inference.py

from __future__ import annotations

import gc
import pathlib
import sys
import tempfile
import os
import gradio as gr
import imageio
import PIL.Image
import torch
from diffusers.utils.import_utils import is_xformers_available
from einops import rearrange
from huggingface_hub import ModelCard
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection, CLIPTextModelWithProjection
from diffusers import AutoencoderKL, DDPMScheduler, DDIMScheduler, PNDMScheduler, ControlNetModel, PriorTransformer, UnCLIPScheduler
from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
from omegaconf import OmegaConf
from typing import Any, Callable, Dict, List, Optional, Union, Tuple

sys.path.append('Make-A-Protagonist')

from makeaprotagonist.models.unet import UNet3DConditionModel
from makeaprotagonist.pipelines.pipeline_stable_unclip_controlavideo import MakeAProtagonistStableUnCLIPPipeline, MultiControlNetModel
from makeaprotagonist.dataset.dataset import MakeAProtagonistDataset
from makeaprotagonist.util import save_videos_grid, ddim_inversion_unclip, ddim_inversion_prior
from experts.grounded_sam_mask_out import mask_out_reference_image


import ipdb

class InferencePipeline:
    def __init__(self, hf_token: str | None = None):
        self.hf_token = hf_token
        self.pipe = None
        self.device = torch.device(
            'cuda:0' if torch.cuda.is_available() else 'cpu')
        self.model_id = None

        self.conditions = None
        self.masks = None
        self.ddim_inv_latent = None
        self.train_dataset, self.sample_indices = None, None

    def clear(self) -> None:
        self.model_id = None
        del self.pipe
        self.pipe = None
        self.conditions = None
        self.masks = None
        self.ddim_inv_latent = None
        self.train_dataset, self.sample_indices = None, None

        torch.cuda.empty_cache()
        gc.collect()

    @staticmethod
    def check_if_model_is_local(model_id: str) -> bool:
        return pathlib.Path(model_id).exists()

    @staticmethod
    def get_model_card(model_id: str,
                       hf_token: str | None = None) -> ModelCard:
        if InferencePipeline.check_if_model_is_local(model_id):
            card_path = (pathlib.Path(model_id) / 'README.md').as_posix()
        else:
            card_path = model_id
        return ModelCard.load(card_path, token=hf_token)

    @staticmethod
    def get_base_model_info(model_id: str, hf_token: str | None = None) -> str:
        card = InferencePipeline.get_model_card(model_id, hf_token)
        return card.data.base_model

    @torch.no_grad()
    def load_pipe(self, model_id: str, n_steps, seed) -> None:
        if model_id == self.model_id:
            return self.conditions, self.masks, self.ddim_inv_latent, self.train_dataset, self.sample_indices

        base_model_id = self.get_base_model_info(model_id, self.hf_token)
        
        pretrained_model_path = 'stabilityai/stable-diffusion-2-1-unclip-small'
        # image encoding components
        feature_extractor = CLIPImageProcessor.from_pretrained(pretrained_model_path, subfolder="feature_extractor")
        image_encoder = CLIPVisionModelWithProjection.from_pretrained(pretrained_model_path, subfolder="image_encoder")
        # image noising components
        image_normalizer = StableUnCLIPImageNormalizer.from_pretrained(pretrained_model_path, subfolder="image_normalizer", torch_dtype=torch.float16,)
        image_noising_scheduler = DDPMScheduler.from_pretrained(pretrained_model_path, subfolder="image_noising_scheduler")
        # regular denoising components
        tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
        text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder", torch_dtype=torch.float16,)
        vae = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae", torch_dtype=torch.float16,)
        self.ddim_inv_scheduler = DDIMScheduler.from_pretrained(pretrained_model_path, subfolder='scheduler')
        self.ddim_inv_scheduler.set_timesteps(n_steps)

        prior_model_id = "kakaobrain/karlo-v1-alpha"
        data_type = torch.float16
        prior = PriorTransformer.from_pretrained(prior_model_id, subfolder="prior", torch_dtype=data_type)

