colomi
0eb61f3e60
包含： - hair_service_sd: 换发型/换发色算法服务 (端口 8801) - photo_service: LoRA 训练调度服务 (端口 32678) - stable-diffusion-webui: SD WebUI 推理服务 (端口 57860) - kohya_ss_home: 训练环境代码 - meidaojia: 监控测试脚本 - setup.sh: 一键部署脚本 (conda环境恢复 + 配置生成 + 完整性检查) - start_all_services.sh: 启动3个服务 - configure.ini.template: 路径模板化 (BASE_DIR自动推导) - conda_envs/py310.yml: py310 环境定义 大文件 (weights/, models/, data/, conda_envs/*.tar.gz 等) 通过 .gitignore 排除， 由网盘单独上传。
463 lines
17 KiB
Python
463 lines
17 KiB
Python
import torch.nn as nn
|
|
import torch.utils.model_zoo as model_zoo
|
|
import torch
|
|
import numpy as np
|
|
import os
|
|
from utils import landmark_processor
|
|
from algorithm_conf import ConfFactory
|
|
from utils.umeyama import umeyama
|
|
|
|
import cv2
|
|
|
|
__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
|
|
'resnet152']
|
|
|
|
model_urls = {
|
|
'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
|
|
'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
|
|
'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
|
|
'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
|
|
'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
|
|
}
|
|
|
|
def conv3x3(in_planes, out_planes, stride=1):
|
|
"""3x3 convolution with padding"""
|
|
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
|
|
padding=1, bias=False)
|
|
|
|
|
|
def conv1x1(in_planes, out_planes, stride=1):
|
|
"""1x1 convolution"""
|
|
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
|
|
|
|
|
|
class BasicBlock(nn.Module):
|
|
expansion = 1
|
|
|
|
def __init__(self, inplanes, planes, stride=1, downsample=None):
|
|
super(BasicBlock, self).__init__()
|
|
self.conv1 = conv3x3(inplanes, planes, stride)
|
|
self.bn1 = nn.BatchNorm2d(planes)
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.conv2 = conv3x3(planes, planes)
|
|
self.bn2 = nn.BatchNorm2d(planes)
|
|
self.downsample = downsample
|
|
self.stride = stride
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
|
|
out = self.conv1(x)
|
|
out = self.bn1(out)
|
|
out = self.relu(out)
|
|
|
|
out = self.conv2(out)
|
|
out = self.bn2(out)
|
|
|
|
if self.downsample is not None:
|
|
identity = self.downsample(x)
|
|
|
|
out += identity
|
|
out = self.relu(out)
|
|
|
|
return out
|
|
|
|
|
|
class Bottleneck(nn.Module):
|
|
expansion = 4
|
|
|
|
def __init__(self, inplanes, planes, stride=1, downsample=None):
|
|
super(Bottleneck, self).__init__()
|
|
self.conv1 = conv1x1(inplanes, planes)
|
|
self.bn1 = nn.BatchNorm2d(planes)
|
|
self.conv2 = conv3x3(planes, planes, stride)
|
|
self.bn2 = nn.BatchNorm2d(planes)
|
|
self.conv3 = conv1x1(planes, planes * self.expansion)
|
|
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.downsample = downsample
|
|
self.stride = stride
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
|
|
out = self.conv1(x)
|
|
out = self.bn1(out)
|
|
out = self.relu(out)
|
|
|
|
out = self.conv2(out)
|
|
out = self.bn2(out)
|
|
out = self.relu(out)
|
|
|
|
out = self.conv3(out)
|
|
out = self.bn3(out)
|
|
|
|
if self.downsample is not None:
|
|
identity = self.downsample(x)
|
|
|
|
out += identity
|
|
out = self.relu(out)
|
|
|
|
return out
|
|
|
|
|
|
class ResNet(nn.Module):
|
|
|
|
def __init__(self, block, layers, num_classes=1000, is_1k=False, zero_init_residual=False):
|
|
super(ResNet, self).__init__()
|
|
self.inplanes = 64
|
|
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
|
|
bias=False)
|
|
self.bn1 = nn.BatchNorm2d(64)
