inference.py

from typing import Any, Dict, Optional
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.schedulers import KarrasDiffusionSchedulers

import numpy
import torch
import torch.nn as nn
import torch.utils.checkpoint
import torch.distributed
import transformers
from collections import OrderedDict
from PIL import Image
from torchvision import transforms
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DiffusionPipeline,
    EulerAncestralDiscreteScheduler,
    UNet2DConditionModel,
    ImagePipelineOutput
)
from diffusers.image_processor import VaeImageProcessor
from diffusers.models.attention_processor import Attention, AttnProcessor, XFormersAttnProcessor, AttnProcessor2_0
from diffusers.utils.import_utils import is_xformers_available


def to_rgb_image(maybe_rgba: Image.Image):
    if maybe_rgba.mode == 'RGB':
        return maybe_rgba
    elif maybe_rgba.mode == 'RGBA':
        rgba = maybe_rgba
        img = numpy.random.randint(127, 128, size=[rgba.size[1], rgba.size[0], 3], dtype=numpy.uint8)
        img = Image.fromarray(img, 'RGB')
        img.paste(rgba, mask=rgba.getchannel('A'))
        return img
    else:
        raise ValueError("Unsupported image type.", maybe_rgba.mode)


class ReferenceOnlyAttnProc(torch.nn.Module):
    def __init__(
        self,
        chained_proc,
        enabled=False,
        name=None
    ) -> None:
        super().__init__()
        self.enabled = enabled
        self.chained_proc = chained_proc
        self.name = name

    def __call__(
        self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None,
        mode="w", ref_dict: dict = None, is_cfg_guidance = False
    ) -> Any:
        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        if self.enabled and is_cfg_guidance:
            res0 = self.chained_proc(attn, hidden_states[:1], encoder_hidden_states[:1], attention_mask)
            hidden_states = hidden_states[1:]
            encoder_hidden_states = encoder_hidden_states[1:]
        if self.enabled:
            if mode == 'w':
                ref_dict[self.name] = encoder_hidden_states
            elif mode == 'r':
                encoder_hidden_states = torch.cat([encoder_hidden_states, ref_dict.pop(self.name)], dim=1)
            elif mode == 'm':
                encoder_hidden_states = torch.cat([encoder_hidden_states, ref_dict[self.name]], dim=1)
            else:
                assert False, mode
        res = self.chained_proc(attn, hidden_states, encoder_hidden_states, attention_mask)
        if self.enabled and is_cfg_guidance:
            res = torch.cat([res0, res])
        return res


class RefOnlyNoisedUNet(torch.nn.Module):
    def __init__(self, unet: UNet2DConditionModel, train_sched: DDPMScheduler, val_sched: EulerAncestralDiscreteScheduler) -> None:
        super().__init__()
        self.unet = unet
        self.train_sched = train_sched
        self.val_sched = val_sched

        unet_lora_attn_procs = dict()
        for name, _ in unet.attn_processors.items():
            if torch.__version__ >= '2.0':
                default_attn_proc = AttnProcessor2_0()
            elif is_xformers_available():
                default_attn_proc = XFormersAttnProcessor()
            else:
                default_attn_proc = AttnProcessor()
            unet_lora_attn_procs[name] = ReferenceOnlyAttnProc(
                default_attn_proc, enabled=name.endswith("attn1.processor"), name=name
            )
        unet.set_attn_processor(unet_lora_attn_procs)

    def __getattr__(self, name: str):
        try:
            return super().__getattr__(name)
        except AttributeError:
            return getattr(self.unet, name)

    def forward_cond(self, noisy_cond_lat, timestep, encoder_hidden_states, class_labels, ref_dict, is_cfg_guidance, **kwargs):
        if is_cfg_guidance:
            encoder_hidden_states = encoder_hidden_states[1:]
            class_labels = class_labels[1:]
        self.unet(
            noisy_cond_lat, timestep,
            encoder_hidden_states=encoder_hidden_states,
            class_labels=class_labels,
            cross_attention_kwargs=dict(mode="w", ref_dict=ref_dict),
            **kwargs
        )

