infer.py

# Copyright 2022 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.

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

import torch
import torch.nn as nn
from packaging import version
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer

from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import deprecate, is_accelerate_available, logging, randn_tensor, replace_example_docstring
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker

from clip_tokenizer import CLIPTextModel

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

import torch
import torch.nn as nn
import torch.nn.functional as F
import os
import clip
from torchvision import transforms

class StyleClip(nn.Module):
    """backbone + projection head"""
    def __init__(self, feat_dim=512, model_path=None):
        super(StyleClip, self).__init__()
        self.clip, self.transform = clip.load("xxx", download_root=model_path)
        dim_in = 512

        model_list = []
        # d_in = dim_in
        # d_out = feat_dim
        # for i in range(4):
        #     model_list += [nn.Linear(d_in, d_out), nn.ReLU(inplace=True)]
        #     d_in = d_out
        # model_list.append(nn.Linear(d_in, feat_dim))
        for i in range(4):
            model_list += [nn.Linear(dim_in, dim_in), nn.ReLU(inplace=True)]
            # d_in = d_out
        model_list.append(nn.Linear(dim_in, feat_dim))
        self.head = nn.Sequential(*model_list)

        self.freeze_encoder()

    def get_transform(self):
        return self.transform

    def get_resolution(self):
        return self.clip.visual.input_resolution

    def freeze_encoder(self, freeze=True):
        for param in self.clip.parameters():
            param.requires_grad = not freeze

    def forward(self, x):
        feat = self.clip.encode_image(x)
        feat = self.head(feat.to(torch.float))
        # print(type(self.head), self.head.state_dict().keys(), self.head)
        # feat = self.head(feat.type(self.head.state_dict()['0.weight'].dtype))
        # feat = F.normalize(self.head(feat), dim=1)
        return feat


class ClipMlp(nn.Module):
    """backbone + projection head"""
    def __init__(self, feat_dim=512, out_dim = 512, model_path=None):
        super(ClipMlp, self).__init__()
        self.encoder = StyleClip(feat_dim=feat_dim, model_path=model_path)
        self.head = nn.Linear(feat_dim, out_dim)
    
    def freeze_encoder(self, freeze=True):
        self.encoder.freeze_encoder(freeze)
    
    def get_transform(self):
        return self.encoder.get_transform()

    def get_resolution(self):
        return self.encoder.get_resolution()

    def forward(self, x):
        feat = self.encoder(x)
        feat = F.normalize(self.head(feat), dim=1)
        return feat



def load_style_encoder(device="cuda", download_root=None, with_head=False, transform_aug=False, style_ckpt=None, model_path=None):

    def _convert_image_to_rgb(image):
        return image.convert("RGB")

    print("load StyleMlp encoder")
    state_dict = torch.load(style_ckpt)
    out_dim = state_dict["model"]['head.weight'].shape[0]
    model = ClipMlp(feat_dim=768, out_dim=out_dim, model_path=model_path)

    state_dict = state_dict["model"]
    state_dict2 = {}
    for key in state_dict:
        key2 = key.replace(".module.",".")
        state_dict2[key2] = state_dict[key]
    model.load_state_dict(state_dict2)
    if not with_head:
        model = model.encoder
    model = model.to(device)
    model.eval()
    resolution = model.get_resolution()

    if transform_aug:
        train_transform = transforms.Compose([
            transforms.RandomRotation(180),
            transforms.RandomResizedCrop(size=[resolution, resolution], scale=(0.2, 1.0)),
            transforms.RandomHorizontalFlip(),
            _convert_image_to_rgb,
            transforms.ToTensor(),
            transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
        ])
    else:
        train_transform = transforms.Compose([
            transforms.Resize([resolution, resolution]),
            _convert_image_to_rgb,
            transforms.ToTensor(),
            transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
        ])

    return model, train_transform


class StyleTokenizer(nn.Module):
    """backbone + projection head"""
    def __init__(self, input_size=None, intermediate_size=512, out_size=768, n_tokens=8, with_placeholder=False, prefix_model="mlp"):
        super(StyleTokenizer, self).__init__()
        if input_size is None:
            input_size=512
        self.n_tokens=n_tokens
        self.with_placeholder=with_placeholder
        self.out_size=out_size
        self.prefix_model = prefix_model
        print(f"StyleTokenizer with prefix_model: {prefix_model}")
        if self.prefix_model == "mlp" or self.prefix_model == "vgg" or self.prefix_model == "clip":
            self.proj = nn.Sequential(
                nn.Linear(input_size, intermediate_size), # apply a fully connected layer with output size intermediate_size
                nn.SiLU(),
                nn.Linear(intermediate_size, out_size * n_tokens),
            )
        elif self.prefix_model == "transformer":
            input_size=768
            self.proj=nn.Linear(input_size, out_size)
        else:
            assert False

        if self.with_placeholder:
            self.start_embedding=nn.Parameter(torch.randn(1, 1, out_size))
            self.end_embedding=nn.Parameter(torch.randn(1, 1, out_size))

    def forward(self, x):
        x = self.proj(x.to(torch.float))
        if self.prefix_model == "mlp" or self.prefix_model == "vgg" or self.prefix_model == "clip":
            x = x.reshape(x.shape[0], self.n_tokens, self.out_size)

        if self.with_placeholder:
            start = self.start_embedding.repeat(x.shape[0], 1, 1)
            end = self.end_embedding.repeat(x.shape[0], 1, 1)
            x = torch.cat([start, x, end], dim=1)

        return x


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import StableDiffusionPipeline

