pipeline_stable_diffusion.py

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

import torch

from diffusers.utils import is_accelerate_available
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,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from diffusers.utils import deprecate, logging
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput

from modules.clip import FrozenCLIPEmbedderWithCustomWords
from models import UNet2DConditionModel, Hypernetwork
import models.shared as shared


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


class StableDiffusionPipeline(DiffusionPipeline):
    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[
            DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler, EulerDiscreteScheduler, EulerAncestralDiscreteScheduler
        ],
    ):
        super().__init__()

        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
        )

    def enable_xformers_memory_efficient_attention(self):
        r"""
        Enable memory efficient attention as implemented in xformers.
        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
        time. Speed up at training time is not guaranteed.
        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
        is used.
        """
        self.unet.set_use_memory_efficient_attention_xformers(True)

    def disable_xformers_memory_efficient_attention(self):
        r"""
        Disable memory efficient attention as implemented in xformers.
        """
        self.unet.set_use_memory_efficient_attention_xformers(False)

    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
        r"""
        Enable sliced attention computation.
        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
        in several steps. This is useful to save some memory in exchange for a small speed decrease.
        Args:
            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
                `attention_head_dim` must be a multiple of `slice_size`.
        """
        if slice_size == "auto":
            slice_size = self.unet.config.attention_head_dim // 2
        self.unet.set_attention_slice(slice_size)

    def disable_attention_slicing(self):
        r"""
        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
        back to computing attention in one step.
        """
        self.enable_attention_slicing(None)

    def enable_sequential_cpu_offload(self):
        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("cuda")

        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
            cpu_offload(cpu_offloaded_model, device)
            
    def load_hypernet(self, filename, hypernet_name="no_name"):
        enable_sizes = [["mid", s] for s in [768, 320, 640, 1280]]
        enable_sizes += [["down", s] for s in [768, 320, 640, 1280]]
        enable_sizes += [["up", s] for s in [768, 320, 640, 1280]]
        
        shared.hypernetworks = Hypernetwork(hypernet_name, enable_sizes)
        shared.hypernetworks.to(self.device)
        shared.hypernetworks.load(filename)

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        height: int = 512,
        width: int = 512,
        num_inference_steps: int = 50,
        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[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,
        use_custom_encoder=True,
        CLIP_stop_at_last_layers=1,
        **kwargs,
    ):
        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 512):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to 512):
                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*):
                A [torch generator](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.
            use_custom_encoder (`bool`, *optional*, defaults to True):
                Use FrozenCLIPEmbedderWithCustomWords or original diffusers embedder
            CLIP_stop_at_last_layers (`int`, *optional*, defaults to 1):
                Select last layer of CLIP model (Valid only use_custom_encoder=True)
        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`.
        """
        if use_custom_encoder:
            self.embedder = FrozenCLIPEmbedderWithCustomWords(self.tokenizer, self.text_encoder, CLIP_stop_at_last_layers=CLIP_stop_at_last_layers)

        if isinstance(prompt, str):
            batch_size = 1
        elif isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            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)}."
            )

        if not use_custom_encoder:
            # get prompt text embeddings
            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids

            if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
                removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
                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}"
                )
                text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
            text_embeddings = self.text_encoder(text_input_ids.to(self.device))[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)
        else:
            if batch_size == 1:
                prompt_str = [prompt]
            text_embeddings = self.embedder(prompt_str)

        # 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
        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""]
            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
            if not use_custom_encoder:
                uncond_input = self.tokenizer(
                    uncond_tokens,
                    padding="max_length",
                    max_length=self.tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                )
                uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
            else:
                uncond_embeddings = self.embedder(uncond_tokens)

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

        # get the initial random noise unless the user supplied it

        # Unlike in other pipelines, latents need to be generated in the target device
        # for 1-to-1 results reproducibility with the CompVis implementation.
        # However this currently doesn't work in `mps`.
        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
        latents_dtype = text_embeddings.dtype
        if latents is None:
            if self.device.type == "mps":
                # randn does not work reproducibly on mps
                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
                    self.device
                )
            else:
                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
        else:
            if latents.shape != latents_shape:
                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
            latents = latents.to(self.device)

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

        # Some schedulers like PNDM have timesteps as arrays
        # It's more optimized to move all timesteps to correct device beforehand
        timesteps_tensor = self.scheduler.timesteps.to(self.device)

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

        # 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

        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
            if do_classifier_free_guidance:
                # expand the latents if we are doing classifier free guidance
                latent_model_input = latents # torch.cat([latents] * 2)
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                # predict the noise residual
                noise_pred_text = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
                noise_pred_uncond = self.unet(latent_model_input, t, encoder_hidden_states=uncond_embeddings).sample
                # perform guidance
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
            else:
                # expand the latents if we are doing classifier free guidance
                latent_model_input = latents
                # predict the noise residual
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                # perform guidance
                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample

            # 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 callback is not None and i % callback_steps == 0:
                callback(i, t, latents)

        latents = 1 / 0.18215 * latents
        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().permute(0, 2, 3, 1).float().numpy()

        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image, None)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)