diffusers/pipelines/kandinsky3/kandinsky3img2img_pipeline.py

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

import numpy as np
import PIL
import PIL.Image
import torch
from transformers import T5EncoderModel, T5Tokenizer

from ...loaders import LoraLoaderMixin
from ...models import Kandinsky3UNet, VQModel
from ...schedulers import DDPMScheduler
from ...utils import (
    is_accelerate_available,
    logging,
)
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput


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


def downscale_height_and_width(height, width, scale_factor=8):
    new_height = height // scale_factor**2
    if height % scale_factor**2 != 0:
        new_height += 1
    new_width = width // scale_factor**2
    if width % scale_factor**2 != 0:
        new_width += 1
    return new_height * scale_factor, new_width * scale_factor


def prepare_image(pil_image):
    arr = np.array(pil_image.convert("RGB"))
    arr = arr.astype(np.float32) / 127.5 - 1
    arr = np.transpose(arr, [2, 0, 1])
    image = torch.from_numpy(arr).unsqueeze(0)
    return image


class Kandinsky3Img2ImgPipeline(DiffusionPipeline, LoraLoaderMixin):
    model_cpu_offload_seq = "text_encoder->unet->movq"

    def __init__(
        self,
        tokenizer: T5Tokenizer,
        text_encoder: T5EncoderModel,
        unet: Kandinsky3UNet,
        scheduler: DDPMScheduler,
        movq: VQModel,
    ):
        super().__init__()

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

    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)

        t_start = max(num_inference_steps - init_timestep, 0)
        timesteps = self.scheduler.timesteps[t_start:]

        return timesteps, num_inference_steps - t_start

    def remove_all_hooks(self):
        if is_accelerate_available():
            from accelerate.hooks import remove_hook_from_module
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

        for model in [self.text_encoder, self.unet]:
            if model is not None:
                remove_hook_from_module(model, recurse=True)

        self.unet_offload_hook = None
        self.text_encoder_offload_hook = None
        self.final_offload_hook = None

    def _process_embeds(self, embeddings, attention_mask, cut_context):
        # return embeddings, attention_mask
        if cut_context:
            embeddings[attention_mask == 0] = torch.zeros_like(embeddings[attention_mask == 0])
            max_seq_length = attention_mask.sum(-1).max() + 1
            embeddings = embeddings[:, :max_seq_length]
            attention_mask = attention_mask[:, :max_seq_length]
        return embeddings, attention_mask

    @torch.no_grad()
    def encode_prompt(
        self,
        prompt,
        do_classifier_free_guidance=True,
        num_images_per_prompt=1,
        device=None,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        _cut_context=False,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
             prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`torch.device`, *optional*):
                torch device to place the resulting embeddings on
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
        """
        if prompt is not None and negative_prompt is not None:
            if 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)}."
                )

        if device is None:
            device = self._execution_device

        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        max_length = 128

        if prompt_embeds is None:
            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids.to(device)
            attention_mask = text_inputs.attention_mask.to(device)
            prompt_embeds = self.text_encoder(
                text_input_ids,
                attention_mask=attention_mask,
            )
            prompt_embeds = prompt_embeds[0]
            prompt_embeds, attention_mask = self._process_embeds(prompt_embeds, attention_mask, _cut_context)
            prompt_embeds = prompt_embeds * attention_mask.unsqueeze(2)

        if self.text_encoder is not None:
            dtype = self.text_encoder.dtype
        else:
            dtype = None

        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
        attention_mask = attention_mask.repeat(num_images_per_prompt, 1)
        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]

