ZeroScope と OpenVINO によるビデオ生成#

%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu "diffusers>=0.18.0" "torch>=2.1" transformers "openvino>=2023.1.0" numpy "gradio>=4.19"

import gc 
from typing 
import Optional, Union, List, Callable 
import base64 
import tempfile 
import warnings 

import diffusers 
import transformers 
import numpy as np 
import IPython 
import ipywidgets as widgets 
import torch 
import PIL 
import gradio as gr 

import openvino as ov

WIDTH = 432 # 8 で割り切れる必要があります 
HEIGHT = 240 # 8 で割り切れる必要があります 
NUM_FRAMES = 16

pipe = diffusers.DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w")

unet = pipe.unet 
unet.eval() 
vae = pipe.vae 
vae.eval() 
text_encoder = pipe.text_encoder 
text_encoder.eval() 
tokenizer = pipe.tokenizer 
scheduler = pipe.scheduler 
vae_scale_factor = pipe.vae_scale_factor 
unet_in_channels = pipe.unet.config.in_channels 
sample_width = WIDTH // vae_scale_factor 
sample_height = HEIGHT // vae_scale_factor 
del pipe 
gc.collect();

def tensor2vid(video: torch.Tensor, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) -> List[np.ndarray]:
    # このコードは、https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/pipelines/multi_modal/text_to_video_synthesis_pipeline.py#L78 からコピーしたものです。 
    # ncfhw に再形成 
    mean = torch.tensor(mean, device=video.device).reshape(1, -1, 1, 1, 1) 
    std = torch.tensor(std, device=video.device).reshape(1, -1, 1, 1, 1) 
    # [0,1] に非正規化 
    video = video.mul_(std).add_(mean) 
    video.clamp_(0, 1) 
    # 最終出力を準備 
    i, c, f, h, w = video.shape 
    images = video.permute(2, 3, 0, 4, 1).reshape(f, h, i * w, c) # 1 番目 (frames, h, batch_size, w, c) 2 番目 (frames, h, batch_size * w, c) 
    images = images.unbind(dim=0) # 個々の (連続したフレーム) のリストを準備 
    images = [(image.cpu().numpy() * 255).astype("uint8") for image in images] # f h w c 
    return images

try:: 
    from diffusers.utils import randn_tensor 
except ImportError: 
    from diffusers.utils.torch_utils import randn_tensor 

class OVTextToVideoSDPipeline(diffusers.DiffusionPipeline): 
    def __init__( 
        self, 
        vae_decoder: ov.CompiledModel, 
        text_encoder: ov.CompiledModel, 
        tokenizer: transformers.CLIPTokenizer, 
        unet: ov.CompiledModel, 
        scheduler: diffusers.schedulers.DDIMScheduler, 
): 
        super().__init__() 

        self.vae_decoder = vae_decoder 
        self.text_encoder = text_encoder 
        self.tokenizer = tokenizer 
        self.unet = unet 
        self.scheduler = scheduler 
        self.vae_scale_factor = vae_scale_factor 
        self.unet_in_channels = unet_in_channels 
        self.width = WIDTH 
        self.height = HEIGHT 
        self.num_frames = NUM_FRAMES 

    def __call__( 
        self, 
        prompt: Union[str, List[str]] = None, 
        num_inference_steps: int = 50, 
        guidance_scale: float = 9.0, 
        negative_prompt: Optional[Union[str, List[str]]] = None, 
        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] = "np", 
        return_dict: bool = True, 
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, 
        callback_steps: int = 1, 
    ): 
        r""" 
        Function invoked when calling the pipeline for generation.

        Args: 
            prompt (`str` or `List[str]`, *optional*): 
                The prompt or prompts to guide the video generation.If not defined, one has to pass `prompt_embeds`. 
                instead.
            num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps.More denoising steps usually lead to a higher quality videos 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 videos that are closely linked to the text `prompt`, 
                usually at the expense of lower video quality. 
            negative_prompt (`str` or `List[str]`, *optional*): 
                The prompt or prompts not to guide the video generation.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`). 
            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` 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 video 
                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`.Latents should be of shape 
                `(batch_size, num_channel, num_frames, height, width)`. 
            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 `"np"`):  
                The output format of the generate video.Choose between `torch.FloatTensor` or `np.array`. 
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] 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. 
        Returns:
            `List[np.ndarray]`: generated video frames 
        """ 

        num_images_per_prompt = 1 

        # 1. 入力を確認。正しくない場合はエラーを発生 
        self.check_inputs( 
            prompt, 
            callback_steps, 
            negative_prompt, 
            prompt_embeds, 
            negative_prompt_embeds, 
        ) 

        # 2. 呼び出しパラメーターを定義 
        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] 

        # ここで `guidance_scale` は Imagen 論文の式 (2) のガイダンス重み`w`と同様に定義されます: 
        # https://arxiv.org/pdf/2205.11487.pdf`guidance_scale = 1` 
        # は、分類器フリーのガイダンスを行わないことに相当 
        do_classifier_free_guidance = guidance_scale > 1.0 
        # 3. 入力プロンプトをエンコード 
        prompt_embeds = self._encode_prompt( 
            prompt, 
            num_images_per_prompt, 
            do_classifier_free_guidance, 
            negative_prompt, 
            prompt_embeds=prompt_embeds, 
            negative_prompt_embeds=negative_prompt_embeds, 
        ) 

        # 4. タイムステップを準備 
        self.scheduler.set_timesteps(num_inference_steps) 
        timesteps = self.scheduler.timesteps 

        # 5. 潜在変数を準備 
        num_channels_latents = self.unet_in_channels 
        latents = self.prepare_latents( 
            batch_size * num_images_per_prompt, 
            num_channels_latents, 
            prompt_embeds.dtype, 
            generator, 
            latents, 
        ) 

