JiT-B-16-SIM/pipeline.py

"""Hub custom pipeline: JiTPipeline for FD-Loss post-trained JiT checkpoints.

Uses FD-Loss sampling (legacy time convention, velocity Euler/Heun, t: 1→0).
See libs/FD-Loss-diffusers and scripts/evaluate_released_ckpt.sh (JiT_B preset).
"""

from __future__ import annotations

import json
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Union

import torch
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from diffusers.utils.torch_utils import randn_tensor

from scheduling_flow_match_fd import FDLossFlowMatchScheduler

RECOMMENDED_NOISE_BY_SIZE = {
    256: 1.0,
    512: 2.0,
}

RECOMMENDED_CFG_BY_VARIANT = {
    "JiT-B": 3.0,
    "JiT-L": 2.4,
    "JiT-H": 2.2,
}


class JiTPipeline(DiffusionPipeline):
    r"""
    Pipeline for FD-Loss post-trained JiT (flow matching, legacy time convention).

    Parameters:
        transformer ([`JiTTransformer2DModel`]):
            Class-conditioned JiT backbone.
        scheduler ([`FDLossFlowMatchScheduler`]):
            Flow timesteps from 1 (noise) to 0 (data).
        legacy_time_convention (`bool`, *optional*, defaults to `True`):
            Flip flow time when passing to the backbone (`t_bb = 1 - t`), as in FD-Loss training.
        id2label (`dict[int, str]`, *optional*):
            ImageNet class id to English label mapping.
    """

    model_cpu_offload_seq = "transformer"

    def __init__(
        self,
        transformer,
        scheduler=None,
        id2label: Optional[Dict[Union[int, str], str]] = None,
        legacy_time_convention: bool = True,
    ):
        super().__init__()
        if scheduler is None:
            scheduler = FDLossFlowMatchScheduler()
        self.register_modules(transformer=transformer, scheduler=scheduler)
        self.legacy_time_convention = legacy_time_convention
        self._id2label = self._normalize_id2label(id2label)
        self.labels = self._build_label2id(self._id2label)
        self._labels_loaded_from_model_index = bool(self._id2label)

    def _backbone_t(self, t: torch.Tensor) -> torch.Tensor:
        if self.legacy_time_convention:
            return 1.0 - t
        return t

    def _normalize_class_labels(
        self,
        class_labels: Union[int, str, List[Union[int, str]]],
    ) -> List[int]:
        if isinstance(class_labels, int):
            return [class_labels]
        if isinstance(class_labels, str):
            return self.get_label_ids(class_labels)
        if class_labels and isinstance(class_labels[0], str):
            return self.get_label_ids(class_labels)
        return list(class_labels)

    def _forward_velocity(
        self,
        z: torch.Tensor,
        t: torch.Tensor,
        labels: torch.Tensor,
        guidance_scale: float,
        cfg_interval: Optional[Tuple[float, float]],
        t_eps: float,
    ) -> torch.Tensor:
        t_view = t.reshape(-1, *([1] * (z.ndim - 1)))
        t_bb = self._backbone_t(t).flatten().expand(z.shape[0])

        x_cond = self.transformer(
            z,
            timestep=t_bb,
            class_labels=labels,
            interpolate_pos_encoding=interpolate_pos_encoding,
        ).sample
        v_cond = (z - x_cond) / t_view.clamp_min(t_eps)

        if guidance_scale == 1.0:
            return v_cond

        null_class = int(
            getattr(self.transformer.config, "num_classes", getattr(self.transformer.config, "num_class_embeds", 1000))
        )
        class_null = torch.full_like(labels, null_class)
        x_uncond = self.transformer(
            z,
            timestep=t_bb,
            class_labels=class_null,
            interpolate_pos_encoding=interpolate_pos_encoding,
        ).sample
        v_uncond = (z - x_uncond) / t_view.clamp_min(t_eps)

        if cfg_interval is None:
            return v_uncond + guidance_scale * (v_cond - v_uncond)

        low, high = cfg_interval
        mask = (t < high) & ((low == 0) | (t > low))
        scale = torch.where(
            mask,
            torch.tensor(guidance_scale, device=z.device, dtype=z.dtype),
            torch.tensor(1.0, device=z.device, dtype=z.dtype),
        )
        while scale.ndim < v_cond.ndim:
            scale = scale.unsqueeze(-1)
        return v_uncond + scale * (v_cond - v_uncond)

    @staticmethod
    def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
        if not id2label:
            return {}
        return {int(key): value for key, value in id2label.items()}

    @staticmethod
    def _read_id2label_from_model_index(variant_path: Optional[str]) -> Dict[int, str]:
        if not variant_path:
            return {}
        variant_dir = Path(variant_path).resolve()
        model_index_path = variant_dir / "model_index.json"
        if not model_index_path.exists():
            return {}
        raw = json.loads(model_index_path.read_text(encoding="utf-8"))
        id2label = raw.get("id2label")
        if not isinstance(id2label, dict):
            return {}
        return {int(key): value for key, value in id2label.items()}

