import torch
from omegaconf import OmegaConf
from safetensors.torch import load_model
import numpy as np
import os
from diffusers.models import AutoencoderKL  
from pipeline.whisper import WhisperAudioProcessor
from pipeline.dwpose import DWposeDetector
from pipeline.utils import A2MEvalDataset
from pipeline.a2m_pipeline import A2M_Pipeline
from model.model_A2M import A2MModel_CrossAtten_Audio_PosePre , A2MModel_CrossAtten_Audio
from model.model_AMD import AMDModel
from typing import Optional
from torch.utils.data import DataLoader
from omegaconf import OmegaConf
class a2m_inferencer:
    def __init__(
        self,
        config,
        device,
        dtype
    ):
        self.config = config
        self.device = device
        self.dtype = dtype
        self.setup()
    def setup(self):
        vae_model = AutoencoderKL.from_pretrained(self.config.vae_path, subfolder="vae").to(self.device, self.dtype).requires_grad_(False)
        a2m_config = OmegaConf.load(self.config.a2m_config_path)
        if self.config.enable_pose:
            a2m_model = A2MModel_CrossAtten_Audio_PosePre(**a2m_config['model']).to(self.device, self.dtype).requires_grad_(False)
        else:
            a2m_model = A2MModel_CrossAtten_Audio(**a2m_config["model"]).to(self.device, self.dtype).requires_grad_(False)
        load_model(a2m_model, self.config.a2m_ckpt_path)
        amd_model = AMDModel.from_config(AMDModel.load_config(self.config.amd_config_path)).to(self.device, self.dtype).requires_grad_(False)
        load_model(amd_model, self.config.amd_ckpt_path)
        dwpose_model = DWposeDetector().to(self.device)
        whisper_model = WhisperAudioProcessor(
            16000, 30,
            self.config.whisper_model_path,
            os.path.dirname(self.config.audio_separator_model_file),
            os.path.basename(self.config.audio_separator_model_file),
            cache_dir = self.config.cache_dir, 
            device = self.device)
        self.pipeline = A2M_Pipeline(
            amd_model,
            a2m_model,
            vae_model,
            dwpose_model,
            whisper_model,
            amd_sample_steps=self.config.amd_sample_steps,
            a2m_sample_steps=self.config.a2m_sample_steps,
            output_dir=self.config.output_dir,
            enable_pose=self.config.enable_pose
        )
    def infer(self, audio_path:str, refimg_path:str):
        video = self.pipeline.run(audio_path, refimg_path)
        return video
    def eval(self, audio_emb_dir:str, dwpose_dir:str, ref_img_dir:str, num_frames:int = 96, audio_dir:Optional[str]=None):
        evalset = A2MEvalDataset(
        audio_emb_dir,
        dwpose_dir,
        ref_img_dir,
        num_frames,
        random_audio=False,
        random_dwpose=False,
        audio_dir=audio_dir,
        # num_evals=4,
        # audio_suffix="wav"
        )
        evalloader = DataLoader(
            evalset, 12, shuffle=False,drop_last=True,collate_fn=evalset.collate,num_workers=16
        )
        self.pipeline.eval(evalloader)

if __name__ == "__main__":
    # TODO add argparse here
    # config_path = "/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/config/inference/a2m_wpose.yaml"
    config_path = "/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/config/inference/a2m.yaml"
    audio_path = "/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/demo/audio21.wav"
    refimg_path = "/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/demo/face36.jpg"
    audio_emb_dir = "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/whisper_embs"
    dwpose_dir = "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/firstframes/fromdwpose"
    ref_img_dir = "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/firstframes/fromvideo"
    audio_dir = "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/audios"
    config = OmegaConf.load(config_path)
    inferencer = a2m_inferencer(config, torch.device("cuda:0"), torch.float16)
    # inferencer.infer(audio_path, refimg_path)
    inferencer.eval(
        audio_emb_dir,
        dwpose_dir,
        ref_img_dir,
        96,
        audio_dir
    )