


if __name__ == "__main__":
    current_path = os.getcwd()
    sys.path.append(current_path)
    main()