-
Notifications
You must be signed in to change notification settings - Fork 0
/
hybridnetsModel.py
258 lines (212 loc) · 11.6 KB
/
hybridnetsModel.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
253
254
255
256
257
258
import torch
import numpy as np
from torch import nn
from PIL import Image, ImageFont, ImageDraw
from _utils.utils import BBoxUtility, calculate_f1score
from network.Hybirdnets import Hybridnets
from custom.customlosses import ConfidenceLoss, BBOXL1Loss
import configure.config as cfg
class HybridnetsModel:
def __init__(self,
fpn_cells: int,
num_layers: int,
num_anchors: int,
num_classes: int,
seg_classes: int,
num_features: int,
conv_channels: list,
out_indices: tuple,
up_scale: tuple,
backbone: str,
priors: np.ndarray,
learning_rate: float,
weight_decay: float,
iou_thresh: float,
nms_thresh: float,
resume_train: bool,
ckpt_path: str,
device: object = None):
'''
:param fpn_cells:
:param num_layers: 分类头和回归头 网络的层数
:param num_anchors:
:param num_classes:
:param seg_classes:
:param num_features:
:param conv_channels:
:param out_indices:
:param up_scale:
:param backbone:
:param priors:
:param learning_rate:
:param weight_decay:
:param iou_thresh:
:param nms_thresh:
:param resume_train:
:param ckpt_path:
:param device:
'''
self.box_utils = BBoxUtility(priors=priors,
num_classes=num_classes - 1,
overlap_threshold=iou_thresh,
nms_thresh=nms_thresh)
self.network = Hybridnets(backbone=backbone,
fpn_cells=fpn_cells,
out_indices=out_indices,
conv_channels=conv_channels,
num_features=num_features,
num_anchors=num_anchors,
num_classes=num_classes,
num_layers=num_layers,
num_seg_classes=seg_classes,
up_scale=up_scale)
if device:
self.device = device
self.network.to(device)
if resume_train:
try:
ckpt = torch.load(ckpt_path)
self.network.load_state_dict(ckpt['state_dict'])
print("model successfully loaded,loss is {:3f}".format(ckpt['loss']))
except FileNotFoundError:
raise ('please enter the right params path')
self.conf_loss = ConfidenceLoss()
self.bbox_loss = BBOXL1Loss()
self.seg_loss = nn.BCELoss(reduction='mean')
weights, bias = self.network.split_weights()
self.optimizer = torch.optim.Adam(params=[{'params': weights, 'weight_decay': weight_decay},
{'params': bias}],
lr=learning_rate)
self.train_loss, self.val_loss = 0, 0
self.train_acc, self.val_acc = 0, 0
self.train_conf_acc, self.val_conf_acc = 0, 0
self.train_f1_score, self.val_f1_score = 0, 0
self.num_classes = num_classes - 1
def train(self, sources, seg_sources, targets):
sources = torch.tensor(sources).float()
seg_sources = torch.tensor(seg_sources).float()
targets = torch.tensor(targets).float()
if self.device:
sources = sources.to(self.device)
seg_sources = seg_sources.to(self.device)
targets = targets.to(self.device)
self.optimizer.zero_grad() #
regressions, classifications, segmentations = self.network(sources)
regressions = torch.reshape(input=regressions,
shape=(regressions.size(0), -1, regressions.size(-1)))
classifications = torch.reshape(input=classifications,
shape=(classifications.size(0), -1, classifications.size(-1)))
conf_loss = self.conf_loss(classifications, targets[..., 4:])
bbox_loss = self.bbox_loss(regressions, targets[..., :5])
seg_loss = self.seg_loss(torch.permute(input=segmentations, dims=(0, 2, 3, 1)), seg_sources)
loss = conf_loss + bbox_loss + seg_loss
loss.backward() #
self.optimizer.step() #
self.train_loss += loss.data.item()
prob_confs = torch.where(condition=torch.ge(classifications[..., 0:1], .5),
input=torch.ones_like(classifications[..., 0:1]),
other=torch.zeros_like(classifications[..., 0:1]))
# torch.prod() 求乘积
# .data 得到一个新的tensor,但是没有梯度；新的tensor改变，旧的tensor也会改变，旧tensor依然可以求导，不安全
# .detach() 得到一个新的tensor,但是没有梯度；新的tensor改变，旧的tensor也会改变，旧tensor不可以求导，安全
# .item()返回一个标准的python数据类型，只能用于tensor有一个元素
# .tolist() 返回一个标准的python数据list，可用于有多个元素的tensor
total_num = torch.prod(torch.tensor(targets[..., 4:5].size())).float().detach().item()
object_num = (1 - targets[..., 4:5]).cpu().sum().data.item()
