model/video_diffusion/pipelines/pipeline_st_stable_diffusion.py

# Copyright 2023 Bytedance Ltd. and/or its affiliates

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

#     http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
from typing import Callable, List, Optional, Union

import torch
from einops import rearrange

from diffusers.utils import is_accelerate_available
from packaging import version
from transformers import CLIPTextModel, CLIPTokenizer

from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.schedulers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from diffusers.utils import deprecate, logging
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput

from ..models.unet_3d_condition import UNetPseudo3DConditionModel
import os, importlib

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


class SpatioTemporalStableDiffusionPipeline(DiffusionPipeline):
    r"""
    Pipeline for text-to-video generation using Spatio-Temporal Stable Diffusion.
    改变了unet的输入, unet换成3d unet, 其他部分完全和原来2D的一致。
    latents的变为 b,c,f,h,w 原来是 b,c,h,w。
    要用VAE的decoder的时候, 把输入reshape 成 (b f) c h w
    """
    _optional_components = []

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNetPseudo3DConditionModel,
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
    ):
        super().__init__()

        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
                " file"
            )
            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
            )
            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["clip_sample"] = False
            scheduler._internal_dict = FrozenDict(new_config)

        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
            version.parse(unet.config._diffusers_version).base_version
        ) < version.parse("0.9.0.dev0")
        is_unet_sample_size_less_64 = (
            hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
        )
        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
            deprecation_message = (
                "The configuration file of the unet has set the default `sample_size` to smaller than"
                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
                " in the config might lead to incorrect results in future versions. If you have downloaded this"
                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
                " the `unet/config.json` file"
            )
            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(unet.config)
            new_config["sample_size"] = 64
            unet._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)

    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding.

        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
        steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_sequential_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
        """
        if is_accelerate_available():
            from accelerate import cpu_offload
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

        device = torch.device(f"cuda:{gpu_id}")

        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
            if cpu_offloaded_model is not None:
                cpu_offload(cpu_offloaded_model, device)

    @property
    def _execution_device(self):
        r"""
        Returns the device on which the pipeline's models will be executed. After calling
        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
        hooks.
        """
        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    def _encode_prompt(
        self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `list(int)`):
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
        """
        batch_size = len(prompt) if isinstance(prompt, list) else 1

        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
            text_input_ids, untruncated_ids
        ):
            removed_text = self.tokenizer.batch_decode(
                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
            )
            logger.warning(
                "The following part of your input was truncated because CLIP can only handle sequences up to"
                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
            )

        if (
            hasattr(self.text_encoder.config, "use_attention_mask")
            and self.text_encoder.config.use_attention_mask
        ):
            attention_mask = text_inputs.attention_mask.to(device)
        else:
            attention_mask = None

        text_embeddings = self.text_encoder(
            text_input_ids.to(self.text_encoder.device),    # FIXME 强制对齐device的位置
            attention_mask=attention_mask,
        )
        text_embeddings = text_embeddings[0]

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        bs_embed, seq_len, _ = text_embeddings.shape
        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            max_length = text_input_ids.shape[-1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            if (
                hasattr(self.text_encoder.config, "use_attention_mask")
                and self.text_encoder.config.use_attention_mask
            ):
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            uncond_embeddings = self.text_encoder(
                uncond_input.input_ids.to(self.text_encoder.device), # 同上，强制位置对齐。
                attention_mask=attention_mask,
            )
            uncond_embeddings = uncond_embeddings[0]

            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = uncond_embeddings.shape[1]
            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)

            # For classifier free guidance, we need to do two forward passes.
            # Here we concatenate the unconditional and text embeddings into a single batch
            # to avoid doing two forward passes
            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

        return text_embeddings

    def decode_latents(self, latents):
        b = latents.shape[0]
        latents = 1 / 0.18215 * latents
        
        is_video = len(latents.shape) == 5
        if is_video:
            latents = rearrange(latents, "b c f h w -> (b f) c h w")

        image = self.vae.decode(latents).sample
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16

        image = image.cpu().float().numpy()
        if is_video:
            image = rearrange(image, "(b f) c h w -> b f h w c", b=b)
        else:
            image = rearrange(image, "b c h w -> b h w c")
        return image

