handler.py

from typing import Dict, Any
import torch
from PIL import Image
import base64
from io import BytesIO
import numpy as np
from diffusers import AutoencoderKL, DDIMScheduler
from einops import repeat
from omegaconf import OmegaConf
from transformers import CLIPVisionModelWithProjection
import cv2
import os
import sys
import skvideo.io
from src.models.pose_guider import PoseGuider
from src.models.unet_2d_condition import UNet2DConditionModel
from src.models.unet_3d import UNet3DConditionModel
from src.pipelines.pipeline_pose2vid_long import Pose2VideoPipeline
from src.utils.util import read_frames, get_fps, save_videos_grid
import roop.globals
from roop.core import start, decode_execution_providers, suggest_max_memory, suggest_execution_threads
from roop.utilities import normalize_output_path
from roop.processors.frame.core import get_frame_processors_modules

import onnxruntime as ort
import gc
import subprocess

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

if device.type != 'cuda':
    raise ValueError("The model requires a GPU for inference.")

class EndpointHandler():
    def __init__(self, path=""):
        base_dir = os.path.dirname(os.path.abspath(__file__))
        config_path = os.path.join(base_dir, 'configs', 'prompts', 'animation.yaml')

        if not os.path.exists(config_path):
            raise FileNotFoundError(f"The configuration file was not found at: {config_path}")

        self.config = OmegaConf.load(config_path)
        self.weight_dtype = torch.float16
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.pipeline = None
        self._initialize_pipeline()

    def _initialize_pipeline(self):
        base_dir = os.path.dirname(os.path.abspath(__file__))
        config_path = os.path.join(base_dir, 'pretrained_weights', 'sd-vae-ft-mse')

        if not os.path.exists(config_path):
            raise FileNotFoundError(f"The sd-vae-ft-mse folder was not found at: {config_path}")

        vae = AutoencoderKL.from_pretrained(config_path).to(self.device, dtype=self.weight_dtype)

        pretrained_base_model_path_unet = os.path.join(base_dir, 'pretrained_weights', 'stable-diffusion-v1-5', 'unet')
        print("model path is " + pretrained_base_model_path_unet)
        reference_unet = UNet2DConditionModel.from_pretrained(
            pretrained_base_model_path_unet
        ).to(dtype=self.weight_dtype, device=self.device)

        inference_config_path = os.path.join(base_dir, 'configs', 'inference', 'inference_v2.yaml')
        motion_module_path = os.path.join(base_dir, 'pretrained_weights', 'motion_module.pth')
        denoising_unet_path = os.path.join(base_dir, 'pretrained_weights', 'denoising_unet.pth')
        reference_unet_path = os.path.join(base_dir, 'pretrained_weights', 'reference_unet.pth')
        pose_guider_path = os.path.join(base_dir, 'pretrained_weights', 'pose_guider.pth')
        image_encoder_path = os.path.join(base_dir, 'pretrained_weights', 'image_encoder')

        infer_config = OmegaConf.load(inference_config_path)
        denoising_unet = UNet3DConditionModel.from_pretrained_2d(
            pretrained_base_model_path_unet,
            motion_module_path,
            unet_additional_kwargs=infer_config.unet_additional_kwargs,
        ).to(self.device, dtype=self.weight_dtype)

        pose_guider = PoseGuider(320, block_out_channels=(16, 32, 96, 256)).to(self.device, dtype=self.weight_dtype)
        image_enc = CLIPVisionModelWithProjection.from_pretrained(image_encoder_path).to(self.device, dtype=self.weight_dtype)
        sched_kwargs = OmegaConf.to_container(infer_config.noise_scheduler_kwargs)
        scheduler = DDIMScheduler(**sched_kwargs)

        denoising_unet.load_state_dict(torch.load(denoising_unet_path, map_location="cpu"), strict=False)
        reference_unet.load_state_dict(torch.load(reference_unet_path, map_location="cpu"))
        pose_guider.load_state_dict(torch.load(pose_guider_path, map_location="cpu"))

        self.pipeline = Pose2VideoPipeline(
            vae=vae,
            image_encoder=image_enc,
            reference_unet=reference_unet,
            denoising_unet=denoising_unet,
            pose_guider=pose_guider,
            scheduler=scheduler
        ).to(self.device, dtype=self.weight_dtype)

    def _crop_face(self, image, save_path="cropped_face.jpg", margin=0.5):
        # Convert image to OpenCV format
        cv_image = np.array(image)
        cv_image = cv_image[:, :, ::-1].copy()

        # Load OpenCV face detector
        face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')

        # Detect faces
        gray = cv2.cvtColor(cv_image, cv2.COLOR_BGR2GRAY)
        faces = face_cascade.detectMultiScale(gray, 1.1, 4)

        if len(faces) == 0:
            raise ValueError("No faces detected in the reference image.")