        >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
        >>> pipe = pipe.to("cuda")

        >>> prompt = "a photo of an astronaut riding a horse on mars"
        >>> image = pipe(prompt).images[0]
        ```
"""


class StableDiffusionPipelineWithStyle(DiffusionPipeline):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion.

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

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
        feature_extractor ([`CLIPFeatureExtractor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """
    _optional_components = ["safety_checker", "feature_extractor"]

    @classmethod
    def load_pretrained(cls, model_path, device="cuda"):
        pipe = StableDiffusionPipelineWithStyle.from_pretrained(model_path)
        pipe = pipe.to(device)
        # pipe.device=device

        style_ckpt=os.path.join(model_path, "xxx.pth")
        style_encoder, style_preprocess = load_style_encoder(device = device, style_ckpt=style_ckpt, model_path=model_path)
        style_encoder.eval()
        pipe.style_encoder = style_encoder
        pipe.style_preprocess = style_preprocess

        fusion_model_path = os.path.join(model_path, "linear.ckpt")
        style_tokenizer_state_dict = torch.load(fusion_model_path)
        if "start_embedding" in style_tokenizer_state_dict.keys():
            with_placeholder = True
        else:
            with_placeholder = False
        style_tokenizer_out_dim = 1024
        n_tokens = style_tokenizer_state_dict["proj.2.bias"].shape[0]//style_tokenizer_out_dim
        style_token_input_size = style_tokenizer_state_dict["proj.0.weight"].shape[1]
        style_tokenizer = StyleTokenizer(input_size=style_token_input_size, out_size=style_tokenizer_out_dim, n_tokens=n_tokens, with_placeholder=with_placeholder, prefix_model="mlp")
        style_tokenizer.load_state_dict(torch.load(fusion_model_path))

        # style_tokenizer = StyleTokenizer(input_size=512, intermediate_size=512, out_size=1024, n_tokens=8, with_placeholder=False)
        # fusion_model_path = os.path.join(model_path, "linear.ckpt")
        # style_tokenizer.load_state_dict(torch.load(fusion_model_path))

        style_tokenizer=style_tokenizer.to(device)
        pipe.style_tokenizer = style_tokenizer
        text_encoder = CLIPTextModel.from_pretrained(
            model_path, subfolder="text_encoder"
        ).to(device)
        pipe.text_encoder = text_encoder
        return pipe

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPFeatureExtractor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
            deprecation_message = (
                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"
            )
            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
                " config accordingly as not setting `clip_sample` in the config might lead 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"
            )
            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["clip_sample"] = False
            scheduler._internal_dict = FrozenDict(new_config)

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
            version.parse(unet.config._diffusers_version).base_version
        ) < version.parse("0.9.0.dev0")
        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
            deprecation_message = (
                "The configuration file of the unet has set the default `sample_size` to smaller than"
                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
                " in the config might lead 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 `unet/config.json` file"
            )
            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(unet.config)
            new_config["sample_size"] = 64
            unet._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
            # style_encoder = style_encoder
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.register_to_config(requires_safety_checker=requires_safety_checker)
        # self.fusion_model = FusionModel()
        self.fc_linear = torch.nn.Linear(1024, 768)

    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding.

        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
        steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_sequential_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
        """
        if is_accelerate_available():
            from accelerate import cpu_offload
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

        device = torch.device(f"cuda:{gpu_id}")

        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
            cpu_offload(cpu_offloaded_model, device)

        if self.safety_checker is not None:
            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)

    @property
    def _execution_device(self):
        r"""
        Returns the device on which the pipeline's models will be executed. After calling
        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
        hooks.
        """
        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    def _encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        style_token_id=None,
        style_latent=None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
             prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                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 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]
        batch_size = 1

        # if prompt_embeds is None:
        assert prompt_embeds is None
        #########
        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
            text_input_ids, untruncated_ids
        ):
            removed_text = self.tokenizer.batch_decode(
                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
            )
            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}"
            )

        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
            attention_mask = text_inputs.attention_mask.to(device)
        else:
            attention_mask = None

        prompt_embeds = self.text_encoder(
            text_input_ids.to(device),
            attention_mask=attention_mask,
            style_embeds=style_latent, 
            style_id=style_token_id
        )
        prompt_embeds = prompt_embeds[0]

