diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py

# 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
from typing import Callable, List, Optional, Union

import numpy as np
import torch

import PIL
from transformers import CLIPFeatureExtractor, CLIPTokenizer

from ...configuration_utils import FrozenDict
from ...onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
from ...pipeline_utils import DiffusionPipeline
from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
from ...utils import PIL_INTERPOLATION, deprecate, logging
from . import StableDiffusionPipelineOutput


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


NUM_UNET_INPUT_CHANNELS = 9
NUM_LATENT_CHANNELS = 4


def prepare_mask_and_masked_image(image, mask, latents_shape):
    image = np.array(image.convert("RGB").resize((latents_shape[1] * 8, latents_shape[0] * 8)))
    image = image[None].transpose(0, 3, 1, 2)
    image = image.astype(np.float32) / 127.5 - 1.0

    image_mask = np.array(mask.convert("L").resize((latents_shape[1] * 8, latents_shape[0] * 8)))
    masked_image = image * (image_mask < 127.5)

    mask = mask.resize((latents_shape[1], latents_shape[0]), PIL_INTERPOLATION["nearest"])
    mask = np.array(mask.convert("L"))
    mask = mask.astype(np.float32) / 255.0
    mask = mask[None, None]
    mask[mask < 0.5] = 0
    mask[mask >= 0.5] = 1

    return mask, masked_image


class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
    r"""
    Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*.

    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, running on a particular device, etc.)

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            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 ([`UNet2DConditionModel`]): 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`], or [`PNDMScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            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`.
    """
    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__()
        logger.info("`OnnxStableDiffusionInpaintPipeline` is experimental and will very likely change in the future.")

        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.register_to_config(requires_safety_checker=requires_safety_checker)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
    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="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}"
            )

        text_embeddings = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[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.int32))[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

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        image: PIL.Image.Image,
        mask_image: PIL.Image.Image,
        height: Optional[int] = 512,
        width: Optional[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[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.
            image (`PIL.Image.Image`):
                `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
                be masked out with `mask_image` and repainted according to `prompt`.
            mask_image (`PIL.Image.Image`):
                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
                to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
                instead of 3, so the expected shape would be `(B, H, W, 1)`.
            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 (`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`.
        """

        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

        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

        # 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

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

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

        # prepare mask and masked_image
        mask, masked_image = prepare_mask_and_masked_image(image, mask_image, latents_shape[-2:])
        mask = mask.astype(latents.dtype)
        masked_image = masked_image.astype(latents.dtype)

        masked_image_latents = self.vae_encoder(sample=masked_image)[0]
        masked_image_latents = 0.18215 * masked_image_latents

        # duplicate mask and masked_image_latents for each generation per prompt
        mask = mask.repeat(batch_size * num_images_per_prompt, 0)
        masked_image_latents = masked_image_latents.repeat(batch_size * num_images_per_prompt, 0)

        mask = np.concatenate([mask] * 2) if do_classifier_free_guidance else mask
        masked_image_latents = (
            np.concatenate([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
        )

        num_channels_mask = mask.shape[1]
        num_channels_masked_image = masked_image_latents.shape[1]

        unet_input_channels = NUM_UNET_INPUT_CHANNELS
        if num_channels_latents + num_channels_mask + num_channels_masked_image != unet_input_channels:
            raise ValueError(
                "Incorrect configuration settings! The config of `pipeline.unet` expects"
                f" {unet_input_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
                " `pipeline.unet` or your `mask_image` or `image` input."
            )

        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * np.float(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 = 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
            # concat latents, mask, masked_image_latnets in the channel dimension
            latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
            latent_model_input = latent_model_input.cpu().numpy()
            latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1)

            # predict the noise residual
            timestep = np.array([t], dtype=timestep_dtype)
            noise_pred = self.unet(
                sample=latent_model_input, timestep=timestep, encoder_hidden_states=text_embeddings
            )[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
        # 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])[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)
            # safety_checker does not support batched inputs yet
            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)