DeepFaceLive/modelhub/onnx/InsightFaceSwap/InsightFaceSwap.py
2023-07-09 22:01:38 +04:00

94 lines
3.1 KiB
Python

from pathlib import Path
from typing import List
import numpy as np
from xlib.file import SplittedFile
from xlib.image import ImageProcessor
from xlib.onnxruntime import (InferenceSession_with_device, ORTDeviceInfo,
get_available_devices_info)
import cv2
import onnx
from onnx import numpy_helper
class InsightFaceSwap:
"""
arguments
device_info ORTDeviceInfo
use LIA.get_available_devices()
to determine a list of avaliable devices accepted by model
raises
Exception
"""
@staticmethod
def get_available_devices() -> List[ORTDeviceInfo]:
return get_available_devices_info()
def __init__(self, device_info : ORTDeviceInfo):
if device_info not in InsightFaceSwap.get_available_devices():
raise Exception(f'device_info {device_info} is not in available devices for InsightFaceSwap')
inswapper_path = Path(__file__).parent / 'inswapper_128.onnx'
SplittedFile.merge(inswapper_path, delete_parts=False)
if not inswapper_path.exists():
raise FileNotFoundError(f'{inswapper_path} not found')
w600k_path = Path(__file__).parent / 'w600k_r50.onnx'
SplittedFile.merge(w600k_path, delete_parts=False)
if not w600k_path.exists():
raise FileNotFoundError(f'{w600k_path} not found')
self._sess_swap = InferenceSession_with_device(str(inswapper_path), device_info)
self._sess_rec = InferenceSession_with_device(str(w600k_path), device_info)
swap_onnx_model = onnx.load(str(inswapper_path))
self._emap = numpy_helper.to_array(swap_onnx_model.graph.initializer[-1])
def get_input_size(self):
"""
returns optimal Width/Height for input images, thus you can resize source image to avoid extra load
"""
return 128
def get_face_vector_input_size(self):
return 112
def get_face_vector(self, img : np.ndarray) -> np.ndarray:
ip = ImageProcessor(img)
ip.fit_in(TW=112, TH=112, pad_to_target=True, allow_upscale=True)
img = ip.ch(3).to_ufloat32().get_image('NCHW')
latent = self._sess_rec.run([self._sess_rec.get_outputs()[0].name], {self._sess_rec.get_inputs()[0].name: img,})[0]
latent = np.dot(latent.reshape(1, -1,), self._emap)
latent /= np.linalg.norm(latent)
return latent
def generate(self, img : np.ndarray, face_vector : np.ndarray):
"""
arguments
img np.ndarray HW HWC 1HWC uint8/float32
face_vector np.ndarray
"""
ip_target = ImageProcessor(img)
dtype = ip_target.get_dtype()
_,H,W,_ = ip_target.get_dims()
out = self._sess_swap.run(['output'], {'target' : ip_target.resize( (128, 128) ).ch(3).to_ufloat32().get_image('NCHW'),
'source' : face_vector}
)[0].transpose(0,2,3,1)[0]
out = ImageProcessor(out).to_dtype(dtype).resize((W,H)).get_image('HWC')
return out