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
|
self.layer1 = self._make_layer(block, 64, layers[0])
|
|
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
|
|
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
|
|
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
|
|
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
|
|
if is_1k:
|
|
self.fc = nn.Sequential(*[nn.Linear(512 * block.expansion, num_classes), nn.Tanh()])
|
|
else:
|
|
self.fc_key = nn.Sequential(*[nn.Linear(256 * block.expansion, 45 * 2), nn.Tanh()])
|
|
self.fc_ctrl = nn.Sequential(*[nn.Linear(256 * block.expansion, 48 * 2), nn.Tanh()])
|
|
self.is_1k = is_1k
|
|
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
|
elif isinstance(m, nn.BatchNorm2d):
|
|
nn.init.constant_(m.weight, 1)
|
|
nn.init.constant_(m.bias, 0)
|
|
|
|
# Zero-initialize the last BN in each residual branch,
|
|
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
|
|
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
|
|
if zero_init_residual:
|
|
for m in self.modules():
|
|
if isinstance(m, Bottleneck):
|
|
nn.init.constant_(m.bn3.weight, 0)
|
|
elif isinstance(m, BasicBlock):
|
|
nn.init.constant_(m.bn2.weight, 0)
|
|
|
|
def _make_layer(self, block, planes, blocks, stride=1):
|
|
downsample = None
|
|
if stride != 1 or self.inplanes != planes * block.expansion:
|
|
downsample = nn.Sequential(
|
|
conv1x1(self.inplanes, planes * block.expansion, stride),
|
|
nn.BatchNorm2d(planes * block.expansion),
|
|
)
|
|
|
|
layers = []
|
|
layers.append(block(self.inplanes, planes, stride, downsample))
|
|
self.inplanes = planes * block.expansion
|
|
for _ in range(1, blocks):
|
|
layers.append(block(self.inplanes, planes))
|
|
|
|
return nn.Sequential(*layers)
|
|
|
|
def forward(self, x):
|
|
x = self.conv1(x)
|
|
x = self.bn1(x)
|
|
x = self.relu(x)
|
|
x = self.maxpool(x)
|
|
|
|
x = self.layer1(x)
|
|
x = self.layer2(x)
|
|
x = self.layer3(x)
|
|
x = self.layer4(x)
|
|
|
|
if self.is_1k:
|
|
key = self.avgpool(x)
|
|
key = key.view(key.size(0), -1)
|
|
key = self.fc(key)
|
|
return key
|
|
else:
|
|
key, ctrl = torch.chunk(x, 2, dim=1)
|
|
|
|
key = self.avgpool(key)
|
|
key = key.view(key.size(0), -1)
|
|
key = self.fc_key(key)
|
|
|
|
ctrl = self.avgpool(ctrl)
|
|
ctrl = ctrl.view(ctrl.size(0), -1)
|
|
ctrl = self.fc_ctrl(ctrl)
|
|
|
|
return key, ctrl
|
|
|
|
def resnet18(pretrained=False, **kwargs):
|
|
"""Constructs a ResNet-18 model.
|
|
|
|
Args:
|
|
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
|
"""
|
|
model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
|
|
if pretrained:
|
|
model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
|
|
return model
|
|
|
|
|
|
def resnet34(pretrained=False, **kwargs):
|
|
"""Constructs a ResNet-34 model.
|
|
|
|
Args:
|
|
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
|
"""
|
|
model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
|
|
if pretrained:
|
|
model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))
|
|
return model
|
|
|
|
|
|
def resnet50(pretrained=False, **kwargs):
|
|
"""Constructs a ResNet-50 model.
|
|
|
|
Args:
|
|
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
|
"""
|
|
model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
|
|
if pretrained:
|
|
model.load_state_dict(model_zoo.load_url(model_urls['resnet50']))
|
|
return model
|
|
|
|
|
|
def resnet101(pretrained=False, **kwargs):
|
|
"""Constructs a ResNet-101 model.
