pipeline_openvino_stable_diffusion.py

# Copyright 2022 The OFA-Sys Team.
# This source code is licensed under the Apache 2.0 license
# found in the LICENSE file in the root directory.
# Copyright 2022 The HuggingFace Inc. team.
# All rights reserved.
# This source code is licensed under the Apache 2.0 license
# found in the LICENSE file in the root directory.

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

import numpy as np
import torch
import os

from transformers import CLIPFeatureExtractor, CLIPTokenizer

from diffusers.configuration_utils import FrozenDict
from diffusers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
from diffusers.utils import deprecate, logging
from diffusers.onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel

from diffusers import OnnxStableDiffusionPipeline, DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from openvino.runtime import Core

logger = logging.get_logger(__name__)


class OpenVINOStableDiffusionPipeline(DiffusionPipeline):
    vae_encoder: OnnxRuntimeModel
    vae_decoder: OnnxRuntimeModel
    text_encoder: OnnxRuntimeModel
    tokenizer: CLIPTokenizer
    unet: OnnxRuntimeModel
    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
    safety_checker: OnnxRuntimeModel
    feature_extractor: CLIPFeatureExtractor

    _optional_components = ["safety_checker", "feature_extractor"]

    def __init__(
        self,
        vae_encoder: OnnxRuntimeModel,
        vae_decoder: OnnxRuntimeModel,
        text_encoder: OnnxRuntimeModel,
        tokenizer: CLIPTokenizer,
        unet: OnnxRuntimeModel,
        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
        safety_checker: OnnxRuntimeModel,
        feature_extractor: CLIPFeatureExtractor,
        requires_safety_checker: bool = True,
    ):
        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                f"The configuration file of this scheduler: {scheduler} is outdated.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)

        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. Both the 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.convert_to_openvino()
        self.register_to_config(
            requires_safety_checker=requires_safety_checker)

    @classmethod
    def from_onnx_pipeline(cls, onnx_pipe: OnnxStableDiffusionPipeline):
        r"""
        Create OpenVINOStableDiffusionPipeline from a onnx stable pipeline.
        Parameters:
            onnx_pipe (OnnxStableDiffusionPipeline)
        """
        return cls(onnx_pipe.vae_encoder, onnx_pipe.vae_decoder,
                   onnx_pipe.text_encoder, onnx_pipe.tokenizer, onnx_pipe.unet,
                   onnx_pipe.scheduler, onnx_pipe.safety_checker,
                   onnx_pipe.feature_extractor, True)

    def convert_to_openvino(self):
        ie = Core()

        # VAE decoder
        vae_decoder_onnx = ie.read_model(
            model=os.path.join(self.vae_decoder.model_save_dir, "model.onnx"))
        vae_decoder = ie.compile_model(model=vae_decoder_onnx,
                                       device_name="CPU")

        # Text encoder
        text_encoder_onnx = ie.read_model(
            model=os.path.join(self.text_encoder.model_save_dir, "model.onnx"))
        text_encoder = ie.compile_model(model=text_encoder_onnx,
                                        device_name="CPU")

        # Unet
        unet_onnx = ie.read_model(
            model=os.path.join(self.unet.model_save_dir, "model.onnx"))
        unet = ie.compile_model(model=unet_onnx, device_name="CPU")

        self.register_modules(vae_decoder=vae_decoder,
                              text_encoder=text_encoder,
                              unet=unet)

    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[str]`):
                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="max_length",
                                         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}")

        prompt_embeds = self.text_encoder(
            {"input_ids":
             text_input_ids.astype(np.int32)})[self.text_encoder.outputs[0]]
        prompt_embeds = np.repeat(prompt_embeds, 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",
            )
            negative_prompt_embeds = self.text_encoder({
                "input_ids":
                uncond_input.input_ids.astype(np.int32)
            })[self.text_encoder.outputs[0]]
            negative_prompt_embeds = np.repeat(negative_prompt_embeds,
                                               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
            prompt_embeds = np.concatenate(
                [negative_prompt_embeds, prompt_embeds])

        return prompt_embeds

    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,
    ):
        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 generator is None:
            generator = np.random

        # 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

        prompt_embeds = self._encode_prompt(prompt, num_images_per_prompt,
                                            do_classifier_free_guidance,
                                            negative_prompt)

        # get the initial random noise unless the user supplied it
        latents_dtype = prompt_embeds.dtype
        latents_shape = (batch_size * num_images_per_prompt, 4, height // 8,
                         width // 8)
        if latents is None:
            latents = generator.randn(*latents_shape).astype(latents_dtype)
        elif latents.shape != latents_shape:
            raise ValueError(
                f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}"
            )

        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

        latents = latents * np.float64(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

        # timestep_dtype = next(
        #     (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
        # )
        timestep_dtype = 'tensor(int64)'
        timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]

        for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = np.concatenate(
                [latents] * 2) if do_classifier_free_guidance else latents
            latent_model_input = self.scheduler.scale_model_input(
                torch.from_numpy(latent_model_input), t)
            latent_model_input = latent_model_input.cpu().numpy()

            # predict the noise residual
            timestep = np.array([t], dtype=timestep_dtype)
            unet_input = {
                "sample": latent_model_input,
                "timestep": timestep,
                "encoder_hidden_states": prompt_embeds
            }
            noise_pred = self.unet(unet_input)[self.unet.outputs[0]]
            # noise_pred = noise_pred[0]

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

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

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

        latents = 1 / 0.18215 * latents
        image2 = self.vae_decoder({"latent_sample":
                                   latents})[self.vae_decoder.outputs[0]]
        # image = self.vae_decoder(latent_sample=latents)[0]
        # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
        image = np.concatenate([
            self.vae_decoder({"latent_sample":
                              latents[i:i + 1]})[self.vae_decoder.outputs[0]]
            for i in range(latents.shape[0])
        ])

        image = np.clip(image / 2 + 0.5, 0, 1)
        image = image.transpose((0, 2, 3, 1))

        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(
                self.numpy_to_pil(image),
                return_tensors="np").pixel_values.astype(image.dtype)

            image, has_nsfw_concepts = self.safety_checker(
                clip_input=safety_checker_input, images=image)

            # 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

        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)