        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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)

        # 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 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)}."
                )
            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

            max_length = prompt_embeds.shape[1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        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=self.text_encoder.dtype, device=device)

            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)

            # 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
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

        return prompt_embeds

    def run_safety_checker(self, image, device, dtype):
        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        else:
            has_nsfw_concept = None
        return image, has_nsfw_concept

    def decode_latents(self, latents):
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents).sample
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (鐣�) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to 鐣� 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,
        height,
        width,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
    ):
        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 (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}."
                )

    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
        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."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        style_image = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        style_guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = 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: Optional[int] = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                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.
            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`).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (鐣�) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](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`.
            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.
            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.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).

        Examples:

        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 style_image is not None:
            style_image = self.style_preprocess(style_image)
            style_image = style_image.unsqueeze(0).to(self.device)

        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
        )

        # 2. Define call parameters
        # 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 or style_guidance_scale > 1.0

        # 3. Encode input prompt
        style_token_str = "<style_token>"
        self.tokenizer.add_tokens([style_token_str])
        style_token_id = self.tokenizer.convert_tokens_to_ids(style_token_str)
        if style_image is not None:
            n_tokens = self.style_tokenizer.n_tokens
            if self.style_tokenizer.with_placeholder:
                n_tokens += 2
            print("n_tokens", self.style_tokenizer.n_tokens, n_tokens)
            style_prompt = style_token_str * n_tokens + prompt
            # if style_guidance_scale == guidance_scale:
            #     prompt = None
                

        if style_image is not None:
            style_latent = self.style_encoder(style_image)
            style_latent = self.style_tokenizer(style_latent)
        else:
            style_latent = None

        multi_cfg = False if guidance_scale == style_guidance_scale else True

        style_prompt_embeds = self._encode_prompt(
            style_prompt,
            device,
            num_images_per_prompt,
            do_classifier_free_guidance,
            negative_prompt,
            prompt_embeds=None,
            negative_prompt_embeds=None,
            style_token_id=style_token_id,
            style_latent=style_latent,
        )

        if multi_cfg:
            prompt_embeds = self._encode_prompt(
                prompt,
                device,
                num_images_per_prompt,
                do_classifier_free_guidance=False,
                negative_prompt=None,
                prompt_embeds=None,
                negative_prompt_embeds=None,
                style_token_id=style_token_id,
                style_latent=None,
            )
            prompt_embeds = torch.cat([style_prompt_embeds, prompt_embeds])
        else:
            prompt_embeds = style_prompt_embeds
        prompt_embeds = self.fc_linear(prompt_embeds)        

        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        num_channels_latents = self.unet.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7. Denoising loop
        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):
                # expand the latents if we are doing classifier free guidance
                num_latents = prompt_embeds.shape[0]
                latent_model_input = torch.cat([latents] * num_latents) if do_classifier_free_guidance else latents
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=cross_attention_kwargs,
                ).sample

                # perform guidance
                if do_classifier_free_guidance:
                    if multi_cfg:
                        noise_pred_uncond, noise_pred_text_style, noise_pred_text = noise_pred.chunk(3)
                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + style_guidance_scale * (noise_pred_text_style - noise_pred_text)
                    else:
                        noise_pred_uncond, noise_pred_text_style = noise_pred.chunk(2)
                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text_style - noise_pred_uncond)
                
                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        callback(i, t, latents)

        if output_type == "latent":
            image = latents
            has_nsfw_concept = None
        elif output_type == "pil":
            # 8. Post-processing
            image = self.decode_latents(latents)

            # 9. Run safety checker
            # image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
            has_nsfw_concept = None
            # 10. Convert to PIL
            image = self.numpy_to_pil(image)
        else:
            # 8. Post-processing
            image = self.decode_latents(latents)

            # 9. Run safety checker
            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)


if __name__=="__main__":
    # ant1.4_freezeunet_class_mlp_styledrop0.1_textdrop0.1_aes6.0_b40_8_lr1e-4_0.1/checkpoint-24000
    from PIL import Image

    model_path = "xxx"
    pipe = StableDiffusionPipelineWithStyle.load_pretrained(model_path, device="cuda")
    pipe.fc_linear.load_state_dict(torch.load(os.path.join(model_path, 'xxx.pth')))
    pipe.fc_linear.to("cuda")#, dtype=torch.float16)
    pipe.fc_linear.eval()
    
    style_image_path = "xxx.jpeg" #"/path/to/style_image"
    style_image = Image.open(style_image_path).convert('RGB')

    prompt = "xxx" # 'your prompt'
    image_out = pipe(
        prompt=prompt, 
        style_image=style_image, 
        guidance_scale=7.5, 
        style_guidance_scale=7.5
    ).images[0]
    image_out.save("xxx.jpg")