code/eval_audioldm2_test_loss.py

"""Sweep audioldm2 ckpts (concat + controlnet variants) and compute test-set
MSE loss for each. Outputs JSON with {ckpt: test_loss}.

Mirrors `train_concat.py:val_loss` and `train_controlnet.py:val_loss` exactly:
fp16 forward, predicted noise vs gt noise MSE, deterministic per-batch
(noise, timesteps) so different ckpts are compared apples-to-apples.

Architecture is auto-detected from the ckpt path:
- `audioldm2_bg2fg_controlnet_rebalance/.../controlnet.pt` → ControlNet sidebranch
- `audioldm2_bg2fg_rebalance/.../unet/diffusion_pytorch_model.safetensors` → 16-ch concat UNet

Usage:
    python eval_audioldm2_test_loss.py --ckpts <hf_path1> <hf_path2> ... --out <json>
"""
import argparse, json, os, re, sys, time
from pathlib import Path

import torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download, snapshot_download
from torch.utils.data import DataLoader
from safetensors.torch import load_file

ROOT = Path("/nfs/turbo/coe-ahowens-nobackup/dingqy/AudioLDM-training-finetuning")
sys.path.insert(0, str(ROOT))

from diffusers import AudioLDM2Pipeline, DDPMScheduler                          # noqa
from audioldm2_v2.dataset import BG2FGDataset, waveform_to_log_mel              # noqa
from audioldm2_v2.empty_prompt_cache import expand_to_batch, load_cache         # noqa
from audioldm2_v2.model_concat import expand_unet_conv_in                       # noqa
from audioldm2_v2.model_controlnet import (
    AudioLDM2ControlNet, unet_forward_with_residuals,                           # noqa
)

HF_REPO = "AE-W/ckpt"
CACHE   = os.environ.get("HF_CACHE",
                         "/nfs/turbo/coe-ahowens-nobackup/dingqy/.cache/huggingface")


def detect_arch(ckpt_rel: str) -> str:
    """Returns 'controlnet' or 'concat' based on HF path conventions.
    For a local abs path (the audioldm2_bg2fg_controlnet_rebalance/.../step_*
    dir), peek at what file is inside to disambiguate."""
    if os.path.isabs(ckpt_rel):
        if (Path(ckpt_rel) / "controlnet.pt").exists():
            return "controlnet"
        if (Path(ckpt_rel) / "unet").is_dir():
            return "concat"
    if "controlnet" in ckpt_rel:
        return "controlnet"
    return "concat"


def download_ckpt_dir(ckpt_rel: str) -> Path:
    """ckpt_rel is the HF dir path like
    `audioldm2_bg2fg_controlnet_rebalance/step_00010000` — or an absolute
    path to a local step_NNNNNNNN dir (used when HF mirror is missing a step
    that's still on turbo)."""
    if os.path.isabs(ckpt_rel) and Path(ckpt_rel).is_dir():
        return Path(ckpt_rel)
    arch = detect_arch(ckpt_rel)
    if arch == "controlnet":
        local = hf_hub_download(repo_id=HF_REPO,
                                filename=f"{ckpt_rel}/controlnet.pt",
                                repo_type="dataset", cache_dir=CACHE)
        return Path(local).parent
    else:
        # concat: needs unet/diffusion_pytorch_model.safetensors
        local = hf_hub_download(repo_id=HF_REPO,
                                filename=f"{ckpt_rel}/unet/diffusion_pytorch_model.safetensors",
                                repo_type="dataset", cache_dir=CACHE)
        return Path(local).parent.parent     # the step_NNNN dir


def collate_test(batch):
    """Stack fg_wav and bg_wav, drop other fields."""
    fg = torch.stack([item["fg_wav"] for item in batch], dim=0)
    bg = torch.stack([item["bg_wav"] for item in batch], dim=0)
    return {"fg_wav": fg, "bg_wav": bg}


@torch.no_grad()
def encode_batch(pipe, fg_wav, bg_wav, vae_scale, device):
    fg_wav = fg_wav.to(device, non_blocking=True)
    bg_wav = bg_wav.to(device, non_blocking=True)
    mel_fg = waveform_to_log_mel(fg_wav)
    mel_bg = waveform_to_log_mel(bg_wav)
    z_fg = pipe.vae.encode(mel_fg).latent_dist.mean * vae_scale
    z_bg = pipe.vae.encode(mel_bg).latent_dist.mean * vae_scale
    return z_fg, z_bg