    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def check_inputs(self, prompt, height, width, callback_steps):
        if not isinstance(prompt, str) and not isinstance(prompt, list):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(
                f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
            )

        if (callback_steps is None) or (
            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
        ):
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        clip_length,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):  
        if clip_length>0:
            shape = (
                batch_size,
                num_channels_latents,
                clip_length,
                height // self.vae_scale_factor,
                width // self.vae_scale_factor,
            )
        else:
            shape = (
                batch_size,
                num_channels_latents,
                height // self.vae_scale_factor,
                width // self.vae_scale_factor,
            )

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            rand_device = "cpu" if device.type == "mps" else device

            if isinstance(generator, list):
                shape = (1,) + shape[1:]
                latents = [
                    torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)
                    for i in range(batch_size)
                ]
                latents = torch.cat(latents, dim=0).to(device)
            else:
                latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype).to(
                    device
                )
        else:
            if latents.shape != shape:
                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        height: Optional[int] = None,
        width: Optional[int] = None,
        fps_labels = None,
        num_inference_steps: int = 50,
        clip_length: int = 8,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
            When returning a tuple, the first element is a list with the generated images, and the second element is a
            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
            (nsfw) content, according to the `safety_checker`.
        """
        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(prompt, height, width, callback_steps)

        # 2. Define call parameters
        batch_size = 1 if isinstance(prompt, str) else len(prompt)
        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        # 3. Encode input prompt
        text_embeddings = self._encode_prompt(
            prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
        )

        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        num_channels_latents = self.unet.in_channels
        
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            clip_length,
            height,
            width,
            text_embeddings.dtype,
            device,
            generator,
            latents,
        )
        latents_dtype = latents.dtype

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                # print(latent_model_input.shape, )
                # predict the noise residual
                if fps_labels:
                    if isinstance(fps_labels, list):
                        fps_labels = torch.tensor(fps_labels).to(self.unet.device)
                    # 控制帧率
                    noise_pred = self.unet(
                        latent_model_input, t, encoder_hidden_states=text_embeddings, fps_labels=fps_labels,
                    ).sample.to(dtype=latents_dtype)
                else:
                    noise_pred = self.unet(
                        latent_model_input, t, encoder_hidden_states=text_embeddings
                    ).sample.to(dtype=latents_dtype)

                # perform guidance
                if do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + guidance_scale * (
                        noise_pred_text - noise_pred_uncond
                    )

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample

                # call the callback, if provided
                if i == len(timesteps) - 1 or (
                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
                ):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        callback(i, t, latents)

        # 8. Post-processing
        image = self.decode_latents(latents)
        # image[:, 1:, :, :, :] = image[:, 1:, :, :, :] + image[:, 0:1, :, :, :]    # 叠加残差

        # 9. Run safety checker
        has_nsfw_concept = None

        # 10. Convert to PIL
        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    @staticmethod
    def numpy_to_pil(images):
        if len(images.shape)==5:
            pil_images = []
            for sequence in images:
                pil_images.append(DiffusionPipeline.numpy_to_pil(sequence))
            return pil_images
        else:
            return DiffusionPipeline.numpy_to_pil(images)
    

    # 改写一下 model_index.json的保存内容, Unet是新定义的，直接保存会导致读取的时候出错～
    def to_json_string(self) -> str:
        from diffusers import __version__
        import json
        import numpy as np 

        config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {}
        config_dict["_class_name"] = self.__class__.__name__
        config_dict["_diffusers_version"] = __version__

        def to_json_saveable(value):
            if isinstance(value, np.ndarray):
                value = value.tolist()
            return value

        config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}
        if 'unet' in config_dict:
            config_dict["unet"] = [
                                    "diffusers",
                                    "UNet2DConditionModel"
                                ]
        if 'controlnet' in config_dict:
            config_dict['controlnet'] = [
                                "diffusers",
                                "UNet2DConditionModel"
                            ]
        # 覆盖
        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"