            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            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 negative_prompt is not None:
                uncond_input = self.tokenizer(
                    uncond_tokens,
                    padding="max_length",
                    max_length=128,
                    truncation=True,
                    return_attention_mask=True,
                    return_tensors="pt",
                )
                text_input_ids = uncond_input.input_ids.to(device)
                negative_attention_mask = uncond_input.attention_mask.to(device)

                negative_prompt_embeds = self.text_encoder(
                    text_input_ids,
                    attention_mask=negative_attention_mask,
                )
                negative_prompt_embeds = negative_prompt_embeds[0]
                negative_prompt_embeds = negative_prompt_embeds[:, : prompt_embeds.shape[1]]
                negative_attention_mask = negative_attention_mask[:, : prompt_embeds.shape[1]]
                negative_prompt_embeds = negative_prompt_embeds * negative_attention_mask.unsqueeze(2)

            else:
                negative_prompt_embeds = torch.zeros_like(prompt_embeds)
                negative_attention_mask = torch.zeros_like(attention_mask)

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

            negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
            if negative_prompt_embeds.shape != prompt_embeds.shape:
                negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
                negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
                negative_attention_mask = negative_attention_mask.repeat(num_images_per_prompt, 1)

            # 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
        else:
            negative_prompt_embeds = None
            negative_attention_mask = None
        return prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask

    def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
            raise ValueError(
                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
            )

        image = image.to(device=device, dtype=dtype)

        batch_size = batch_size * num_images_per_prompt

        if image.shape[1] == 4:
            init_latents = image

        else:
            if isinstance(generator, list) and len(generator) != batch_size:
                raise ValueError(
                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
                )

            elif isinstance(generator, list):
                init_latents = [
                    self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
                ]
                init_latents = torch.cat(init_latents, dim=0)
            else:
                init_latents = self.movq.encode(image).latent_dist.sample(generator)

            init_latents = self.movq.config.scaling_factor * init_latents

        init_latents = torch.cat([init_latents], dim=0)

        shape = init_latents.shape
        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)

        # get latents
        init_latents = self.scheduler.add_noise(init_latents, noise, timestep)

        latents = init_latents

        return latents

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, generator, 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

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def check_inputs(
        self,
        prompt,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
    ):
        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 prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (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 negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None,
        strength: float = 0.3,
        num_inference_steps: int = 25,
        guidance_scale: float = 3.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        latents=None,
    ):
        cut_context = True
        # 1. Check inputs. Raise error if not correct
        self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)

        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        device = self._execution_device

        # 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
        prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = self.encode_prompt(
            prompt,
            do_classifier_free_guidance,
            num_images_per_prompt=num_images_per_prompt,
            device=device,
            negative_prompt=negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            _cut_context=cut_context,
        )

        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
            attention_mask = torch.cat([negative_attention_mask, attention_mask]).bool()
        if not isinstance(image, list):
            image = [image]
        if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image):
            raise ValueError(
                f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support  PIL image and pytorch tensor"
            )

        image = torch.cat([prepare_image(i) for i in image], dim=0)
        image = image.to(dtype=prompt_embeds.dtype, device=device)
        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
        # 5. Prepare latents
        latents = self.movq.encode(image)["latents"]
        latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
        latents = self.prepare_latents(
            latents, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
        )
        if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
            self.text_encoder_offload_hook.offload()

        # 7. Denoising loop
        # TODO(Yiyi): Correct the following line and use correctly
        # num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    encoder_attention_mask=attention_mask,
                )[0]
                if do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)

                    noise_pred = (guidance_scale + 1.0) * noise_pred_text - guidance_scale * noise_pred_uncond

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(
                    noise_pred,
                    t,
                    latents,
                    generator=generator,
                ).prev_sample
                progress_bar.update()
                if callback is not None and i % callback_steps == 0:
                    step_idx = i // getattr(self.scheduler, "order", 1)
                    callback(step_idx, t, latents)
            # post-processing
            image = self.movq.decode(latents, force_not_quantize=True)["sample"]

            if output_type not in ["pt", "np", "pil"]:
                raise ValueError(
                    f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}"
                )

            if output_type in ["np", "pil"]:
                image = image * 0.5 + 0.5
                image = image.clamp(0, 1)
                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,)

            return ImagePipelineOutput(images=image)