image_to_image_and_score.py

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

import numpy as np
import torch

from PIL import Image
from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    DiffusionPipeline,
    PNDMScheduler,
    LMSDiscreteScheduler,
    UNet2DConditionModel,
)
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
from diffusers.pipelines.stable_diffusion.safety_checker import cosine_distance
from tqdm.auto import tqdm
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer


def preprocess_init_image(image: Image, width: int, height: int):
    image = image.resize((width, height), resample=Image.LANCZOS)
    image = np.array(image).astype(np.float32) / 255.0
    image = image[None].transpose(0, 3, 1, 2)
    image = torch.from_numpy(image)
    return 2.0 * image - 1.0


def preprocess_mask(mask: Image, width: int, height: int):
    mask = mask.convert("L")
    mask = mask.resize((width // 8, height // 8), resample=Image.LANCZOS)
    mask = np.array(mask).astype(np.float32) / 255.0
    mask = np.tile(mask, (4, 1, 1))
    mask = mask[None].transpose(0, 1, 2, 3)  # what does this step do?
    mask = torch.from_numpy(mask)
    return mask


class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
    """
    From https://github.com/huggingface/diffusers/pull/241
    """

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPFeatureExtractor,
    ):
        super().__init__()
        scheduler = scheduler.set_format("pt")
        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        init_image: Optional[torch.FloatTensor],
        mask: Optional[torch.FloatTensor],
        width: int,
        height: int,
        prompt_strength: float = 0.8,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        eta: float = 0.0,
        generator: Optional[torch.Generator] = None,
    ) -> Image:
        if isinstance(prompt, str):
            batch_size = 1
            prompt = [prompt]
        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 prompt_strength < 0 or prompt_strength > 1:
            raise ValueError(
                f"The value of prompt_strength should in [0.0, 1.0] but is {prompt_strength}"
            )

        if mask is not None and init_image is None:
            raise ValueError(
                "If mask is defined, then init_image also needs to be defined"
            )

        if width % 8 != 0 or height % 8 != 0:
            raise ValueError("Width and height must both be divisible by 8")

        # set timesteps
        accepts_offset = "offset" in set(
            inspect.signature(self.scheduler.set_timesteps).parameters.keys()
        )
        extra_set_kwargs = {}
        offset = 0
        if accepts_offset:
            offset = 1
            extra_set_kwargs["offset"] = 1

        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)

        text_embeddings = self.embed_text(prompt=prompt)
        uncond_embeddings = self.embed_text(
            prompt=[""] * batch_size
        )  # only need to call this once

        if init_image is not None:
            init_latents_orig, latents, init_timestep = self.latents_from_init_image(
                init_image,
                prompt_strength,
                offset,
                num_inference_steps,
                batch_size,
                generator,
            )
        else:
            latents = torch.randn(
                (batch_size, self.unet.in_channels, height // 8, width // 8),
                generator=generator,
                device=self.device,
            )
            init_timestep = num_inference_steps

        do_classifier_free_guidance = guidance_scale > 1.0
        if do_classifier_free_guidance:
            sd_hidden_states = torch.cat([uncond_embeddings, text_embeddings], dim=0)
        else:
            sd_hidden_states = text_embeddings

        # 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

        mask_noise = torch.randn(latents.shape, generator=generator, device=self.device)

        # if we use LMSDiscreteScheduler, let's make sure latents are mulitplied by sigmas
        if isinstance(self.scheduler, LMSDiscreteScheduler):
            latents = latents * self.scheduler.sigmas[0]

        t_start = max(num_inference_steps - init_timestep + offset, 0)
        for i, t in tqdm(enumerate(self.scheduler.timesteps[t_start:])):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = (
                torch.cat([latents] * 2) if do_classifier_free_guidance else latents
            )

            if isinstance(self.scheduler, LMSDiscreteScheduler):
                sigma = self.scheduler.sigmas[i]
                latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)

            # predict the noise residual
            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=sd_hidden_states,  # hidden state/context is the uncond embedding and the text embedding
            )["sample"]

