ppdiffusers/pipelines/stable_diffusion/pipeline_fastdeploy_stable_diffusion.py

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# 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
import time
from typing import Callable, List, Optional, Union

import numpy as np
import paddle

from paddlenlp.transformers import CLIPFeatureExtractor, CLIPTokenizer

from ...fastdeploy_utils import FastDeployRuntimeModel
from ...pipeline_utils import DiffusionPipeline
from ...schedulers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from ...schedulers.preconfig import (
    PreconfigEulerAncestralDiscreteScheduler,
    PreconfigLMSDiscreteScheduler,
)
from ...utils import logging
from . import StableDiffusionPipelineOutput

logger = logging.get_logger(__name__)


class FastDeployStableDiffusionPipeline(DiffusionPipeline):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion.

    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
    library implements for all the pipelines (such as downloading or saving etc.)

    Args:
        vae_encoder ([`FastDeployRuntimeModel`]):
            Variational Auto-Encoder (VAE) Model to encode images to latent representations.
        vae_decoder ([`FastDeployRuntimeModel`]):
            Variational Auto-Encoder (VAE) Model to decode images from latent representations.
        text_encoder ([`FastDeployRuntimeModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`FastDeployRuntimeModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`PNDMScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`]
            or [`DPMSolverMultistepScheduler`].
        safety_checker ([`FastDeployRuntimeModel`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
        feature_extractor ([`CLIPFeatureExtractor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """
    _optional_components = ["vae_encoder", "safety_checker", "feature_extractor"]

    def __init__(
        self,
        vae_encoder: FastDeployRuntimeModel,
        vae_decoder: FastDeployRuntimeModel,
        text_encoder: FastDeployRuntimeModel,
        tokenizer: CLIPTokenizer,
        unet: FastDeployRuntimeModel,
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            PreconfigLMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            PreconfigEulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
        safety_checker: FastDeployRuntimeModel,
        feature_extractor: CLIPFeatureExtractor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()
        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. PaddleNLP team, diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )
        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        self.register_modules(
            vae_encoder=vae_encoder,
            vae_decoder=vae_decoder,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    def _encode_prompt(self, prompt, 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
            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

        # get prompt text embeddings
        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="np",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="np").input_ids

        if not np.array_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}"
            )

        text_embeddings = self.text_encoder(input_ids=text_input_ids.astype(np.int64))[0]
        text_embeddings = np.repeat(text_embeddings, num_images_per_prompt, axis=0)
        # 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] * batch_size
            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="np",
            )
            uncond_embeddings = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int64))[0]
            uncond_embeddings = np.repeat(uncond_embeddings, num_images_per_prompt, axis=0)

            # 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 = np.concatenate([uncond_embeddings, text_embeddings])

        return text_embeddings

    def run_safety_checker(self, image, dtype):
        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(
                self.numpy_to_pil(image), return_tensors="np"
            ).pixel_values.astype(dtype)
            # There will throw an error if use safety_checker batchsize>1
            images, has_nsfw_concept = [], []
            for i in range(image.shape[0]):
                image_i, has_nsfw_concept_i = self.safety_checker(
                    clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
                )
                images.append(image_i)
                has_nsfw_concept.append(has_nsfw_concept_i[0])
            image = np.concatenate(images)
        else:
            has_nsfw_concept = None
        return image, has_nsfw_concept

    def decode_latents(self, latents):
        latents = 1 / 0.18215 * latents
        latents_shape = latents.shape
        vae_output_shape = [latents_shape[0], 3, latents_shape[2] * 8, latents_shape[3] * 8]
        images_vae = paddle.zeros(vae_output_shape, dtype="float32")

        vae_input_name = self.vae_decoder.model.get_input_info(0).name
        vae_output_name = self.vae_decoder.model.get_output_info(0).name

        self.vae_decoder.zero_copy_infer(
            prebinded_inputs={vae_input_name: latents},
            prebinded_outputs={vae_output_name: images_vae},
            share_with_raw_ptr=True,
        )