        prior_text_model_id = "openai/clip-vit-large-patch14"
        prior_tokenizer = CLIPTokenizer.from_pretrained(prior_text_model_id)
        prior_text_model = CLIPTextModelWithProjection.from_pretrained(prior_text_model_id, torch_dtype=data_type)
        prior_scheduler = UnCLIPScheduler.from_pretrained(prior_model_id, subfolder="prior_scheduler")
        prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config)

        controlnet_model_id = ['controlnet-2-1-unclip-small-openposefull', 'controlnet-2-1-unclip-small-depth']
        controlnet = MultiControlNetModel( [ControlNetModel.from_pretrained('Make-A-Protagonist/controlnet-2-1-unclip-small', subfolder=subfolder_id, torch_dtype=torch.float16) for subfolder_id in controlnet_model_id] )

        unet = UNet3DConditionModel.from_pretrained(
            model_id,
            subfolder='unet',
            torch_dtype=torch.float16,
            use_auth_token=self.hf_token)

        # Freeze vae and text_encoder and adapter
        vae.requires_grad_(False)
        text_encoder.requires_grad_(False)

        ## freeze image embed
        image_encoder.requires_grad_(False)

        unet.requires_grad_(False)
        ## freeze controlnet
        controlnet.requires_grad_(False)

        ## freeze prior
        prior.requires_grad_(False)
        prior_text_model.requires_grad_(False)

        config_file = os.path.join('Make-A-Protagonist/configs', model_id.split('/')[-1] + '.yaml')
        self.cfg = OmegaConf.load(config_file)

        # def source_parsing(self, n_steps):
        # ipdb.set_trace()
        train_dataset = MakeAProtagonistDataset(**self.cfg)
        train_dataset.preprocess_img_embedding(feature_extractor, image_encoder)
        train_dataloader = torch.utils.data.DataLoader(
        train_dataset, batch_size=1, num_workers=0, 
        )
        image_encoder.to(dtype=data_type)
        pipe = MakeAProtagonistStableUnCLIPPipeline(
                prior_tokenizer=prior_tokenizer,
                prior_text_encoder=prior_text_model,
                prior=prior,
                prior_scheduler=prior_scheduler,
                feature_extractor=feature_extractor,
                image_encoder=image_encoder,
                image_normalizer=image_normalizer,
                image_noising_scheduler=image_noising_scheduler,
                vae=vae, 
                text_encoder=text_encoder, 
                tokenizer=tokenizer, 
                unet=unet,
                controlnet=controlnet,
                scheduler=DDIMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
            )

        pipe = pipe.to(self.device)
        
        if is_xformers_available():
            pipe.unet.enable_xformers_memory_efficient_attention()
            pipe.controlnet.enable_xformers_memory_efficient_attention()
        self.pipe = pipe
        self.model_id = model_id  # type: ignore
        self.vae = vae
        # self.feature_extractor = feature_extractor
        # self.image_encoder = image_encoder
        ## ddim inverse for source video

        batch = next(iter(train_dataloader))
        weight_dtype = torch.float16
        pixel_values = batch["pixel_values"].to(weight_dtype).to(self.device)
        video_length = pixel_values.shape[1]
        pixel_values = rearrange(pixel_values, "b f c h w -> (b f) c h w")
        latents = self.vae.encode(pixel_values).latent_dist.sample()
        latents = rearrange(latents, "(b f) c h w -> b c f h w", f=video_length)
        latents = latents * self.vae.config.scaling_factor
        # ControlNet
        # ipdb.set_trace()
        conditions = [_condition.to(weight_dtype).to(self.device) for _, _condition in batch["conditions"].items()] # b f c h w
        masks = batch["masks"].to(weight_dtype).to(self.device) # b,f,1,h,w
        emb_dim = train_dataset.img_embeddings[0].size(0)
        key_frame_embed = torch.zeros((1, emb_dim)).to(device=latents.device, dtype=latents.dtype) ## this is dim 0
        # ipdb.set_trace()
        ddim_inv_latent = ddim_inversion_unclip(
            self.pipe, self.ddim_inv_scheduler, video_latent=latents,
            num_inv_steps=n_steps, prompt="", image_embed=key_frame_embed, noise_level=0, seed=seed)[-1].to(weight_dtype)
        self.conditions = conditions
        self.masks = masks
        self.ddim_inv_latent = ddim_inv_latent
        self.train_dataset = train_dataset
        self.sample_indices = batch["sample_indices"][0]
        return conditions, masks, ddim_inv_latent, train_dataset, batch["sample_indices"][0]
                