            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.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 isinstance(self.scheduler, LMSDiscreteScheduler):
                latents = self.scheduler.step(
                    noise_pred, i, latents, **extra_step_kwargs
                )["prev_sample"]
            else:
                latents = self.scheduler.step(
                    noise_pred, t, latents, **extra_step_kwargs
                )["prev_sample"]

            # replace the unmasked part with original latents, with added noise
            if mask is not None:
                timesteps = self.scheduler.timesteps[t_start + i]
                timesteps = torch.tensor(
                    [timesteps] * batch_size, dtype=torch.long, device=self.device
                )
                noisy_init_latents = self.scheduler.add_noise(
                    init_latents_orig, mask_noise, timesteps
                )
                latents = noisy_init_latents * mask + latents * (1 - mask)

        # scale and decode the image latents with vae
        latents = 1 / 0.18215 * latents
        decoded_image = self.vae.decode(latents)
        decoded_image = (decoded_image / 2 + 0.5).clamp(0, 1)
        decoded_image = decoded_image.cpu().permute(0,2,3,1).numpy()

        # only need the "first" prompt since it is repeated batch_size times
        pooled_text_embeddings = self.embed_text(prompt=[prompt[0]], pooled_output=True)

        # get pooled image embeddings for cosine similarity (1, 768)
        clip_image_tensors = self.feature_extractor(
            self.numpy_to_pil(decoded_image), return_tensors="pt"
        ).to(self.device)

        clip_image_tensors = clip_image_tensors.pixel_values

        pooled_image_embeddings = self.image_to_embedding(image_tensors=clip_image_tensors)

        scores = (
            cosine_distance(pooled_image_embeddings, pooled_text_embeddings)
        ).detach().cpu().numpy()

        pil_image = self.numpy_to_pil(decoded_image)

        return {
            "sample": pil_image,
            "score": scores
            # "nsfw_content_detected": has_nsfw_concept,
        }
        # # check safety
        # _, has_nsfw_concept = self.safety_checker(
        #     images=image, clip_input=image_tensors
        # )  # TODO: check if this is correct

    def latents_from_init_image(
        self,
        init_image: torch.FloatTensor,
        prompt_strength: float,
        offset: int,
        num_inference_steps: int,
        batch_size: int,
        generator: Optional[torch.Generator],
    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, int]:
        # encode the init image into latents and scale the latents
        init_latents = self.vae.encode(init_image.to(self.device)).sample()
        init_latents = 0.18215 * init_latents
        init_latents_orig = init_latents

        # prepare init_latents noise to latents
        init_latents = torch.cat([init_latents] * batch_size)

        # get the original timestep using init_timestep
        init_timestep = int(num_inference_steps * prompt_strength) + offset
        init_timestep = min(init_timestep, num_inference_steps)
        timesteps = self.scheduler.timesteps[-init_timestep]
        timesteps = torch.tensor(
            [timesteps] * batch_size, dtype=torch.long, device=self.device
        )

        # add noise to latents using the timesteps
        noise = torch.randn(init_latents.shape, generator=generator, device=self.device)
        init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)

        return init_latents_orig, init_latents, init_timestep

    def embed_text(
        self,
        prompt: List[str],
        pooled_output: bool = False,
    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
        """Embed text using CLIP model."""
        assert isinstance(
            prompt, list
        ), f"prompt {prompt} must be a list of strings, try wrapping it in a list e.g. [prompt]"
        # get prompt text embeddings
        text_input = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_embeds = self.text_encoder(
            text_input.input_ids.to(self.device)
        )  # last_hiden_state (batch_size, seq_len, dim)
        if pooled_output:
            return text_embeds[1]
        else:
            return text_embeds[0]

    def image_to_embedding(self, image_tensors: torch.FloatTensor) -> torch.FloatTensor:
        pooled_image_embeddings = self.safety_checker.vision_model(
            pixel_values=image_tensors
        ).pooler_output
        pooled_image_embeddings = self.safety_checker.visual_projection(
            pooled_image_embeddings
        )  # we have to project the pooled image embeddings to the same dimension as the text embeddings
        return pooled_image_embeddings