        images_vae = paddle.clip(images_vae / 2 + 0.5, 0, 1)
        images = images_vae.transpose([0, 2, 3, 1])
        return images.numpy()

    def prepare_extra_step_kwargs(self, 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
        return extra_step_kwargs

    def check_var_kwargs_of_scheduler_func(self, scheduler_func):
        sig = inspect.signature(scheduler_func)
        params = sig.parameters.values()
        has_kwargs = any([True for p in params if p.kind == p.VAR_KEYWORD])
        return has_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, height, width, dtype, generator, latents=None):
        if generator is None:
            generator = np.random

        latents_shape = (batch_size, num_channels_latents, height // 8, width // 8)
        if latents is None:
            latents = generator.randn(*latents_shape).astype(dtype)
        elif latents.shape != latents_shape:
            raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")

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

    def __call__(
        self,
        prompt: Union[str, List[str]],
        height: Optional[int] = 512,
        width: Optional[int] = 512,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: Optional[float] = 0.0,
        generator: Optional[np.random.RandomState] = None,
        latents: Optional[np.ndarray] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, np.ndarray], 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*, 512):
                The height in pixels of the generated image.
            width (`int`, *optional*, 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 (`np.random.RandomState`, *optional*):
                A np.random.RandomState to make generation deterministic.
            latents (`np.ndarray`, *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: np.ndarray)`.
            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`.
        """
        # 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)

        # 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
        start_time_encode_prompt = time.perf_counter()
        text_embeddings = self._encode_prompt(
            prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
        )
        print("_encode_prompt latency:", time.perf_counter() - start_time_encode_prompt)
        # 4. Prepare timesteps
        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        num_channels_latents = 4
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            text_embeddings.dtype,
            generator,
            latents,
        )
        if isinstance(latents, np.ndarray):
            latents = paddle.to_tensor(latents)
        # 6. Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(eta)
        # 7. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        scheduler_support_kwagrs_scale_input = self.check_var_kwargs_of_scheduler_func(
            self.scheduler.scale_model_input
        )
        scheduler_support_kwagrs_step = self.check_var_kwargs_of_scheduler_func(self.scheduler.step)

        unet_output_name = self.unet.model.get_output_info(0).name
        unet_input_names = [self.unet.model.get_input_info(i).name for i in range(self.unet.model.num_inputs())]
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            text_embeddings = paddle.to_tensor(text_embeddings, dtype="float32")
            for i, t in enumerate(timesteps):
                noise_pred_unet = paddle.zeros(
                    [2 * batch_size * num_images_per_prompt, 4, height // 8, width // 8], dtype="float32"
                )
                # expand the latents if we are doing classifier free guidance
                latent_model_input = paddle.concat([latents] * 2) if do_classifier_free_guidance else latents
                if scheduler_support_kwagrs_scale_input:
                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t, step_index=i)
                else:
                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                # predict the noise residual
                self.unet.zero_copy_infer(
                    prebinded_inputs={
                        unet_input_names[0]: latent_model_input,
                        unet_input_names[1]: t,
                        unet_input_names[2]: text_embeddings,
                    },
                    prebinded_outputs={unet_output_name: noise_pred_unet},
                    share_with_raw_ptr=True,
                )
                # perform guidance
                if do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred_unet.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
                if scheduler_support_kwagrs_step:
                    scheduler_output = self.scheduler.step(
                        noise_pred, t, latents, step_index=i, return_pred_original_sample=False, **extra_step_kwargs
                    )
                else:
                    scheduler_output = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs)
                latents = scheduler_output.prev_sample
                if i == num_inference_steps - 1:
                    # sync for accuracy it/s measure
                    paddle.device.cuda.synchronize()
                # call the callback, if provided
                if i == num_inference_steps - 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
        time_start_decoder = time.perf_counter()
        image = self.decode_latents(latents)
        print("decoder latency:", time.perf_counter() - time_start_decoder)
        # 9. Run safety checker
        image, has_nsfw_concept = self.run_safety_checker(image, text_embeddings.dtype)

        # 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)