fcdm_diffae/model.py

"""FCDMDiffAE: standalone HuggingFace-compatible diffusion autoencoder."""

from __future__ import annotations

from pathlib import Path

import torch
from torch import Tensor, nn

from .config import FCDMDiffAEConfig, FCDMDiffAEInferenceConfig
from .decoder import Decoder
from .encoder import Encoder, EncoderPosterior
from .samplers import run_ddim, run_dpmpp_2m
from .vp_diffusion import get_schedule, make_initial_state, sample_noise


def _resolve_model_dir(
    path_or_repo_id: str | Path,
    *,
    revision: str | None,
    cache_dir: str | Path | None,
) -> Path:
    """Resolve a local path or HuggingFace Hub repo ID to a local directory."""
    local = Path(path_or_repo_id)
    if local.is_dir():
        return local
    repo_id = str(path_or_repo_id)
    try:
        from huggingface_hub import snapshot_download
    except ImportError:
        raise ImportError(
            f"'{repo_id}' is not an existing local directory. "
            "To download from HuggingFace Hub, install huggingface_hub: "
            "pip install huggingface_hub"
        )
    cache_dir_str = str(cache_dir) if cache_dir is not None else None
    local_dir = snapshot_download(
        repo_id,
        revision=revision,
        cache_dir=cache_dir_str,
    )
    return Path(local_dir)


class FCDMDiffAE(nn.Module):
    """Standalone FCDM DiffAE model for HuggingFace distribution.

    A diffusion autoencoder built on FCDM (Fully Convolutional Diffusion Model)
    blocks. Encodes images to compact 128-channel spatial latents via a
    VP-parameterized diagonal Gaussian posterior, and decodes them back via
    iterative VP diffusion with a skip-concat decoder.

    Usage::

        model = FCDMDiffAE.from_pretrained("path/to/weights")
        model = model.to("cuda", dtype=torch.bfloat16)

        # Encode (returns posterior mode by default)
        latents = model.encode(images)  # images: [B,3,H,W] in [-1,1]

        # Decode (1 step by default — PSNR-optimal)
        recon = model.decode(latents, height=H, width=W)

        # Reconstruct (encode + 1-step decode)
        recon = model.reconstruct(images)
    """

    _LATENT_NORM_EPS: float = 1e-4

    def __init__(self, config: FCDMDiffAEConfig) -> None:
        super().__init__()
        self.config = config

        # Latent running stats for whitening/dewhitening (at exported latent channels)
        self.register_buffer(
            "latent_norm_running_mean",
            torch.zeros((config.latent_channels,), dtype=torch.float32),
        )
        self.register_buffer(
            "latent_norm_running_var",
            torch.ones((config.latent_channels,), dtype=torch.float32),
        )

        self.encoder = Encoder(
            in_channels=config.in_channels,
            patch_size=config.patch_size,
            model_dim=config.model_dim,
            depth=config.encoder_depth,
            bottleneck_dim=config.bottleneck_dim,
            mlp_ratio=config.mlp_ratio,
            depthwise_kernel_size=config.depthwise_kernel_size,
            bottleneck_posterior_kind=config.bottleneck_posterior_kind,
            bottleneck_norm_mode=config.bottleneck_norm_mode,
        )

        self.decoder = Decoder(
            in_channels=config.in_channels,
            patch_size=config.patch_size,
            model_dim=config.model_dim,
            depth=config.decoder_depth,
            start_block_count=config.decoder_start_blocks,
            end_block_count=config.decoder_end_blocks,
            bottleneck_dim=config.bottleneck_dim,
            mlp_ratio=config.mlp_ratio,
            depthwise_kernel_size=config.depthwise_kernel_size,
            adaln_low_rank_rank=config.adaln_low_rank_rank,
        )

    @classmethod
    def from_pretrained(
        cls,
        path_or_repo_id: str | Path,
        *,
        dtype: torch.dtype = torch.bfloat16,
        device: str | torch.device = "cpu",
        revision: str | None = None,
        cache_dir: str | Path | None = None,
    ) -> FCDMDiffAE:
        """Load a pretrained model from a local directory or HuggingFace Hub.

        The directory (or repo) should contain:
        - config.json: Model architecture config.
        - model.safetensors (preferred) or model.pt: Model weights.

        Args:
            path_or_repo_id: Local directory path or HuggingFace Hub repo ID.
            dtype: Load weights in this dtype (float32 or bfloat16).
            device: Target device.
            revision: Git revision for Hub downloads.
            cache_dir: Where to cache Hub downloads.