    def forward(
        self, sample, timestep, encoder_hidden_states, class_labels=None,
        *args, cross_attention_kwargs,
        down_block_res_samples=None, mid_block_res_sample=None,
        **kwargs
    ):
        cond_lat = cross_attention_kwargs['cond_lat']
        is_cfg_guidance = cross_attention_kwargs.get('is_cfg_guidance', False)
        noise = torch.randn_like(cond_lat)
        if self.training:
            noisy_cond_lat = self.train_sched.add_noise(cond_lat, noise, timestep)
            noisy_cond_lat = self.train_sched.scale_model_input(noisy_cond_lat, timestep)
        else:
            noisy_cond_lat = self.val_sched.add_noise(cond_lat, noise, timestep.reshape(-1))
            noisy_cond_lat = self.val_sched.scale_model_input(noisy_cond_lat, timestep.reshape(-1))
        ref_dict = {}
        self.forward_cond(
            noisy_cond_lat, timestep,
            encoder_hidden_states, class_labels,
            ref_dict, is_cfg_guidance, **kwargs
        )
        weight_dtype = self.unet.dtype
        return self.unet(
            sample, timestep,
            encoder_hidden_states, *args,
            class_labels=class_labels,
            cross_attention_kwargs=dict(mode="r", ref_dict=ref_dict, is_cfg_guidance=is_cfg_guidance),
            down_block_additional_residuals=[
                sample.to(dtype=weight_dtype) for sample in down_block_res_samples
            ] if down_block_res_samples is not None else None,
            mid_block_additional_residual=(
                mid_block_res_sample.to(dtype=weight_dtype)
                if mid_block_res_sample is not None else None
            ),
            **kwargs
        )


def scale_latents(latents):
    latents = (latents - 0.22) * 0.75
    return latents


def unscale_latents(latents):
    latents = latents / 0.75 + 0.22
    return latents


def scale_image(image):
    image = image * 0.5 / 0.8
    return image


def unscale_image(image):
    image = image / 0.5 * 0.8
    return image


class DepthControlUNet(torch.nn.Module):
    def __init__(self, unet: RefOnlyNoisedUNet) -> None:
        super().__init__()
        self.unet = unet
        self.controlnet = diffusers.ControlNetModel.from_unet(unet.unet)
        DefaultAttnProc = AttnProcessor2_0
        if is_xformers_available():
            DefaultAttnProc = XFormersAttnProcessor
        self.controlnet.set_attn_processor(DefaultAttnProc())

    def __getattr__(self, name: str):
        try:
            return super().__getattr__(name)
        except AttributeError:
            return getattr(self.unet, name)

    def forward(self, sample, timestep, encoder_hidden_states, class_labels=None, *args, cross_attention_kwargs: dict, **kwargs):
        cross_attention_kwargs = dict(cross_attention_kwargs)
        control_depth = cross_attention_kwargs.pop('control_depth')
        down_block_res_samples, mid_block_res_sample = self.controlnet(
            sample,
            timestep,
            encoder_hidden_states=encoder_hidden_states,
            controlnet_cond=control_depth,
            return_dict=False,
        )
        return self.unet(
            sample,
            timestep,
            encoder_hidden_states=encoder_hidden_states,
            down_block_res_samples=down_block_res_samples,
            mid_block_res_sample=mid_block_res_sample,
            cross_attention_kwargs=cross_attention_kwargs
        )


class ModuleListDict(torch.nn.Module):
    def __init__(self, procs: dict) -> None:
        super().__init__()
        self.keys = sorted(procs.keys())
        self.values = torch.nn.ModuleList(procs[k] for k in self.keys)

    def __getitem__(self, key):
        return self.values[self.keys.index(key)]


class SuperNet(torch.nn.Module):
    def __init__(self, state_dict: Dict[str, torch.Tensor]):
        super().__init__()
        state_dict = OrderedDict((k, state_dict[k]) for k in sorted(state_dict.keys()))
        self.layers = torch.nn.ModuleList(state_dict.values())
        self.mapping = dict(enumerate(state_dict.keys()))
        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}

        # .processor for unet, .self_attn for text encoder
        self.split_keys = [".processor", ".self_attn"]

        # we add a hook to state_dict() and load_state_dict() so that the
        # naming fits with `unet.attn_processors`
        def map_to(module, state_dict, *args, **kwargs):
            new_state_dict = {}
            for key, value in state_dict.items():
                num = int(key.split(".")[1])  # 0 is always "layers"
                new_key = key.replace(f"layers.{num}", module.mapping[num])
                new_state_dict[new_key] = value

            return new_state_dict

        def remap_key(key, state_dict):
            for k in self.split_keys:
                if k in key:
                    return key.split(k)[0] + k
            return key.split('.')[0]

        def map_from(module, state_dict, *args, **kwargs):
            all_keys = list(state_dict.keys())
            for key in all_keys:
                replace_key = remap_key(key, state_dict)
                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
                state_dict[new_key] = state_dict[key]
                del state_dict[key]

