tsi-org/tango

1"""2    modeled after the textual_inversion.py / train_dreambooth.py and the work3    of justinpinkney here: https://github.com/justinpinkney/stable-diffusion/blob/main/notebooks/imagic.ipynb4"""5import inspect6import warnings7from typing import List, Optional, Union8 9import numpy as np10import PIL11import torch12import torch.nn.functional as F13from accelerate import Accelerator14 15# TODO: remove and import from diffusers.utils when the new version of diffusers is released16from packaging import version17from tqdm.auto import tqdm18from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer19 20from diffusers import DiffusionPipeline21from diffusers.models import AutoencoderKL, UNet2DConditionModel22from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput23from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker24from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler25from diffusers.utils import logging26 27 28if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):29    PIL_INTERPOLATION = {30        "linear": PIL.Image.Resampling.BILINEAR,31        "bilinear": PIL.Image.Resampling.BILINEAR,32        "bicubic": PIL.Image.Resampling.BICUBIC,33        "lanczos": PIL.Image.Resampling.LANCZOS,34        "nearest": PIL.Image.Resampling.NEAREST,35    }36else:37    PIL_INTERPOLATION = {38        "linear": PIL.Image.LINEAR,39        "bilinear": PIL.Image.BILINEAR,40        "bicubic": PIL.Image.BICUBIC,41        "lanczos": PIL.Image.LANCZOS,42        "nearest": PIL.Image.NEAREST,43    }44# ------------------------------------------------------------------------------45 46logger = logging.get_logger(__name__)  # pylint: disable=invalid-name47 48 49def preprocess(image):50    w, h = image.size51    w, h = (x - x % 32 for x in (w, h))  # resize to integer multiple of 3252    image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])53    image = np.array(image).astype(np.float32) / 255.054    image = image[None].transpose(0, 3, 1, 2)55    image = torch.from_numpy(image)56    return 2.0 * image - 1.057 58 59class ImagicStableDiffusionPipeline(DiffusionPipeline):60    r"""61    Pipeline for imagic image editing.62    See paper here: https://arxiv.org/pdf/2210.09276.pdf63 64    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the65    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)66    Args:67        vae ([`AutoencoderKL`]):68            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.69        text_encoder ([`CLIPTextModel`]):70            Frozen text-encoder. Stable Diffusion uses the text portion of71            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically72            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.73        tokenizer (`CLIPTokenizer`):74            Tokenizer of class75            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).76        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.77        scheduler ([`SchedulerMixin`]):78            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of79            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].80        safety_checker ([`StableDiffusionSafetyChecker`]):81            Classification module that estimates whether generated images could be considered offsensive or harmful.82            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.83        feature_extractor ([`CLIPImageProcessor`]):84            Model that extracts features from generated images to be used as inputs for the `safety_checker`.85    """86 87    def __init__(88        self,89        vae: AutoencoderKL,90        text_encoder: CLIPTextModel,91        tokenizer: CLIPTokenizer,92        unet: UNet2DConditionModel,93        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],94        safety_checker: StableDiffusionSafetyChecker,95        feature_extractor: CLIPImageProcessor,96    ):97        super().__init__()98        self.register_modules(99            vae=vae,100            text_encoder=text_encoder,101            tokenizer=tokenizer,102            unet=unet,103            scheduler=scheduler,104            safety_checker=safety_checker,105            feature_extractor=feature_extractor,106        )107 108    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):109        r"""110        Enable sliced attention computation.111        When this option is enabled, the attention module will split the input tensor in slices, to compute attention112        in several steps. This is useful to save some memory in exchange for a small speed decrease.113        Args:114            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):115                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If116                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,117                `attention_head_dim` must be a multiple of `slice_size`.118        """119        if slice_size == "auto":120            # half the attention head size is usually a good trade-off between121            # speed and memory122            slice_size = self.unet.config.attention_head_dim // 2123        self.unet.set_attention_slice(slice_size)124 125    def disable_attention_slicing(self):126        r"""127        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go128        back to computing attention in one step.129        """130        # set slice_size = `None` to disable `attention slicing`131        self.enable_attention_slicing(None)132 133    def train(134        self,135        prompt: Union[str, List[str]],136        image: Union[torch.FloatTensor, PIL.Image.Image],137        height: Optional[int] = 512,138        width: Optional[int] = 512,139        generator: Optional[torch.Generator] = None,140        embedding_learning_rate: float = 