        # Crop the first face found with a margin
        x, y, w, h = faces[0]
        x_margin = int(margin * w)
        y_margin = int(margin * h)

        x1 = max(0, x - x_margin)
        y1 = max(0, y - y_margin // 2)  # Less margin at the top
        x2 = min(cv_image.shape[1], x + w + x_margin)
        y2 = min(cv_image.shape[0], y + h + y_margin)  # More margin at the bottom

        cropped_face = cv_image[y1:y2, x1:x2]

        # Convert back to PIL format
        cropped_face = Image.fromarray(cropped_face[:, :, ::-1]).convert("RGB")

        # Save the cropped face
        cropped_face.save(save_path, format="JPEG", quality=95)

        return cropped_face

    def _swap_face(self, source_image, target_video_path):
        source_path = "input.jpg"
        source_image.save(source_path, format="JPEG", quality=95)
        output_path = "output.mp4"

        roop.globals.source_path = source_path
        roop.globals.target_path = target_video_path
        roop.globals.output_path = normalize_output_path(roop.globals.source_path, roop.globals.target_path, output_path)
        roop.globals.frame_processors = ["face_swapper", "face_enhancer"]
        roop.globals.headless = True
        roop.globals.keep_fps = True
        roop.globals.keep_audio = True
        roop.globals.keep_frames = False
        roop.globals.many_faces = False
        roop.globals.video_encoder = "libx264"
        roop.globals.video_quality = 100
        roop.globals.max_memory = suggest_max_memory()

        # Set GPU execution provider
        roop.globals.execution_providers = decode_execution_providers(["CUDAExecutionProvider"])
        roop.globals.execution_threads = suggest_execution_threads()

        # Ensure onnxruntime is using the GPU
        ort.set_default_logger_severity(3)  # Suppress verbose logging
        providers = ['CUDAExecutionProvider']
        options = ort.SessionOptions()
        options.intra_op_num_threads = 1

        for frame_processor in get_frame_processors_modules(roop.globals.frame_processors):
            if hasattr(frame_processor, 'onnx_session'):
                frame_processor.onnx_session.set_providers(providers, options)

        # Clear CUDA cache before starting the face swapping process
        torch.cuda.empty_cache()

        start()

        # Clear CUDA cache after the face swapping process
        for frame_processor in roop.globals.frame_processors:
            del frame_processor
        torch.cuda.empty_cache()

        return os.path.join(os.getcwd(), output_path)

    def print_memory_stat_for_stuff(self, phase, log_file="memory_stats.log"):
        with open(log_file, "a") as f:
            f.write(f"Memory Stats - {phase}:\n")
            f.write(f"Allocated memory: {torch.cuda.memory_allocated() / 1024**2:.2f} MB\n")
            f.write(f"Reserved memory: {torch.cuda.memory_reserved() / 1024**2:.2f} MB\n")
            f.write(f"Max allocated memory: {torch.cuda.max_memory_allocated() / 1024**2:.2f} MB\n")
            f.write(f"Max reserved memory: {torch.cuda.max_memory_reserved() / 1024**2:.2f} MB\n")
            f.write("="*30 + "\n")

    def convert_to_playable_format(self, input_path, output_path):
        command = [
            "ffmpeg",
            "-i", input_path,
            "-c:v", "libx264",
            "-preset", "fast",
            "-crf", "18",
            "-y",  # Overwrite output file if it exists
            output_path
        ]
        result = subprocess.run(command, capture_output=True, text=True)
        print("Conversion STDOUT:", result.stdout)
        print("Conversion STDERR:", result.stderr)
        
        if result.returncode != 0:
            raise RuntimeError(f"FFmpeg conversion failed with exit code {result.returncode}")

    def run_rife_interpolation(self, video_path, output_path, multi=2, scale=1.0):
        base_dir = os.path.dirname(os.path.abspath(__file__))
        directory = os.path.join(base_dir, "Practical-RIFE", "inference_video.py")
        model_directory = os.path.join(base_dir, "Practical-RIFE", "train_log")
        command = [
            "python",
            directory,
            f"--video={video_path}",
            f"--output={output_path}",
            f"--multi={multi}",
            f"--scale={scale}",
            f"--model={model_directory}",
        ]
        
