WIP: img2img

1 files changed, 342 insertions, 0 deletions

diff --git a/pipelines/stable_diffusion/vlpn_stable_diffusion.py b/pipelines/stable_diffusion/vlpn_stable_diffusion.py
new file mode 100644
index 0000000..4c793a8
--- /dev/null
+++ b/pipelines/stable_diffusion/vlpn_stable_diffusion.py
@@ -0,0 +1,342 @@
+import inspect
+import warnings
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+import PIL
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers import AutoencoderKL, DiffusionPipeline, DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.utils import logging
+from transformers import CLIPTextModel, CLIPTokenizer
+from schedulers.scheduling_euler_a import EulerAScheduler, CFGDenoiserForward
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def preprocess(image, w, h):
+    image = image.resize((w, h), resample=PIL.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
+
+
+class VlpnStableDiffusion(DiffusionPipeline):
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler, EulerAScheduler],
+        **kwargs,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            warnings.warn(
+                f"The configuration file of this scheduler: {scheduler} is outdated.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",
+                DeprecationWarning,
+            )
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        strength: float = 0.8,
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[Union[torch.FloatTensor, PIL.Image.Image]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            strength (`float`, *optional*, defaults to 0.8):
+                Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1.
+                `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The
+                number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
+                noise will be maximum and the denoising process will run for the full number of iterations specified in
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`.
+            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.
+            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 (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *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.
+
+        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 negative_prompt is None:
+            negative_prompt = [""] * batch_size
+        elif isinstance(negative_prompt, str):
+            negative_prompt = [negative_prompt] * batch_size
+        elif isinstance(negative_prompt, list):
+            if len(negative_prompt) != batch_size:
+                raise ValueError(
+                    f"`prompt` and `negative_prompt` have to be the same length, but are {len(prompt)} and {len(negative_prompt)}")
+        else:
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_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 strength < 0 or strength > 1:
+            raise ValueError(f"`strength` should in [0.0, 1.0] but is {strength}")
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        offset = self.scheduler.config.get("steps_offset", 0)
+
+        if latents is not None and isinstance(latents, PIL.Image.Image):
+            latents = preprocess(latents, width, height)
+            latent_dist = self.vae.encode(latents.to(self.device)).latent_dist
+            latents = latent_dist.sample(generator=generator)
+            latents = 0.18215 * latents
+            latents = torch.cat([latents] * batch_size)
+
+            # get the original timestep using init_timestep
+            init_timestep = int(num_inference_steps * strength) + offset
+            init_timestep = min(init_timestep, num_inference_steps)
+
+            if isinstance(self.scheduler, LMSDiscreteScheduler):
+                timesteps = torch.tensor(
+                    [num_inference_steps - init_timestep] * batch_size, dtype=torch.long, device=self.device
+                )
+            elif isinstance(self.scheduler, EulerAScheduler):
+                timesteps = self.scheduler.timesteps[-init_timestep]
+                timesteps = torch.tensor([timesteps] * batch_size, device=self.device)
+            else:
+                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(latents.shape, generator=generator, device=self.device)
+            latents = self.scheduler.add_noise(latents, noise, timesteps)
+        else:
+            init_timestep = num_inference_steps + offset
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])
+            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}"
+            )
+            print(f"Too many tokens: {removed_text}")
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # 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
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                negative_prompt, padding="max_length", max_length=max_length, return_tensors="pt"
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[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 = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_device = "cpu" if self.device.type == "mps" else self.device
+        latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
+        if latents is None:
+            latents = torch.randn(
+                latents_shape,
+                generator=generator,
+                device=latents_device,
+                dtype=text_embeddings.dtype,
+            )
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        t_start = max(num_inference_steps - init_timestep + offset, 0)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimzed to move all timesteps to correct device beforehand
+        timesteps_tensor = torch.tensor(self.scheduler.timesteps[t_start:], device=self.device)
+
+        # if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas
+        if isinstance(self.scheduler, LMSDiscreteScheduler):
+            latents = latents * self.scheduler.sigmas[0]
+        elif isinstance(self.scheduler, EulerAScheduler):
+            sigma = self.scheduler.timesteps[0]
+            latents = latents * 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]
+        scheduler_step_args = set(inspect.signature(self.scheduler.step).parameters.keys())
+        accepts_eta = "eta" in scheduler_step_args
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+        accepts_generator = "generator" in scheduler_step_args
+        if generator is not None and accepts_generator:
+            extra_step_kwargs["generator"] = generator
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            t_index = t_start + i
+
+            # 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[t_index]
+                # the model input needs to be scaled to match the continuous ODE formulation in K-LMS
+                latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
+
+            noise_pred = None
+            if isinstance(self.scheduler, EulerAScheduler):
+                sigma = t.reshape(1)
+                sigma_in = torch.cat([sigma] * latent_model_input.shape[0])
+                # noise_pred = model(latent_model_input,sigma_in,uncond_embeddings, text_embeddings,guidance_scale)
+                noise_pred = CFGDenoiserForward(self.unet, latent_model_input, sigma_in,
+                                                text_embeddings, guidance_scale, quantize=True, DSsigmas=self.scheduler.DSsigmas)
+                # noise_pred = self.unet(latent_model_input, sigma_in, encoder_hidden_states=text_embeddings).sample
+            else:
+                # predict the noise residual
+                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).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, t_index, latents, **extra_step_kwargs).prev_sample
+            elif isinstance(self.scheduler, EulerAScheduler):
+                if t_index < self.scheduler.timesteps.shape[0] - 1:  # avoid out of bound error
+                    t_prev = self.scheduler.timesteps[t_index+1]
+                    latents = self.scheduler.step(noise_pred, t, t_prev, latents, **extra_step_kwargs).prev_sample
+            else:
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+        # scale and decode the image latents with vae
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).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, None)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)