0.001,141        diffusion_model_learning_rate: float = 2e-6,142        text_embedding_optimization_steps: int = 500,143        model_fine_tuning_optimization_steps: int = 1000,144        **kwargs,145    ):146        r"""147        Function invoked when calling the pipeline for generation.148        Args:149            prompt (`str` or `List[str]`):150                The prompt or prompts to guide the image generation.151            height (`int`, *optional*, defaults to 512):152                The height in pixels of the generated image.153            width (`int`, *optional*, defaults to 512):154                The width in pixels of the generated image.155            num_inference_steps (`int`, *optional*, defaults to 50):156                The number of denoising steps. More denoising steps usually lead to a higher quality image at the157                expense of slower inference.158            guidance_scale (`float`, *optional*, defaults to 7.5):159                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).160                `guidance_scale` is defined as `w` of equation 2. of [Imagen161                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >162                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,163                usually at the expense of lower image quality.164            eta (`float`, *optional*, defaults to 0.0):165                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to166                [`schedulers.DDIMScheduler`], will be ignored for others.167            generator (`torch.Generator`, *optional*):168                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation169                deterministic.170            latents (`torch.FloatTensor`, *optional*):171                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image172                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents173                tensor will ge generated by sampling using the supplied random `generator`.174            output_type (`str`, *optional*, defaults to `"pil"`):175                The output format of the generate image. Choose between176                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `nd.array`.177            return_dict (`bool`, *optional*, defaults to `True`):178                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a179                plain tuple.180        Returns:181            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:182            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.183            When returning a tuple, the first element is a list with the generated images, and the second element is a184            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"185            (nsfw) content, according to the `safety_checker`.186        """187        accelerator = Accelerator(188            gradient_accumulation_steps=1,189            mixed_precision="fp16",190        )191 192        if "torch_device" in kwargs:193            device = kwargs.pop("torch_device")194            warnings.warn(195                "`torch_device` is deprecated as an input argument to `__call__` and will be removed in v0.3.0."196                " Consider using `pipe.to(torch_device)` instead."197            )198 199            if device is None:200                device = "cuda" if torch.cuda.is_available() else "cpu"201            self.to(device)202 203        if height % 8 != 0 or width % 8 != 0:204            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")205 206        # Freeze vae and unet207        self.vae.requires_grad_(False)208        self.unet.requires_grad_(False)209        self.text_encoder.requires_grad_(False)210        self.unet.eval()211        self.vae.eval()212        self.text_encoder.eval()213 214        if accelerator.is_main_process:215            accelerator.init_trackers(216                "imagic",217                config={218                    "embedding_learning_rate": embedding_learning_rate,219                    "text_embedding_optimization_steps": text_embedding_optimization_steps,220                },221            )222 223        # get text embeddings for prompt224        text_input = self.tokenizer(225            prompt,226            padding="max_length",227            max_length=self.tokenizer.model_max_length,228            truncation=True,229            return_tensors="pt",230        )231        text_embeddings = torch.nn.Parameter(232            self.text_encoder(text_input.input_ids.to(self.device))[0], requires_grad=True233        )234        text_embeddings = text_embeddings.detach()235        text_embeddings.requires_grad_()236        text_embeddings_orig = text_embeddings.clone()237 238        # Initialize the optimizer239        optimizer = torch.optim.Adam(240            [text_embeddings],  # only optimize the embeddings241            lr=embedding_learning_rate,242        )243 244        if isinstance(image, PIL.Image.Image):245            image = preprocess(image)246 247        latents_dtype = text_embeddings.dtype248        image = image.to(device=self.device, dtype=latents_dtype)249        init_latent_image_dist = self.vae.encode(image).latent_dist250        image_latents = init_latent_image_dist.sample(generator=generator)251        image_latents = 0.18215 * image_latents252 253        progress_bar = tqdm(range(text_embedding_optimization_steps), disable=not accelerator.is_local_main_process)254        progress_bar.set_description("Steps")255 256        global_step = 0257 258        logger.info("First optimizing the text embedding to better reconstruct the init image")259        for _ in range(text_embedding_optimization_steps):260            with accelerator.accumulate(text_embeddings):261                # Sample noise that we'll add to the latents262                noise = torch.randn(image_latents.shape).to(image_latents.device)263                timesteps = torch.randint(1000, (1,), device=image_latents.device)264 265                # Add noise to the latents according to the noise magnitude at each timestep266                # (this is the forward diffusion process)267                noisy_latents = self.scheduler.add_noise(image_latents, noise, timesteps)268 269                # Predict the noise residual270                noise_pred = self.unet(noisy_latents, timesteps, text_embeddings).sample271 272                loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()273                accelerator.backward(loss)274 275                optimizer.step()276                optimizer.zero_grad()277 278            # Checks if the accelerator has performed an optimization step behind the scenes279            if accelerator.sync_gradients:280                progress_bar.update(1)281                global_step += 1282 283            logs = {"loss": loss.detach().item()}  # , "lr": lr_scheduler.get_last_lr()[0]}284            progress_bar.set_postfix(**logs)285            accelerator.log(logs, step=global_step)286 287        accelerator.wait_for_everyone()288 289        text_embeddings.requires_grad_(False)290 291        # Now we fine tune the unet to better reconstruct the image292        self.unet.requires_grad_(True)293        self.unet.train()294        optimizer = torch.optim.Adam(295            self.unet.parameters(),  # only optimize unet296            lr=diffusion_model_learning_rate,297        )298        progress_bar = tqdm(range(model_fine_tuning_optimization_steps), disable=not accelerator.is_local_main_process)299 300        logger.info("Next fine tuning the entire model to better reconstruct the init image")301        for _ in range(model_fine_tuning_optimization_steps):302            with accelerator.accumulate(self.unet.parameters()):303                # Sample noise that we'll add to the latents304                noise = torch.randn(image_latents.shape).to(image_latents.device)305                timesteps = torch.randint(1000, (1,), device=image_latents.device)306 307                # Add noise to the latents according to the noise magnitude at each timestep308                # (this is the forward diffusion process)309                noisy_latents = self.scheduler.add_noise(image_latents, noise, timesteps)310 311                # Predict the noise residual312                noise_pred = self.unet(noisy_latents, timesteps, text_embeddings).sample313 314                loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()315                accelerator.backward(loss)316 317                optimizer.step()318                optimizer.zero_grad()319 320            # Checks if the accelerator has performed an optimization step behind the scenes321            if accelerator.sync_gradients:322                progress_bar.update(1)323                global_step += 1324 325            logs = {"loss": loss.detach().item()}  # , "lr": lr_scheduler.get_last_lr()[0]}326            progress_bar.set_postfix(**logs)327            accelerator.log(logs, step=global_step)328 329        accelerator.wait_for_everyone()330        self.text_embeddings_orig = text_embeddings_orig331        self.text_embeddings = text_embeddings332 333    @torch.no_grad()334    def __call__(335        self,336        alpha: float = 1.2,337        height: Optional[int] = 512,338        width: Optional[int] = 512,339        num_inference_steps: Optional[int] = 50,340        generator: Optional[torch.Generator] = None,341        output_type: Optional[str] = "pil",342        return_dict: bool = True,343        guidance_scale: float = 7.5,344        eta: float = 0.0,345    ):346        r"""347        Function invoked when calling the pipeline for generation.348        Args:349            prompt (`str` or `List[str]`):350                The prompt or prompts to guide the image generation.351            height (`int`, *optional*, defaults to 512):352                The height in pixels of the generated image.353            width (`int`, *optional*, defaults to 512):354                The width in pixels of the generated image.355            num_inference_steps (`int`, *optional*, defaults to 50):356                The number of denoising steps. More denoising steps usually lead to a higher quality image at the357                expense of slower inference.358            guidance_scale (`float`, *optional*, defaults to 7.5):359                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).360                `guidance_scale` is defined as `w` of equation 2. of [Imagen361                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >362                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,363                usually at the expense of lower image quality.364            eta (`float`, *optional*, defaults to 0.0):365                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to366                [`schedulers.DDIMScheduler`], will be ignored for others.367            generator (`torch.Generator`, *optional*):368                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation369                deterministic.370            latents (`torch.FloatTensor`, *optional*):371                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image372                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents373                tensor will ge generated by sampling using the supplied random `generator`.374            output_type (`str`, *optional*, defaults to `"pil"`):375                The output format of the generate