    def run(
        self,
        model_id: str,
        prompt: str,
        video_length: int,
        fps: int,
        seed: int,
        n_steps: int,
        guidance_scale: float,
        ref_image: PIL.Image.Image,
        ref_pro_prompt: str,
        noise_level: int,
        start_step: int, 
        control_pose: float,
        control_depth: float,
        source_pro: int = 0, # 0 or 1
        source_bg: int = 0,
    ) -> PIL.Image.Image:

        if not torch.cuda.is_available():
            raise gr.Error('CUDA is not available.')

        torch.cuda.empty_cache()
        
        conditions, masks, ddim_inv_latent, _, _ = self.load_pipe(model_id, n_steps, seed)
        ## conditions [1,F,3,H,W]
        ## masks [1,F,1,H,W]
        ## ddim_inv_latent [1,4,F,H,W]
        ## NOTE this is to deal with video length
        conditions = [_condition[:,:video_length] for _condition in conditions]
        masks = masks[:, :video_length]
        ddim_inv_latent = ddim_inv_latent[:,:,:video_length]

        generator = torch.Generator(device=self.device).manual_seed(seed)

        ## TODO mask out reference image
        # ipdb.set_trace()
        ref_image = mask_out_reference_image(ref_image, ref_pro_prompt)
        controlnet_conditioning_scale = [control_pose, control_depth]

        prior_denoised_embeds = None
        image_embed = None
        if source_bg:
            ## using source background and changing the protagonist
            prior_denoised_embeds = self.train_dataset.img_embeddings[0][None].to(device=ddim_inv_latent.device, dtype=ddim_inv_latent.dtype) # 1, 768 for UnCLIP-small
        
        if source_pro:
            # using source protagonist and changing the background
            sample_indices = self.sample_indices
            image_embed = [self.train_dataset.img_embeddings[idx] for idx in sample_indices]
            image_embed = torch.stack(image_embed, dim=0).to(device=ddim_inv_latent.device, dtype=ddim_inv_latent.dtype) # F, 768 for UnCLIP-small # F,C
            image_embed = image_embed[:video_length]
            ref_image = None

        # ipdb.set_trace()
        out = self.pipe(
            image=ref_image, 
            prompt=prompt, 
            control_image=conditions, 
            video_length=video_length,
            width=768,
            height=768,
            num_inference_steps=n_steps,
            guidance_scale=guidance_scale,
            generator=generator,
            ## ddim inversion
            latents=ddim_inv_latent, 
            ## ref image embeds 
            noise_level=noise_level, 
            ## controlnet
            controlnet_conditioning_scale=controlnet_conditioning_scale,
            ## mask
            masks=masks, 
            mask_mode='all',
            mask_latent_fuse_mode = 'all',
            start_step=start_step,
            ## edit bg and pro
            prior_latents=None, 
            image_embeds=image_embed, # keep pro
            prior_denoised_embeds=prior_denoised_embeds # keep bg
            )

        frames = rearrange(out.videos[0], 'c t h w -> t h w c')
        frames = (frames * 255).to(torch.uint8).numpy()

        out_file = tempfile.NamedTemporaryFile(suffix='.mp4', delete=False)
        writer = imageio.get_writer(out_file.name, fps=fps)
        for frame in frames:
            writer.append_data(frame)
        writer.close()

        return out_file.name