@torch.no_grad()
def forward_concat(pipe, z_fg, z_bg, noise, t, cond, noise_scheduler):
    z_noisy = noise_scheduler.add_noise(z_fg, noise, t)
    z_in = torch.cat([z_noisy, z_bg], dim=1)
    eps = pipe.unet(
        z_in, t,
        encoder_hidden_states=cond["encoder_hidden_states"],
        encoder_hidden_states_1=cond["encoder_hidden_states_1"],
        encoder_attention_mask_1=cond["encoder_attention_mask_1"],
        return_dict=False,
    )[0]
    return eps


@torch.no_grad()
def forward_controlnet(pipe, controlnet, z_fg, z_bg, noise, t, cond, noise_scheduler):
    z_noisy = noise_scheduler.add_noise(z_fg, noise, t)
    down_res, mid_res = controlnet(
        z_bg, t,
        encoder_hidden_states=cond["encoder_hidden_states"],
        encoder_hidden_states_1=cond["encoder_hidden_states_1"],
        encoder_attention_mask_1=cond["encoder_attention_mask_1"],
    )
    eps = unet_forward_with_residuals(
        pipe.unet, z_noisy, t,
        encoder_hidden_states=cond["encoder_hidden_states"],
        encoder_hidden_states_1=cond["encoder_hidden_states_1"],
        encoder_attention_mask_1=cond["encoder_attention_mask_1"],
        down_block_additional_residuals=down_res,
        mid_block_additional_residual=mid_res,
    )
    return eps


@torch.no_grad()
def test_loss_one(arch, pipe, controlnet, test_loader, cache, noise_scheduler,
                  vae_scale, device, num_train_timesteps=1000, seed_base=42):
    pipe.unet.eval()
    if controlnet is not None:
        controlnet.eval()
    total = 0.0
    n     = 0
    t0    = time.time()
    for i, batch in enumerate(test_loader):
        with torch.cuda.amp.autocast(dtype=torch.float16):
            z_fg, z_bg = encode_batch(pipe, batch["fg_wav"], batch["bg_wav"],
                                      vae_scale, device)
        bs = z_fg.shape[0]
        gen = torch.Generator(device=device).manual_seed(seed_base + i)
        noise = torch.randn(z_fg.shape, generator=gen, device=device, dtype=z_fg.dtype)
        timesteps = torch.randint(0, num_train_timesteps, (bs,),
                                  generator=gen, device=device, dtype=torch.long)
        cond = expand_to_batch(cache, bs)
        with torch.cuda.amp.autocast(dtype=torch.float16):
            if arch == "concat":
                eps = forward_concat(pipe, z_fg, z_bg, noise, timesteps, cond,
                                     noise_scheduler)
            else:
                eps = forward_controlnet(pipe, controlnet, z_fg, z_bg, noise,
                                         timesteps, cond, noise_scheduler)
            loss = F.mse_loss(eps.float(), noise.float(), reduction="mean")
        total += float(loss.item()) * bs
        n     += bs
        if i % 50 == 0:
            print(f"    batch {i:>4} cum_n={n}  rolling_loss={total/n:.5f}  ({time.time()-t0:.0f}s)",
                  flush=True)
    return total / max(n, 1), n


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--ckpts", nargs="+", required=True,
                    help="HF dir paths under AE-W/ckpt, e.g. audioldm2_bg2fg_controlnet_rebalance/step_00010000")
    ap.add_argument("--out", required=True)
    ap.add_argument("--manifest",
                    default="/nfs/turbo/coe-ahowens-nobackup/ymdou/hidingsound/data/noise_guidance_out_latent/manifest.json")
    ap.add_argument("--empty-cache",
                    default=str(ROOT / "audioldm2_v2/empty_prompt_cache_large.pt"))
    ap.add_argument("--model-id", default="cvssp/audioldm2-large")
    ap.add_argument("--batch-size", type=int, default=8)
    ap.add_argument("--num-workers", type=int, default=4)
    ap.add_argument("--limit", type=int, default=0,
                    help="cap test pairs (0 = all)")
    ap.add_argument("--seed-base", type=int, default=42)
    args = ap.parse_args()

    device = "cuda"