    @staticmethod
    def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
        label2id: Dict[str, int] = {}
        for class_id, value in id2label.items():
            for synonym in value.split(","):
                synonym = synonym.strip()
                if synonym:
                    label2id[synonym] = int(class_id)
        return dict(sorted(label2id.items()))

    def _ensure_labels_loaded(self) -> None:
        if self._labels_loaded_from_model_index:
            return
        loaded = self._read_id2label_from_model_index(getattr(self.config, "_name_or_path", None))
        if loaded:
            self._id2label = loaded
            self.labels = self._build_label2id(self._id2label)
        self._labels_loaded_from_model_index = True

    @property
    def id2label(self) -> Dict[int, str]:
        self._ensure_labels_loaded()
        return self._id2label

    def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
        self._ensure_labels_loaded()
        label2id = self.labels
        if not label2id:
            raise ValueError("No English labels loaded. Ensure `id2label` exists in model_index.json.")
        if isinstance(label, str):
            label = [label]
        missing = [item for item in label if item not in label2id]
        if missing:
            preview = ", ".join(list(label2id.keys())[:8])
            raise ValueError(f"Unknown English label(s): {missing}. Example valid labels: {preview}, ...")
        return [label2id[item] for item in label]

    @torch.inference_mode()
    def __call__(
        self,
        class_labels: Union[int, str, List[Union[int, str]]],
        num_inference_steps: int = 1,
        guidance_scale: float = 3.0,
        guidance_interval_min: float = 0.1,
        guidance_interval_max: float = 1.0,
        sampling_method: str = "euler",
        noise_scale: Optional[float] = None,
        t_eps: float = 5e-2,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        interpolate_pos_encoding: bool = True,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
    ) -> Union[ImagePipelineOutput, Tuple]:
        if num_inference_steps < 1:
            raise ValueError("num_inference_steps must be >= 1.")
        if output_type not in {"pil", "np", "pt"}:
            raise ValueError("output_type must be one of: 'pil', 'np', 'pt'.")
        if sampling_method not in {"euler", "heun"}:
            raise ValueError("sampling_method must be 'euler' or 'heun'.")

        class_label_ids = self._normalize_class_labels(class_labels)
        batch_size = len(class_label_ids)
        image_size = int(self.transformer.config.sample_size)
        patch_size = int(self.transformer.config.patch_size)
        height = int(height or image_size)
        width = int(width or image_size)
        if height % patch_size != 0 or width % patch_size != 0:
            raise ValueError(
                f"height and width must be divisible by patch_size={patch_size}. Got {(height, width)}."
            )
        channels = int(self.transformer.config.in_channels)

        if noise_scale is None:
            noise_scale = RECOMMENDED_NOISE_BY_SIZE.get(max(height, width), 1.0)

        z = randn_tensor(
            shape=(batch_size, channels, height, width),
            generator=generator,
            device=self._execution_device,
            dtype=self.transformer.dtype,
        ) * noise_scale

        labels = torch.tensor(class_label_ids, device=self._execution_device, dtype=torch.long).reshape(-1)
        null_class_val = int(
            getattr(self.transformer.config, "num_classes", getattr(self.transformer.config, "num_class_embeds", 1000))
        )
        labels = labels.clamp(0, null_class_val - 1)

        cfg_interval = [
            float(self._backbone_t(torch.tensor(guidance_interval_min, device=z.device))),
            float(self._backbone_t(torch.tensor(guidance_interval_max, device=z.device))),
        ]
        cfg_interval = (min(cfg_interval), max(cfg_interval))

        timesteps = self.scheduler.set_timesteps(num_inference_steps, device=self._execution_device)
        ts = timesteps.view(-1, *([1] * z.ndim)).expand(-1, batch_size, -1, -1, -1)

        for i in self.progress_bar(range(num_inference_steps - 1)):
            t_cur = ts[i]
            t_next = ts[i + 1]
            if sampling_method == "heun":
                dt = t_next - t_cur
                v1 = self._forward_velocity(z, t_cur, labels, guidance_scale, cfg_interval, t_eps)
                z_mid = z + dt * v1
                v2 = self._forward_velocity(z_mid, t_next, labels, guidance_scale, cfg_interval, t_eps)
                z = z + dt * 0.5 * (v1 + v2)
            else:
                v = self._forward_velocity(z, t_cur, labels, guidance_scale, cfg_interval, t_eps)
                z = z + (t_next - t_cur) * v

        if num_inference_steps >= 1:
            t_cur = ts[-2]
            t_next = ts[-1]
            v = self._forward_velocity(z, t_cur, labels, guidance_scale, cfg_interval, t_eps)
            z = z + (t_next - t_cur) * v

        images_pt = ((z.float().clamp(-1, 1) + 1.0) / 2.0).cpu()
        if output_type == "pt":
            images = images_pt
        elif output_type == "np":
            images = images_pt.permute(0, 2, 3, 1).numpy()
        else:
            images = self.numpy_to_pil(images_pt.permute(0, 2, 3, 1).numpy())

        self.maybe_free_model_hooks()

        if not return_dict:
            return (images,)
        return ImagePipelineOutput(images=images)