image. Choose between376                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `nd.array`.377            return_dict (`bool`, *optional*, defaults to `True`):378                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a379                plain tuple.380        Returns:381            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:382            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.383            When returning a tuple, the first element is a list with the generated images, and the second element is a384            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"385            (nsfw) content, according to the `safety_checker`.386        """387        if height % 8 != 0 or width % 8 != 0:388            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")389        if self.text_embeddings is None:390            raise ValueError("Please run the pipe.train() before trying to generate an image.")391        if self.text_embeddings_orig is None:392            raise ValueError("Please run the pipe.train() before trying to generate an image.")393 394        text_embeddings = alpha * self.text_embeddings_orig + (1 - alpha) * self.text_embeddings395 396        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)397        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`398        # corresponds to doing no classifier free guidance.399        do_classifier_free_guidance = guidance_scale > 1.0400        # get unconditional embeddings for classifier free guidance401        if do_classifier_free_guidance:402            uncond_tokens = [""]403            max_length = self.tokenizer.model_max_length404            uncond_input = self.tokenizer(405                uncond_tokens,406                padding="max_length",407                max_length=max_length,408                truncation=True,409                return_tensors="pt",410            )411            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]412 413            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method414            seq_len = uncond_embeddings.shape[1]415            uncond_embeddings = uncond_embeddings.view(1, seq_len, -1)416 417            # For classifier free guidance, we need to do two forward passes.418            # Here we concatenate the unconditional and text embeddings into a single batch419            # to avoid doing two forward passes420            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])421 422        # get the initial random noise unless the user supplied it423 424        # Unlike in other pipelines, latents need to be generated in the target device425        # for 1-to-1 results reproducibility with the CompVis implementation.426        # However this currently doesn't work in `mps`.427        latents_shape = (1, self.unet.in_channels, height // 8, width // 8)428        latents_dtype = text_embeddings.dtype429        if self.device.type == "mps":430            # randn does not exist on mps431            latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(432                self.device433            )434        else:435            latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)436 437        # set timesteps438        self.scheduler.set_timesteps(num_inference_steps)439 440        # Some schedulers like PNDM have timesteps as arrays441        # It's more optimized to move all timesteps to correct device beforehand442        timesteps_tensor = self.scheduler.timesteps.to(self.device)443 444        # scale the initial noise by the standard deviation required by the scheduler445        latents = latents * self.scheduler.init_noise_sigma446 447        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature448        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.449        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502450        # and should be between [0, 1]451        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())452        extra_step_kwargs = {}453        if accepts_eta:454            extra_step_kwargs["eta"] = eta455 456        for i, t in enumerate(self.progress_bar(timesteps_tensor)):457            # expand the latents if we are doing classifier free guidance458            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents459            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)460 461            # predict the noise residual462            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample463 464            # perform guidance465            if do_classifier_free_guidance:466                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)467                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)468 469            # compute the previous noisy sample x_t -> x_t-1470            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample471 472        latents = 1 / 0.18215 * latents473        image = self.vae.decode(latents).sample474 475        image = (image / 2 + 0.5).clamp(0, 1)476 477        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16478        image = image.cpu().permute(0, 2, 3, 1).float().numpy()479 480        if self.safety_checker is not None:481            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(482                self.device483            )484            image, has_nsfw_concept = self.safety_checker(485                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)486            )487        else:488            has_nsfw_concept = None489 490        if output_type == "pil":491            image = self.numpy_to_pil(image)492 493        if not return_dict:494            return (image, has_nsfw_concept)495 496        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)497