|
|
|
|
Args:
|
|
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
|
"""
|
|
model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
|
|
if pretrained:
|
|
model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
|
|
return model
|
|
|
|
|
|
def resnet152(pretrained=False, **kwargs):
|
|
"""Constructs a ResNet-152 model.
|
|
|
|
Args:
|
|
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
|
"""
|
|
model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
|
|
if pretrained:
|
|
model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
|
|
return model
|
|
|
|
class Model1k(nn.Module):
|
|
def __init__(self, gpu_id=None):
|
|
super(Model1k, self).__init__()
|
|
|
|
self.device = torch.device('cuda:{}'.format(gpu_id) if gpu_id is not None else 'cpu')
|
|
|
|
self.face_alignment_net = resnet18(pretrained=False, num_classes=1000 * 2, is_1k=True)
|
|
|
|
self.model_dir = ConfFactory.getModelValue("model_dir")
|
|
weights = torch.load(os.path.join(self.model_dir, 'face_alignment_1k.pth'), map_location=lambda storage, loc: storage)
|
|
|
|
self.load_state_dict(weights)
|
|
self.to(self.device)
|
|
self.eval()
|
|
|
|
def forward(self, imgs):
|
|
pred_key_pts = self.face_alignment_net(imgs)
|
|
pred_key_pts = pred_key_pts + 0.5
|
|
return pred_key_pts
|
|
|
|
class MomocvFaceAlignment1K(object):
|
|
def __init__(self, gpu_id=None):
|
|
self.gpu_id = gpu_id
|
|
|
|
self.device = torch.device('cuda:{}'.format(gpu_id) if gpu_id is not None else 'cpu')
|
|
|
|
self.face_alignment_net = Model1k(gpu_id)
|
|
|
|
self.trackingFaceRects = []
|
|
|
|
print('MomocvFaceAlignment1K success')
|
|
|
|
def forward(self, img_tensor):
|
|
fullyconnected1 = self.face_alignment_net(img_tensor).detach().cpu().numpy()
|
|
return fullyconnected1
|
|
|
|
def detect(self, img, landmarks):
|
|
dst_size = 256
|
|
landmarks_res = []
|
|
with torch.no_grad():
|
|
input_numpy = np.zeros((len(landmarks), 3, dst_size, dst_size), dtype=np.float32)
|
|
all_mat = []
|
|
for ix, landmark in enumerate(landmarks):
|
|
M = landmark_processor.get_transform_mat_full_face(landmark, dst_size)
|
|
all_mat.append(M)
|
|
tmp = cv2.warpAffine(img, M, (dst_size, dst_size))
|
|
|
|
# cv2.imshow('inp', tmp)
|
|
# cv2.waitKey()
|
|
|
|
input_numpy[ix, :, :, :] = tmp.transpose((2, 0, 1)).astype(np.float32) / 255
|
|
in_tensor = torch.from_numpy(input_numpy)
|
|
in_tensor = in_tensor.to(self.device)
|
|
fullyconnected1 = self.face_alignment_net(in_tensor).detach().cpu().numpy()
|
|
for ix, pts in enumerate(fullyconnected1):
|
|
orig_pts = (np.reshape(pts, (2, 1000)).transpose((1, 0)) * dst_size)
|
|
orig_pts = landmark_processor.transform_points(orig_pts, all_mat[ix], invert=True)
|
|
landmarks_res.append(orig_pts)
|
|
return landmarks_res
|
|
|
|
def detect_single_face(self, img, crop_M):
|
|
dst_size = 256
|
|
with torch.no_grad():
|
