108 lines
3.1 KiB
Python
108 lines
3.1 KiB
Python
# -*- encoding: utf-8 -*-
|
||
'''
|
||
@Email : liaoxiju@inspur.com
|
||
'''
|
||
import os
|
||
import re
|
||
import sys
|
||
import uvicorn
|
||
from fastapi import FastAPI, Request
|
||
from modelscope.pipelines import pipeline
|
||
import requests
|
||
import tempfile
|
||
import torch
|
||
import time
|
||
|
||
if torch.cuda.is_available():
|
||
torch.cuda.set_device(0)
|
||
|
||
app = FastAPI()
|
||
|
||
model_path = os.getenv('MODEL_PATH','./wwd123/sadtalker')
|
||
|
||
inference = pipeline('talking-head', model=model_path, model_revision='v1.0.0', device='cuda', use_gpu=True, cache="./")
|
||
|
||
def upload_to_s3(filepath):
|
||
timestap = int(time.time())
|
||
os.system(f's3cmd put --expiry-days 10 {filepath} -r s3://ihp/tmp/liao/service/sadtalker/results/{timestap}.mp4')
|
||
return f"results/{timestap}.mp4"
|
||
|
||
def download_wav(wav_url):
|
||
try:
|
||
if not os.path.exists("results"):
|
||
os.system("mkdir -p results")
|
||
res = requests.get(wav_url)
|
||
content = res.content
|
||
tmp_dir = tempfile.TemporaryDirectory(dir="results").name
|
||
if not os.path.exists(tmp_dir):
|
||
os.system(f"mkdir -p {tmp_dir}")
|
||
wav_path = os.path.join(tmp_dir, "input_audio.wav")
|
||
fout = open(wav_path, "wb")
|
||
fout.write(content)
|
||
fout.close()
|
||
return wav_path, tmp_dir
|
||
except:
|
||
import traceback
|
||
print("download error", traceback.format_exc())
|
||
return None
|
||
|
||
#@app.post("/hairuo/audiodrivenvido")
|
||
@app.post("/agentstore/api/v1/multimodal_models/dh/dighthuman")
|
||
async def digithuman(request: Request, req: dict):
|
||
'''
|
||
获取音频url-wav音频，然后使用sadtalker
|
||
'''
|
||
body_data = await request.body()
|
||
body_data = body_data.decode("utf-8")
|
||
try:
|
||
timestamp = int(time.time())
|
||
save_path = f"results_{timestamp}"
|
||
if not os.path.exists(save_path):
|
||
os.system(f"mkdir -p {save_path}")
|
||
|
||
audio_url = req.get("audio_url")
|
||
wav_path, tmp_dir = download_wav(audio_url)
|
||
source_image = 'nv.jpg'
|
||
|
||
kwargs = {
|
||
'preprocess' : 'crop', # 'crop', 'resize', 'full'
|
||
'still_mode' : True,
|
||
'use_enhancer' : False,
|
||
'batch_size' : 1,
|
||
'size' : 256, # 256, 512
|
||
'pose_style' : 0,
|
||
'exp_scale' : 1,
|
||
'result_dir': save_path
|
||
}
|
||
|
||
video_path = inference(source_image, driven_audio=wav_path, **kwargs)
|
||
|
||
video_name = upload_to_s3(video_path)
|
||
|
||
##清除临时文件
|
||
os.system(f"rm -rf {tmp_dir}")
|
||
os.system(f"rm {video_path}")
|
||
os.system(f"rm -rf {save_path}")
|
||
|
||
video_url = os.path.join("https://ihp.oss.cn-north-4.inspurcloudoss.com/tmp/liao/service/sadtalker/", video_name)
|
||
resp = {
|
||
"code": "0",
|
||
"message": "success",
|
||
"result": video_url
|
||
}
|
||
return resp
|
||
|
||
except:
|
||
import traceback
|
||
print(traceback.format_exc())
|
||
resp = {
|
||
"code": "1",
|
||
"message": "audiodrivenvido model call error",
|
||
"result": ""
|
||
}
|
||
return resp
|
||
|
||
|
||
if __name__ == "__main__":
|
||
uvicorn.run(app, host="0.0.0.0", port=14041)
|