    print(f"loading pipeline {args.model_id}...", flush=True)
    # Mirror training-time setup (train_controlnet.py:384-393): pipe in fp32,
    # AudioLDM2ControlNet built on CPU first (its __init__ runs a CPU dummy
    # through base_unet.time_embedding/conv_in to size zero-convs), THEN
    # both moved to device. Pre-moving pipe.unet to GPU breaks the builder.
    pipe = AudioLDM2Pipeline.from_pretrained(args.model_id)
    pipe.vae.eval(); pipe.unet.eval()
    vae_scale = pipe.vae.config.scaling_factor

    noise_scheduler = DDPMScheduler.from_pretrained(args.model_id, subfolder="scheduler")
    num_T = noise_scheduler.config.num_train_timesteps

    cache = load_cache(args.empty_cache, device=device)

    print(f"loading test split from {args.manifest}", flush=True)
    test_ds = BG2FGDataset(
        manifest_path=args.manifest,
        split="test",
        prompt_stats_path=None,
        load_prompts_for_weights=False,
    )
    if args.limit:
        test_ds.pairs = test_ds.pairs[: args.limit]
    print(f"  {len(test_ds)} test pairs", flush=True)
    test_loader = DataLoader(
        test_ds, batch_size=args.batch_size, shuffle=False,
        num_workers=args.num_workers, collate_fn=collate_test, pin_memory=True,
    )

    results = {}
    if Path(args.out).exists():
        results = json.load(open(args.out))

    # Track current architecture state — pipe.unet's conv_in is mutated when
    # we switch from controlnet (8-ch) to concat (16-ch). We don't switch
    # back, so process all controlnet ckpts FIRST, then all concat.
    sorted_ckpts = sorted(args.ckpts, key=lambda c: (detect_arch(c) != "controlnet", c))
    pipe_is_concat = False
    controlnet_module = None

    for ckpt_rel in sorted_ckpts:
        if ckpt_rel in results:
            print(f"\n[skip] {ckpt_rel}  (cached test_loss={results[ckpt_rel]['test_loss']:.5f})",
                  flush=True)
            continue
        arch = detect_arch(ckpt_rel)
        print(f"\n=== {ckpt_rel}  ({arch}) ===", flush=True)
        ckpt_dir = download_ckpt_dir(ckpt_rel)

        if arch == "controlnet":
            assert not pipe_is_concat, "can't go concat→controlnet without re-init"
            sd = torch.load(ckpt_dir / "controlnet.pt", map_location="cpu",
                            weights_only=False)
            if controlnet_module is None:
                # Build on CPU first (matches train_controlnet.py order), then
                # move both to GPU together.
                controlnet_module = AudioLDM2ControlNet(pipe.unet)
                pipe.to(device)
                controlnet_module = controlnet_module.to(device)
                print(f"  built controlnet shell + moved pipe+cn to {device}")
            controlnet_module.load_state_dict(sd)
            controlnet_module.eval()
            print(f"  controlnet.pt loaded ({sum(v.numel() for v in sd.values()):,} params)")
            del sd
        else:
            if not pipe_is_concat:
                expand_unet_conv_in(pipe.unet, new_in_channels=16)
                pipe.to(device)
                pipe_is_concat = True
                controlnet_module = None    # not used for concat
                print(f"  expanded conv_in 8→16, moved pipe to {device}")
            sd = load_file(str(ckpt_dir / "unet" / "diffusion_pytorch_model.safetensors"))
            pipe.unet.load_state_dict(sd)
            pipe.unet.eval()
            print(f"  unet weights loaded ({sum(v.numel() for v in sd.values()):,} params)")
            del sd

        loss, n = test_loss_one(arch, pipe, controlnet_module, test_loader,
                                cache, noise_scheduler, vae_scale, device,
                                num_train_timesteps=num_T,
                                seed_base=args.seed_base)
        print(f"  test_loss = {loss:.5f}  (n={n})", flush=True)
        results[ckpt_rel] = {"test_loss": loss, "n": n, "arch": arch}
        with open(args.out, "w") as f:
            json.dump(results, f, indent=2)
        print(f"  saved → {args.out}", flush=True)

    # Final summary
    print("\n=== summary (sorted by test_loss) ===")
    sorted_results = sorted(results.items(), key=lambda x: x[1]["test_loss"])
    for ckpt_rel, info in sorted_results:
        print(f"  {info['test_loss']:.5f}  ({info['arch']:>10})  {ckpt_rel}")
    if sorted_results:
        best = sorted_results[0]
        print(f"\n>>> best: {best[0]}  test_loss={best[1]['test_loss']:.5f}")


if __name__ == "__main__":
    main()