pipeline_reconstruction.py

# Copyright 2023 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.


from typing import List, Optional, Tuple, Union

import torch
import numpy as np

from diffusers import DDPMScheduler, DDIMScheduler
from diffusers.utils import randn_tensor
from diffusers.pipeline_utils import DiffusionPipeline, ImagePipelineOutput


class ReconstructionPipeline(DiffusionPipeline):
    r"""
    Pipeline for image generation.

    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
    implemented for all pipelines (downloading, saving, running on a particular device, etc.).

    Parameters:
        unet ([`UNet2DModel`]):
            A `UNet2DModel` to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image. Can be one of
            [`DDPMScheduler`], or [`DDIMScheduler`].
    """

    def __init__(self, unet, scheduler: Union[DDPMScheduler, DDIMScheduler]):
        super().__init__()
        self.register_modules(unet=unet, scheduler=scheduler)

    @torch.no_grad()
    def __call__(
        self,
        original_images:torch.Tensor,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        start_at_timestep: int = 200,
        num_inference_steps: int = 200,
        output_type: Optional[str] = "pil",
        return_dict: bool = True
    ) -> Union[ImagePipelineOutput, Tuple]:
        r"""
        The call function to the pipeline for generation.

        Args:
            original_images:
                Images to add noise to and then reconstruct.
            generator (`torch.Generator`, *optional*):
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
            start_at_timestep (`int`, *optional*, defaults to 200):
                Determine to which degree/timestep the image should be noised and then denoised.,
            num_inference_steps (`int`, *optional*, defaults to 800):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference. start_at_timestep + num_inferenc_steps must be smaller than num_trainsteps
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.

        Example:

        ```py
        >>> from diffusers import DDPMPipeline

        >>> # load model and scheduler
        >>> pipe = ReconstructionPipeline.from_pretrained("google/ddpm-cat-256")

        >>> # run pipeline in inference (sample random noise and denoise)
        >>> image = pipe().images[0]

        >>> # save image
        >>> image.save("ddpm_generated_image.png")
        ```

        Returns:
            [`~pipelines.ImagePipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
                returned where the first element is a list with the generated images
        """
        # Sample gaussian noise to begin loop
        image_shape = original_images.shape
        if self.device.type == "mps":
            # randn does not work reproducibly on mps
            image = randn_tensor(image_shape, generator=generator)
            image = image.to(self.device)
            raise Exception("not correctly implemented for mps yet")
        else:
            noise = randn_tensor(image_shape, generator=generator, device=self.device)
            starting_step = torch.ones((image_shape[0],), dtype=torch.int64, device=self.device) * start_at_timestep
            image = self.scheduler.add_noise(original_images.to(self.device), noise, starting_step)

        # set step values
        # step_ratio = self.config.num_train_timesteps // self.num_inference_steps
        # creates integer timesteps by multiplying by ratio
        # casting to int to avoid issues when num_inference_step is power of 3
        timesteps = (np.arange(0, start_at_timestep)).round()[::-1].copy().astype(np.int64)
        self.scheduler.set_timesteps(timesteps=timesteps)

        for t in self.progress_bar(self.scheduler.timesteps):
            # 1. predict noise model_output
            model_output = self.unet(image, t).sample

            # 2. compute previous image: x_t -> x_t-1
            image = self.scheduler.step(model_output, t, image, original_images, generator=generator).prev_sample

        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).numpy()
        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)