        result = subprocess.run(command, capture_output=True, text=True)
        print(result)
        print(result.stdout)
        print(result.stderr)
        
        if result.returncode != 0:
            raise RuntimeError(f"RIFE interpolation failed with exit code {result.returncode}")
        self.convert_to_playable_format(output_path, "completed_playable.mp4")

    def speed_up_video(self, input_path, output_path, factor=4):
        command = [
            "ffmpeg",
            "-i", input_path,
            "-filter:v", f"setpts=PTS/{factor}",
            "-an",  # Remove audio
            output_path
        ]
        result = subprocess.run(command, capture_output=True, text=True)
        print("Speed Up Video STDOUT:", result.stdout)
        print("Speed Up Video STDERR:", result.stderr)
        
        if result.returncode != 0:
            raise RuntimeError(f"FFmpeg speed up failed with exit code {result.returncode}")

    def slow_down_video(self, input_path, output_path, factor=4):
        command = [
            "ffmpeg",
            "-i", input_path,
            "-filter:v", f"setpts={factor}*PTS",
            "-an",  # Remove audio
            output_path
        ]
        result = subprocess.run(command, capture_output=True, text=True)
        print("Slow Down Video STDOUT:", result.stdout)
        print("Slow Down Video STDERR:", result.stderr)
        
        if result.returncode != 0:
            raise RuntimeError(f"FFmpeg slow down failed with exit code {result.returncode}")

    def __call__(self, data: Any) -> Dict[str, str]:
        inputs = data.get("inputs", {})
        ref_image_base64 = inputs.get("ref_image", "")
        pose_video_path = inputs.get("pose_video_path", "")
        width = inputs.get("width", 512)
        height = inputs.get("height", 768)
        length = inputs.get("length", 24)
        num_inference_steps = inputs.get("num_inference_steps", 25)
        cfg = inputs.get("cfg", 3.5)
        seed = inputs.get("seed", 123)

        ref_image = Image.open(BytesIO(base64.b64decode(ref_image_base64)))

        # Get the base directory of the current file
        base_dir = os.path.dirname(os.path.abspath(__file__))
        
        # Update pose_video_path to use the base directory
        pose_video_path = os.path.join(base_dir, pose_video_path)
        
        if not os.path.exists(pose_video_path):
            raise FileNotFoundError(f"The pose video was not found at: {pose_video_path}")

        # Speed up the pose video by 4x
        sped_up_pose_video_path = os.path.join(base_dir, "sped_up_pose_video.mp4")
        self.speed_up_video(pose_video_path, sped_up_pose_video_path, factor=4)

        torch.manual_seed(seed)
        pose_images = read_frames(sped_up_pose_video_path)
        src_fps = get_fps(sped_up_pose_video_path)
        
        pose_list = []
        total_length = min(length, len(pose_images))
        for pose_image_pil in pose_images[:total_length]:
            pose_list.append(pose_image_pil)

        video = self.pipeline(
            ref_image,
            pose_list,
            width=width,
            height=height,
            video_length=total_length,
            num_inference_steps=num_inference_steps,
            guidance_scale=cfg
        ).videos

        save_dir = os.path.join(base_dir, "output", "gradio")
        if not os.path.exists(save_dir):
            os.makedirs(save_dir, exist_ok=True)
        animation_path = os.path.join(save_dir, "animation_output.mp4")
        save_videos_grid(video, animation_path, n_rows=1, fps=src_fps)

        # Crop the face from the reference image and save it
        cropped_face_path = os.path.join(save_dir, "cropped_face.jpg")
        cropped_face = self._crop_face(ref_image, save_path=cropped_face_path)

        # Delete the pipeline and clear CUDA cache to free up memory
        del self.pipeline
        torch.cuda.empty_cache()

        # Perform face swapping
        swapped_face_video_path = self._swap_face(cropped_face, animation_path)

        # Slow down the produced video by 4x
        slowed_down_animation_path = os.path.join(save_dir, "slowed_down_animation_output.mp4")
        self.slow_down_video(swapped_face_video_path, slowed_down_animation_path, factor=4)

        # Clear CUDA cache before RIFE interpolation
        torch.cuda.empty_cache()

        # Perform RIFE interpolation
        rife_output_path = os.path.join(save_dir, "completed_result.mp4")
        self.run_rife_interpolation(slowed_down_animation_path, rife_output_path, multi=2, scale=0.5)

        # Encode the final video in base64
        with open(rife_output_path, "rb") as video_file:
            video_base64 = base64.b64encode(video_file.read()).decode("utf-8")

        torch.cuda.empty_cache()

        return {"video": video_base64}