        # 6. 追加のステップ kwargs を準備します。TODO: ロジックは理想的にはパイプラインから外すべきです 
        extra_step_kwargs = {"generator": generator, "eta": eta} 

        # 7. ノイズ除去ループ 
        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):
                # 分類器フリーのガイダンスを行う場合は潜在変数を拡張 
                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents 
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) 

                # ノイズ残留を予測 
                noise_pred = self.unet( 
                    { 
                        "sample": latent_model_input, 
                        "timestep": t, 
                        "encoder_hidden_states": prompt_embeds, 
                    } 
                )[0] 
                noise_pred = torch.tensor(noise_pred) 

                # ガイダンスを実行 
                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) 

                # 潜在性を再形成 
                bsz, channel, frames, width, height = latents.shape 
                latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) 
                noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) 

                # 前のノイズサンプルを計算 x_t -> x_t-1 
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample 

                # 潜在性を再形成 
                latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4) 

                # コールバックが提供されている場合はそれを呼び出す 
                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) 

        video_tensor = self.decode_latents(latents) 

        if output_type == "pt": 
            video = video_tensor 
        else: 
            video = tensor2vid(video_tensor) 

        if not return_dict: 
            return (video,) 
        return {"frames": video} 

    # diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt からコピー 
    def _encode_prompt( 
        self, 
        prompt, 
        num_images_per_prompt, 
        do_classifier_free_guidance, 
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None, 
        negative_prompt_embeds: Optional[torch.FloatTensor] = None, 
    ): 
        r""" 
        Encodes the prompt into text encoder hidden states.

        Args: 
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded 
            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.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] 

        if 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]) 
                print( 
                    "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}" 
                ) 

            prompt_embeds = self.text_encoder(text_input_ids) 
            prompt_embeds = prompt_embeds[0] 
            prompt_embeds = torch.tensor(prompt_embeds) 

        bs_embed, seq_len, _ = prompt_embeds.shape 
        # mps フレンドリーな方法を使用して、プロンプトごとに各世代のテキスト埋め込みを複製 
        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) 

        # 分類器の無条件埋め込みを取得するフリーガイダンス 
        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", 
            ) 

            negative_prompt_embeds = self.text_encoder(uncond_input.input_ids) 
            negative_prompt_embeds = negative_prompt_embeds[0] 
            negative_prompt_embeds = torch.tensor(negative_prompt_embeds)
 
        if do_classifier_free_guidance:
            # mps フレンドリーな方法を使用して、プロンプトごとに各世代の無条件埋め込みを複製 
            seq_len = negative_prompt_embeds.shape[1] 

            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) 

            # 分類器フリーのガイダンスでは、2 回のフォワードパスを実行する必要があります
            # ここでは、無条件埋め込みとテキスト埋め込みを 1 つのバッチに連結して、 
            # 2 回のフォワードパスの実行を回避します 
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) 

        return prompt_embeds 

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

        # スケジューラーが要求する標準偏差で初期ノイズをスケール 
        latents = latents * self.scheduler.init_noise_sigma 
        return latents 

    def check_inputs( 
        self, 
        prompt, 
        callback_steps, 
        negative_prompt=None, 
        prompt_embeds=None, 
        negative_prompt_embeds=None, 
    ): 
        if self.height % 8 != 0 or self.width % 8 != 0: 
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {self.height} and {self.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 decode_latents(self, latents): 
        scale_factor = 0.18215 
        latents = 1 / scale_factor * latents 

        batch_size, channels, num_frames, height, width = latents.shape 
        latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) 
        image = self.vae_decoder(latents)[0] 
        image = torch.tensor(image) 
        video = ( 
            image[None, :]
            .reshape( 
                ( 
                    batch_size, 
                    num_frames, 
                    -1, 
                ) 
                + image.shape[2:]
            ) 
            .permute(0, 2, 1, 3, 4) 
        ) 
        # 常に float32 にキャストします。これは大きなオーバーヘッドを発生せず、bfloat16 と互換性があるためです 
        video = video.float() 
        return video

core = ov.Core()

def generate(prompt, seed, num_inference_steps, _=gr.Progress(track_tqdm=True)): 
    generator = torch.Generator().manual_seed(seed) 
    frames = ov_pipe( 
        prompt, 
        num_inference_steps=num_inference_steps, 
        generator=generator, 
    )["frames"] 
    out_file = tempfile.NamedTemporaryFile(suffix=".gif", delete=False) 
    images = [PIL.Image.fromarray(frame) for frame in frames] 
    images[0].save(out_file, save_all=True, append_images=images[1:], duration=125, loop=0) 
    return out_file.name 

demo = gr.Interface( 
    generate, 
    [ 
        gr.Textbox(label="Prompt"), 
        gr.Slider(0, 1000000, value=42, label="Seed", step=1), 
        gr.Slider(10, 50, value=25, label="Number of inference steps", step=1), 
    ], 
    gr.Image(label="Result"), 
    examples=[ 
        ["An astronaut riding a horse.", 0, 25], 
        ["A panda eating bamboo on a rock.", 0, 25], 
        ["Spiderman is surfing.", 0, 25],
    ], 
    allow_flagging="never", 
) 

try: 
    demo.queue().launch(debug=False) 
except Exception: 
    demo.queue().launch(share=True, debug=False) 
# リモートで起動する場合は、server_name と server_port を指定 
# demo.launch(server_name='your server name', server_port='server port in int') 
# 詳細はドキュメントをご覧ください: https://gradio.app/docs/