|
|
|
# add for change crop for 1024
|
|
h, w, _ = img.shape
|
|
if h == 768:
|
|
crop_img = img[104:img.shape[0] - 104, 104:img.shape[1] - 104, :]
|
|
tmp = cv2.resize(crop_img, (dst_size, dst_size))
|
|
|
|
else:
|
|
tmp = cv2.warpAffine(img, crop_M, (dst_size, dst_size), flags=cv2.INTER_CUBIC)
|
|
|
|
# cv2.imshow("img_paf_test_crop: ", tmp)
|
|
# cv2.waitKey()
|
|
# crop_img = img[220:img.shape[0] - 220, 266:img.shape[1] - 266, :] # 220 266
|
|
# tmp = cv2.resize(crop_img, (dst_size, dst_size))
|
|
|
|
# cv2.imshow("detect face: h:{:d}".format(h), tmp)
|
|
# cv2.waitKey()
|
|
|
|
input_numpy = np.zeros((1, 3, dst_size, dst_size), dtype=np.float32)
|
|
input_numpy[0, :, :, :] = tmp.transpose((2, 0, 1)).astype(np.float32) / 255
|
|
in_tensor = torch.from_numpy(input_numpy)
|
|
in_tensor = in_tensor.to(self.device)
|
|
fullyconnected1 = self.face_alignment_net(in_tensor).detach().cpu().numpy()
|
|
|
|
orig_pts = np.reshape(fullyconnected1[0], (2, 1000)).transpose((1, 0))
|
|
|
|
if h == 768:
|
|
orig_pts[:, 0] = orig_pts[:, 0] * crop_img.shape[1]
|
|
orig_pts[:, 1] = orig_pts[:, 1] * crop_img.shape[0]
|
|
orig_pts[:, 0] += 104
|
|
orig_pts[:, 1] += 104
|
|
else:
|
|
orig_pts[:, 0] = orig_pts[:, 0] * dst_size
|
|
orig_pts[:, 1] = orig_pts[:, 1] * dst_size
|
|
orig_pts = landmark_processor.transform_points(orig_pts, crop_M, invert=True)
|
|
# orig_pts[:, 0] = orig_pts[:, 0] * crop_img.shape[1]
|
|
# orig_pts[:, 1] = orig_pts[:, 1] * crop_img.shape[0]
|
|
# orig_pts[:, 0] += 266
|
|
# orig_pts[:, 1] += 220
|
|
|
|
return orig_pts
|
|
|
|
def detect_single_face_old(self, img):
|
|
dst_size = 256
|
|
with torch.no_grad():
|
|
tmp = cv2.resize(img, (dst_size, dst_size))
|
|
input_numpy = np.zeros((1, 3, dst_size, dst_size), dtype=np.float32)
|
|
input_numpy[0, :, :, :] = tmp.transpose((2, 0, 1)).astype(np.float32) / 255
|
|
in_tensor = torch.from_numpy(input_numpy)
|
|
in_tensor = in_tensor.to(self.device)
|
|
fullyconnected1 = self.face_alignment_net(in_tensor).detach().cpu().numpy()
|
|
orig_pts = (np.reshape(fullyconnected1[0], (2, 1000)).transpose((1, 0)) * img.shape[0])
|
|
return orig_pts
|
|
|
|
def detect_according_5pts(self, img, pts5):
|
|
dst_size = 256
|
|
with torch.no_grad():
|
|
input_numpy = np.zeros((1, 3, dst_size, dst_size), dtype=np.float32)
|
|
eye_dis = 0.34
|
|
mouth_dis = 0.34
|
|
g_Average_5point_180 = np.array([
|
|
eye_dis, 0.3,
|
|
1 - eye_dis, 0.3,
|
|
0.5, 0.6,
|
|
mouth_dis, 0.63,
|
|
1 - mouth_dis, 0.63
|
|
])
|
|
# print(g_Average_5point_180)
|
|
left_eye = np.array([pts5[0], pts5[5]])
|
|
right_eye = np.array([pts5[1], pts5[6]])
|
|
nose = np.array([pts5[2], pts5[7]])
|
|
left_mouth = np.array([pts5[3], pts5[8]])
|
|
right_mouth = np.array([pts5[4], pts5[9]])
|
|
|
|
pts5_src = np.vstack((left_eye, right_eye,
|
|
nose,
|
|
left_mouth, right_mouth))
|
|