        Returns:
            Loaded model in eval mode.
        """
        model_dir = _resolve_model_dir(
            path_or_repo_id, revision=revision, cache_dir=cache_dir
        )
        config = FCDMDiffAEConfig.load(model_dir / "config.json")
        model = cls(config)

        safetensors_path = model_dir / "model.safetensors"
        pt_path = model_dir / "model.pt"

        if safetensors_path.exists():
            try:
                from safetensors.torch import load_file

                state_dict = load_file(str(safetensors_path), device=str(device))
            except ImportError:
                raise ImportError(
                    "safetensors package required to load .safetensors files. "
                    "Install with: pip install safetensors"
                )
        elif pt_path.exists():
            state_dict = torch.load(
                str(pt_path), map_location=device, weights_only=True
            )
        else:
            raise FileNotFoundError(
                f"No model weights found in {model_dir}. "
                "Expected model.safetensors or model.pt."
            )

        model.load_state_dict(state_dict)
        model = model.to(dtype=dtype, device=torch.device(device))
        model.eval()
        return model

    def _latent_norm_stats(self) -> tuple[Tensor, Tensor]:
        """Return (mean, std) tensors for latent whitening, shaped [1,C,1,1]."""
        mean = self.latent_norm_running_mean.view(1, -1, 1, 1)
        var = self.latent_norm_running_var.view(1, -1, 1, 1)
        std = torch.sqrt(var.to(torch.float32) + self._LATENT_NORM_EPS)
        return mean.to(torch.float32), std

    def whiten(self, latents: Tensor) -> Tensor:
        """Whiten encoder latents using per-channel running stats.

        Use this before passing latents to a downstream latent-space
        diffusion model. The whitened latents have approximately zero mean
        and unit variance per channel.

        Args:
            latents: [B, bottleneck_dim, h, w] raw encoder output.

        Returns:
            Whitened latents [B, bottleneck_dim, h, w] in float32.
        """
        z = latents.to(torch.float32)
        mean, std = self._latent_norm_stats()
        return (z - mean.to(device=z.device)) / std.to(device=z.device)

    def dewhiten(self, latents: Tensor) -> Tensor:
        """Undo whitening to recover raw encoder latent scale.

        Use this before passing whitened latents back to ``decode()``.

        Args:
            latents: [B, bottleneck_dim, h, w] whitened latents.

        Returns:
            Dewhitened latents [B, bottleneck_dim, h, w] in float32.
        """
        z = latents.to(torch.float32)
        mean, std = self._latent_norm_stats()
        return z * std.to(device=z.device) + mean.to(device=z.device)

    def _patchify(self, z: Tensor) -> Tensor:
        """2x2 patchify: [B, C, H, W] -> [B, 4C, H/2, W/2]."""
        b, c, h, w = z.shape
        z = z.reshape(b, c, h // 2, 2, w // 2, 2)
        z = z.permute(0, 1, 3, 5, 2, 4)
        return z.reshape(b, c * 4, h // 2, w // 2)

    def _unpatchify(self, z: Tensor) -> Tensor:
        """2x2 unpatchify: [B, 4C, H/2, W/2] -> [B, C, H, W]."""
        b, c, h, w = z.shape
        z = z.reshape(b, c // 4, 2, 2, h, w)
        z = z.permute(0, 1, 4, 2, 5, 3)
        return z.reshape(b, c // 4, h * 2, w * 2)

    def encode(self, images: Tensor) -> Tensor:
        """Encode images to whitened latents (posterior mode).

        Returns latents whitened using per-channel running stats, ready for
        use by downstream latent-space diffusion models.

        Args:
            images: [B, 3, H, W] in [-1, 1], H and W divisible by
                effective_patch_size.

        Returns:
            Whitened latents [B, latent_channels, H/effective_patch, W/effective_patch].
        """
        eff_patch = self.config.effective_patch_size
        h, w = int(images.shape[2]), int(images.shape[3])
        if h % eff_patch != 0 or w % eff_patch != 0:
            raise ValueError(
                f"Image height={h} and width={w} must be divisible by "
                f"effective_patch_size={eff_patch}"
            )
        try:
            model_dtype = next(self.parameters()).dtype
        except StopIteration:
            model_dtype = torch.float32
        z = self.encoder(images.to(dtype=model_dtype))
        if self.config.bottleneck_patchify_mode == "patch_2x2":
            z = self._patchify(z)
        return self.whiten(z).to(dtype=model_dtype)

    def encode_posterior(self, images: Tensor) -> EncoderPosterior:
        """Encode images and return the full posterior (mean + logsnr).

