src/builder.py

import torch
import torch.nn as nn
from pathlib import Path

from .stage1.medarc_architecture import MultiSubjectConvLinearEncoder
from .stage2.CFM import CFM

def build_models(cfg, ckpt_dir: Path, all_features, subjects, device, stage2_ckpt_name=None):
    """Build and load stage 1 and stage 2 models from checkpoints."""
    sample_episode = next(iter(all_features[0]))
    feat_dims = [feats[sample_episode].shape[-1] for feats in all_features]
    print(f"  Feature dims: {feat_dims}")

    stage1_model = MultiSubjectConvLinearEncoder(
        num_subjects=len(subjects),
        feat_dims=feat_dims,
        **cfg.stage1.model,
    ).to(device)

    stage1_path = ckpt_dir / "stage1_best.pt"
    print(f"  Loading stage 1: {stage1_path}")
    stage1_model.load_state_dict(torch.load(stage1_path, map_location=device, weights_only=True))
    stage1_model.eval()

    target_dim = 1000
    cfm_params = cfg.stage2.cfm
    decoder_params = cfg.stage2.decoder
    latent_dim = cfg.stage2.get("latent_dim", 128)

    stage2_models = nn.ModuleDict()
    for sub in subjects:
        sub_key = str(sub)
        cfm_model = CFM(
            in_channels=2 * target_dim,
            out_channel=target_dim,
            cfm_params=cfm_params,
            decoder_params=decoder_params,
            n_spks=1,
            voxel_dim=target_dim,
            latent_dim=latent_dim,
        )
        stage2_models[sub_key] = cfm_model

    if stage2_ckpt_name:
        stage2_path = ckpt_dir / stage2_ckpt_name
    else:
        stage2_paths = sorted(ckpt_dir.glob("stage2_epoch_*.pt"))
        if not stage2_paths:
            raise FileNotFoundError(f"No stage2 checkpoints found in {ckpt_dir}")
        stage2_path = stage2_paths[-1]

    print(f"  Loading stage 2: {stage2_path}")
    stage2_models.load_state_dict(torch.load(stage2_path, map_location=device, weights_only=True))
    stage2_models = stage2_models.to(device)
    stage2_models.eval()

    return stage1_model, stage2_models