-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
252 lines (196 loc) · 8.61 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
import argparse
#pytorch
from concurrent.futures import thread
import subprocess
from xmlrpc.client import Boolean
from sqlalchemy import null
import torch
from torchvision import transforms
from threading import Thread
#other lib
import sys
import numpy as np
import os
import cv2
import shutil
sys.path.insert(0, "yolov5_face")
from models.experimental import attempt_load
from utils.datasets import letterbox
from utils.general import check_img_size, non_max_suppression_face, scale_coords
# Check device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Get model detect
## Case 1:
# model = attempt_load("yolov5_face/yolov5s-face.pt", map_location=device)
## Case 2:
model = attempt_load("yolov5_face/yolov5n-0.5.pt", map_location=device)
# Get model recognition
## Case 1:
from insightface.insight_face import iresnet100
weight = torch.load("insightface/resnet100_backbone.pth", map_location = device)
model_emb = iresnet100()
## Case 2:
# from insightface.insight_face import iresnet18
# weight = torch.load("insightface/resnet18_backbone.pth", map_location = device)
# model_emb = iresnet18()
model_emb.load_state_dict(weight)
model_emb.to(device)
model_emb.eval()
face_preprocess = transforms.Compose([
transforms.ToTensor(), # input PIL => (3,56,56), /255.0
transforms.Resize((112, 112)),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])
def resize_image(img0, img_size):
h0, w0 = img0.shape[:2] # orig hw
r = img_size / max(h0, w0) # resize image to img_size
if r != 1: # always resize down, only resize up if training with augmentation
interp = cv2.INTER_AREA if r < 1 else cv2.INTER_LINEAR
img0 = cv2.resize(img0, (int(w0 * r), int(h0 * r)), interpolation=interp)
imgsz = check_img_size(img_size, s=model.stride.max()) # check img_size
img = letterbox(img0, new_shape=imgsz)[0]
# Convert
img = img[:, :, ::-1].transpose(2, 0, 1).copy() # BGR to RGB, to 3x416x416
img = torch.from_numpy(img).to(device)
img = img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
return img
def get_face(input_image):
# Parameters
size_convert = 256
conf_thres = 0.4
iou_thres = 0.5
# Resize image
img = resize_image(input_image.copy(), size_convert)
# Via yolov5-face
with torch.no_grad():
pred = model(img[None, :])[0]
# Apply NMS
det = non_max_suppression_face(pred, conf_thres, iou_thres)[0]
bboxs = np.int32(scale_coords(img.shape[1:], det[:, :4], input_image.shape).round().cpu().numpy())
return bboxs
def get_feature(face_image, training = True):
# Convert to RGB
face_image = cv2.cvtColor(face_image, cv2.COLOR_BGR2RGB)
# Preprocessing image BGR
face_image = face_preprocess(face_image).to(device)
# Via model to get feature
with torch.no_grad():
if training:
emb_img_face = model_emb(face_image[None, :])[0].cpu().numpy()
else:
emb_img_face = model_emb(face_image[None, :]).cpu().numpy()
# Convert to array
images_emb = emb_img_face/np.linalg.norm(emb_img_face)
return images_emb
def read_features(root_fearure_path):
try:
data = np.load(root_fearure_path + ".npz", allow_pickle=True)
images_name = data["arr1"]
images_emb = data["arr2"]
return images_name, images_emb
except:
return null
def training(full_training_dir, additional_training_dir,
faces_save_dir, features_save_dir, is_add_user):
# Init results output
images_name = []
images_emb = []
# Check mode full training or additidonal
if is_add_user == True:
source = additional_training_dir
else:
source = full_training_dir
# Read train folder, get and save face
for name_person in os.listdir(source):
person_image_path = os.path.join(source, name_person)
# Create path save person face
person_face_path = os.path.join(faces_save_dir, name_person)
os.makedirs(person_face_path, exist_ok=True)
# Read existing features (if exists)
features = read_features(features_save_dir)
if features == null or is_add_user== False:
processed_files = []
else:
# Read features
old_images_name, old_images_emb = features
# Get processed files
processed_files = list(old_images_name)
# Process new files
for image_name in os.listdir(person_image_path):
if image_name.endswith(("png", 'jpg', 'jpeg')):
# Skip already processed files
if image_name in processed_files:
continue
image_path = person_image_path + f"/{image_name}"
input_image = cv2.imread(image_path) # BGR
# Get faces
bboxs = get_face(input_image)
# Get boxs
for i in range(len(bboxs)):
# Get number files in person path
number_files = len(os.listdir(person_face_path))
# Get location face
x1, y1, x2, y2 = bboxs[i]
# Get face from location
face_image = input_image[y1:y2, x1:x2]
# Path save face
path_save_face = person_face_path + f"/{number_files}.jpg"
# Save to face database
cv2.imwrite(path_save_face, face_image)
# Get feature from face
images_emb.append(get_feature(face_image, training=True))
images_name.append(name_person)
# Convert to array
images_emb = np.array(images_emb)
images_name = np.array(images_name)
# Merge with existing features (if exists)
features = read_features(features_save_dir)
if features == null or is_add_user == False:
pass
else:
# Read features
old_images_name, old_images_emb = features
# Add feature and name of image to feature database
images_name = np.hstack((old_images_name, images_name))
images_emb = np.vstack((old_images_emb, images_emb))
print("Update feature!")
# Save features
np.savez_compressed(features_save_dir,
arr1 = images_name, arr2 = images_emb)
# Move additional data to full train data
if is_add_user == True:
for sub_dir in os.listdir(additional_training_dir):
dir_to_move = os.path.join(additional_training_dir, sub_dir)
dest_dir = os.path.join(full_training_dir, sub_dir)
if os.path.exists(dest_dir):
# Copy existing files
for file_name in os.listdir(dir_to_move):
src_file_path = os.path.join(dir_to_move, file_name)
dest_file_path = os.path.join(dest_dir, file_name)
shutil.copy2(src_file_path, dest_file_path)
os.remove(src_file_path)
# Remove additional training directory
shutil.rmtree(dir_to_move)
else:
shutil.move(dir_to_move, dest_dir)
def parse_opt():
parser = argparse.ArgumentParser()
parser.add_argument('--full-training-dir', type=str, default='./dataset/full-training-datasets/', help='dir folder full training')
parser.add_argument('--additional-training-dir', type=str, default='./dataset/additional-training-datasets/', help='dir folder additional training')
parser.add_argument('--faces-save-dir', type=str, default='./dataset/face-datasets/', help='dir folder save face features')
parser.add_argument('--features-save-dir', type=str, default='./feature/face_features', help='dir folder save face features')
parser.add_argument('--is-add-user', type=bool, default=True, help='Mode add user or full training')
opt = parser.parse_args()
return opt
def main(opt):
output1 = subprocess.run(["find",".","-name","\"*.DS_Store\"","-type","f","-delete"], capture_output=True)
print(output1.stdout.decode())
training(full_training_dir=opt.full_training_dir,
additional_training_dir=opt.additional_training_dir,
faces_save_dir=opt.faces_save_dir,
features_save_dir=opt.features_save_dir,
is_add_user=opt.is_add_user)
if __name__ == "__main__":
opt = parse_opt()
main(opt)