        In patch-32 mode, the posterior is returned in the patchified space
        (512ch at H/32), consistent with encode() and whiten().

        Args:
            images: [B, 3, H, W] in [-1, 1], H and W divisible by
                effective_patch_size.

        Returns:
            EncoderPosterior with mean and logsnr tensors in the exported
            latent space.
        """
        try:
            model_dtype = next(self.parameters()).dtype
        except StopIteration:
            model_dtype = torch.float32
        posterior = self.encoder.encode_posterior(images.to(dtype=model_dtype))
        if self.config.bottleneck_patchify_mode == "patch_2x2":
            return EncoderPosterior(
                mean=self._patchify(posterior.mean),
                logsnr=self._patchify(posterior.logsnr),
            )
        return posterior

    @torch.no_grad()
    def decode(
        self,
        latents: Tensor,
        height: int,
        width: int,
        *,
        inference_config: FCDMDiffAEInferenceConfig | None = None,
    ) -> Tensor:
        """Decode whitened latents to images via VP diffusion.

        Latents are dewhitened and (if applicable) unpatchified internally
        before being passed to the decoder.

        Args:
            latents: [B, latent_channels, h, w] whitened encoder latents.
            height: Output image height (divisible by effective_patch_size).
            width: Output image width (divisible by effective_patch_size).
            inference_config: Optional inference parameters.

        Returns:
            Reconstructed images [B, 3, H, W] in float32.
        """
        cfg = inference_config or FCDMDiffAEInferenceConfig()
        config = self.config
        batch = int(latents.shape[0])
        device = latents.device

        try:
            model_dtype = next(self.parameters()).dtype
        except StopIteration:
            model_dtype = torch.float32

        eff_patch = config.effective_patch_size
        if height % eff_patch != 0 or width % eff_patch != 0:
            raise ValueError(
                f"height={height} and width={width} must be divisible by "
                f"effective_patch_size={eff_patch}"
            )

        # Dewhiten and unpatchify back to raw encoder scale for the decoder
        latents = self.dewhiten(latents)
        if config.bottleneck_patchify_mode == "patch_2x2":
            latents = self._unpatchify(latents)
        latents = latents.to(dtype=model_dtype)

        shape = (batch, config.in_channels, height, width)
        noise = sample_noise(
            shape,
            noise_std=config.pixel_noise_std,
            seed=cfg.seed,
            device=torch.device("cpu"),
            dtype=torch.float32,
        )

        schedule = get_schedule(cfg.schedule, cfg.num_steps).to(device=device)
        initial_state = make_initial_state(
            noise=noise.to(device=device),
            t_start=schedule[0:1],
            logsnr_min=config.logsnr_min,
            logsnr_max=config.logsnr_max,
        )

        device_type = "cuda" if device.type == "cuda" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            latents_in = latents.to(device=device)

            def _forward_fn(
                x_t: Tensor,
                t: Tensor,
                latents: Tensor,
                *,
                drop_middle_blocks: bool = False,
                mask_latent_tokens: bool = False,
            ) -> Tensor:
                return self.decoder(
                    x_t.to(dtype=model_dtype),
                    t,
                    latents.to(dtype=model_dtype),
                    drop_middle_blocks=drop_middle_blocks,
                )

            pdg_mode = "path_drop" if cfg.pdg else "disabled"

            if cfg.sampler == "ddim":
                sampler_fn = run_ddim
            elif cfg.sampler == "dpmpp_2m":
                sampler_fn = run_dpmpp_2m
            else:
                raise ValueError(
                    f"Unsupported sampler: {cfg.sampler!r}. Use 'ddim' or 'dpmpp_2m'."
                )

            result = sampler_fn(
                forward_fn=_forward_fn,
                initial_state=initial_state,
                schedule=schedule,
                latents=latents_in,
                logsnr_min=config.logsnr_min,
                logsnr_max=config.logsnr_max,
                pdg_mode=pdg_mode,
                pdg_strength=cfg.pdg_strength,
                device=device,
            )

        return result

    @torch.no_grad()
    def reconstruct(
        self,
        images: Tensor,
        *,
        inference_config: FCDMDiffAEInferenceConfig | None = None,
    ) -> Tensor:
        """Encode then decode. Convenience wrapper.

        Args:
            images: [B, 3, H, W] in [-1, 1].
            inference_config: Optional inference parameters.

        Returns:
            Reconstructed images [B, 3, H, W] in float32.
        """
        latents = self.encode(images)
        _, _, h, w = images.shape
        return self.decode(
            latents, height=h, width=w, inference_config=inference_config
        )