pts5_src = np.array(pts5_src).astype(np.int32)
|
|
pts5_dst = g_Average_5point_180.reshape((5, -1)) * dst_size
|
|
|
|
mat = umeyama(pts5_src, pts5_dst, True)[0:2]
|
|
|
|
tmp = cv2.warpAffine(img, mat, (dst_size, dst_size))
|
|
# cv2.imshow("tmp", tmp)
|
|
# cv2.waitKey()
|
|
input_numpy[0, :, :, :] = tmp.transpose((2, 0, 1)).astype(np.float32) / 255
|
|
in_tensor = torch.from_numpy(input_numpy)
|
|
in_tensor = in_tensor.to(self.device)
|
|
fullyconnected1 = self.face_alignment_net(in_tensor).detach().cpu().numpy()
|
|
orig_pts = (np.reshape(fullyconnected1[0], (2, 1000)).transpose((1, 0)) * dst_size)
|
|
orig_pts = landmark_processor.transform_points(orig_pts, mat, invert=True)
|
|
return orig_pts
|
|
|
|
|
|
def stable_forward(self, image, detected_faces, reset=False):
|
|
if reset is True:
|
|
self.trackingFaceRects = []
|
|
|
|
if len(self.trackingFaceRects) == 0:
|
|
for face_rect in detected_faces:
|
|
new_tracking_rect = [face_rect, True, [0, 0], 0, None]
|
|
self.trackingFaceRects.append(new_tracking_rect)
|
|
|
|
with torch.no_grad():
|
|
landmarks = []
|
|
for ix, tracking_face_rect in enumerate(self.trackingFaceRects):
|
|
if tracking_face_rect[1] == True:
|
|
d = tracking_face_rect[0]
|
|
src_center = np.array([d[2] - (d[2] - d[0]) / 2.0, d[3] - (d[3] - d[1]) / 2.0])
|
|
rotate_degree = tracking_face_rect[3]
|
|
scale = 256 * 0.6 / min(d[2] - d[0], d[3] - d[1])
|
|
dst_center = np.array([0.5, 0.5]) * 256
|
|
offset = dst_center - src_center
|
|
M = cv2.getRotationMatrix2D((src_center[0], src_center[1]), rotate_degree, scale)
|
|
M[:, 2] += offset
|
|
else:
|
|
rotate_degree = 0
|
|
M = landmark_processor.get_transform_mat_mmcv_bigger(tracking_face_rect[4], 256)
|
|
inp = cv2.warpAffine(image, M, (256, 256))
|
|
|
|
# cv2.imshow('inp_{}'.format(ix), inp)
|
|
# cv2.waitKey()
|
|
orig_inp = inp
|
|
|
|
inp = inp.transpose((2, 0, 1)).astype(np.float32)
|
|
inp = inp[np.newaxis, :, :, :] / 255
|
|
|
|
in_tensor = torch.from_numpy(inp)
|
|
in_tensor = in_tensor.cuda(0)
|
|
fullyconnected1 = self.forward(in_tensor)
|
|
fullyconnected1 = fullyconnected1[0]
|
|
orig_pts = (np.reshape(fullyconnected1, (2, 1000)).transpose((1, 0))) * 256
|
|
|
|
t2 = cv2.getTickCount()
|
|
orig_pts = landmark_processor.transform_points(orig_pts, M, invert=True)
|
|
|
|
# orig_pts = orig_pts.transpose((1, 0))
|
|
fullyconnected1 = orig_pts
|
|
|
|
# update tracking infos
|
|
tracking_face_rect[1] = False
|
|
tracking_face_rect[2] = None
|
|
tracking_face_rect[3] = rotate_degree
|
|
tracking_face_rect[4] = fullyconnected1
|
|
|
|
# fullyconnected1 = landmark_processor.pts_1k_to_137(fullyconnected1)
|
|
|
|
# eye_landmark = self.detect_eye(image, fullyconnected1)
|
|
# fullyconnected1[87:104] = eye_landmark[0]
|
|
# fullyconnected1[104:121] = eye_landmark[1]
|
|
|
|
landmarks.append(fullyconnected1)
|
|
return landmarks
|