        self._register_state_dict_hook(map_to)
        self._register_load_state_dict_pre_hook(map_from, with_module=True)


class Zero123PlusPipeline(diffusers.StableDiffusionPipeline):
    tokenizer: transformers.CLIPTokenizer
    text_encoder: transformers.CLIPTextModel
    vision_encoder: transformers.CLIPVisionModelWithProjection

    feature_extractor_clip: transformers.CLIPImageProcessor
    unet: UNet2DConditionModel
    scheduler: diffusers.schedulers.KarrasDiffusionSchedulers

    vae: AutoencoderKL
    ramping: nn.Linear

    feature_extractor_vae: transformers.CLIPImageProcessor

    depth_transforms_multi = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.5], [0.5])
    ])

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        vision_encoder: transformers.CLIPVisionModelWithProjection,
        feature_extractor_clip: CLIPImageProcessor, 
        feature_extractor_vae: CLIPImageProcessor,
        ramping_coefficients: Optional[list] = None,
        safety_checker=None,
    ):
        DiffusionPipeline.__init__(self)

        self.register_modules(
            vae=vae, text_encoder=text_encoder, tokenizer=tokenizer,
            unet=unet, scheduler=scheduler, safety_checker=None,
            vision_encoder=vision_encoder,
            feature_extractor_clip=feature_extractor_clip,
            feature_extractor_vae=feature_extractor_vae
        )
        self.register_to_config(ramping_coefficients=ramping_coefficients)
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)

    def prepare(self):
        train_sched = DDPMScheduler.from_config(self.scheduler.config)
        if isinstance(self.unet, UNet2DConditionModel):
            self.unet = RefOnlyNoisedUNet(self.unet, train_sched, self.scheduler).eval()

    def add_controlnet(self):
        self.unet = DepthControlUNet(self.unet)
        return SuperNet(OrderedDict([('controlnet', self.unet.controlnet)]))

    def encode_condition_image(self, image: torch.Tensor):
        image = self.vae.encode(image).latent_dist.sample()
        return image

    @torch.no_grad()
    def __call__(
        self,
        image: Image.Image = None,
        prompt = "",
        *args,
        num_images_per_prompt: Optional[int] = 1,
        guidance_scale=4.0,
        depth_image: Image.Image = None,
        output_type: Optional[str] = "pil",
        width=640,
        height=960,
        num_inference_steps=28,
        return_dict=True,
        **kwargs
    ):
        self.prepare()
        if image is None:
            raise ValueError("Inputting embeddings not supported for this pipeline. Please pass an image.")
        assert not isinstance(image, torch.Tensor)
        image_1 = self.feature_extractor_vae(images=image, return_tensors="pt").pixel_values
        image_2 = self.feature_extractor_clip(images=image, return_tensors="pt").pixel_values
        if depth_image is not None and hasattr(self.unet, "controlnet"):
            depth_image = self.depth_transforms_multi(depth_image).to(
                device=self.unet.controlnet.device, dtype=self.unet.controlnet.dtype
            )
        image = image_1.to(device=self.vae.device, dtype=self.vae.dtype)
        image_2 = image_2.to(device=self.vae.device, dtype=self.vae.dtype)
        cond_lat = self.encode_condition_image(image)
        if guidance_scale > 1:
            negative_lat = self.encode_condition_image(torch.zeros_like(image))
            cond_lat = torch.cat([negative_lat, cond_lat])
        encoded = self.vision_encoder(image_2, output_hidden_states=False)
        global_embeds = encoded.image_embeds
        global_embeds = global_embeds.unsqueeze(-2)
        
        encoder_hidden_states = self._encode_prompt(
            prompt,
            self.device,
            num_images_per_prompt,
            False
        )
        ramp = global_embeds.new_tensor(self.config.ramping_coefficients).unsqueeze(-1)
        encoder_hidden_states = encoder_hidden_states + global_embeds * ramp
        cak = dict(cond_lat=cond_lat)
        if hasattr(self.unet, "controlnet"):
            cak['control_depth'] = depth_image
        latents: torch.Tensor = super().__call__(
            None,
            *args,
            cross_attention_kwargs=cak,
            guidance_scale=guidance_scale,
            num_images_per_prompt=num_images_per_prompt,
            prompt_embeds=encoder_hidden_states,
            num_inference_steps=num_inference_steps,
            output_type='latent',
            width=width,
            height=height,
            **kwargs
        ).images
        latents = unscale_latents(latents)
        if not output_type == "latent":
            image = unscale_image(self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0])
        else:
            image = latents

        image = self.image_processor.postprocess(image, output_type=output_type)
        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)