correct_conf_num = torch.eq(targets[..., 4:5], prob_confs).float().detach().cpu().sum().item()
self.train_conf_acc += correct_conf_num / total_num # 是否区别背景和目标的准确率
object_mask = (1 - targets[..., 4:5]).squeeze(dim=-1).bool()
prob_class = classifications[..., 1:][object_mask].argmax(dim=-1)
real_class = targets[..., 5:][object_mask].argmax(dim=-1)
correct_class_num = torch.eq(real_class, prob_class).float().detach().sum().cpu().item()
self.train_acc += correct_class_num / object_num # 目标的查全率
self.train_f1_score += calculate_f1score(y_true=targets[..., 5:],
y_pred=classifications[..., 1:],
object_mask=object_mask,
num_classes=self.num_classes)
def validate(self, sources, seg_sources, targets):
sources = torch.tensor(sources).float()
seg_sources = torch.tensor(seg_sources).float()
targets = torch.tensor(targets).float()
if self.device:
sources = sources.to(self.device)
seg_sources = seg_sources.to(self.device)
targets = targets.to(self.device)
regressions, classifications, segmentations = self.network(sources)
regressions = torch.reshape(input=regressions,
shape=(regressions.size(0), -1, regressions.size(-1)))
classifications = torch.reshape(input=classifications,
shape=(classifications.size(0), -1, classifications.size(-1)))
conf_loss = self.conf_loss(classifications, targets[..., 4:])
bbox_loss = self.bbox_loss(regressions, targets[..., :5])
seg_loss = self.seg_loss(torch.permute(input=segmentations, dims=(0, 2, 3, 1)), seg_sources)
loss = conf_loss + bbox_loss + seg_loss
self.val_loss += loss.data.item()
prob_confs = torch.where(torch.ge(classifications[..., 0:1], .5),
torch.ones_like(classifications[..., 0:1]),
torch.zeros_like(classifications[..., 0:1]))
total_num = torch.prod(torch.tensor(targets[..., 4:5].size())).data.item()
object_num = (1 - targets[..., 4:5]).cpu().sum().data.item()
correct_conf_num = torch.eq(targets[..., 4:5], prob_confs).float().detach().cpu().sum().data.item()
self.val_conf_acc += correct_conf_num / total_num
object_mask = (1 - targets[..., 4:5]).squeeze(dim=-1).bool()
prob_class = classifications[..., 1:][object_mask].argmax(dim=-1)
real_class = targets[..., 5:][object_mask].argmax(dim=-1)
correct_class_num = torch.eq(real_class, prob_class).float().detach().cpu().sum().data.item()
self.val_acc += correct_class_num / object_num
self.val_f1_score += calculate_f1score(targets[..., 5:], classifications[..., 1:],
object_mask, self.num_classes)
def generate_sample(self, sources, batch, sample_path):
"""
Drawing and labeling
"""
sources = torch.tensor(sources).float()
if self.device:
sources = sources.to(self.device)
regressions, classifications, segmentations = self.network(sources)
regressions = regressions.reshape(regressions.size(0), -1, regressions.size(-1))
classifications = classifications.reshape(classifications.size(0), -1, classifications.size(-1))
regressions = regressions.detach().cpu().numpy()
classifications = classifications.detach().cpu().numpy()
segmentations = segmentations.detach().cpu().numpy()
index = np.random.choice(sources.size(0), 1)
out_boxes, out_scores, out_classes = self.box_utils.detection_out([regressions[index],
classifications[index]])
out_boxes = np.array(out_boxes).squeeze(axis=0)
out_scores = np.array(out_scores).squeeze(axis=0)
out_classes = np.array(out_classes).squeeze(axis=0)
source = sources[index[0]].cpu().numpy().transpose([1, 2, 0])
segmentation = segmentations[index[0]].transpose([1, 2, 0])
image = Image.fromarray(np.uint8(source * 255))
if out_boxes.shape[0]:
out_boxes = self.box_utils.correct_boxes(out_boxes, np.array(cfg.input_size),
np.array(cfg.input_size))
out_boxes *= np.tile(np.array(cfg.input_size)[::-1], (2,))
for coordinate, out_score, out_class in zip(out_boxes.astype('int'),
out_scores,
out_classes):
left, top = coordinate[:2].tolist()
right, bottom = coordinate[2:].tolist()
font = ImageFont.truetype(font=cfg.font_path,
size=np.floor(4e-2 * image.size[1] + 0.5).astype('int32'))
label = '{:s}: {:.2f}'.format(cfg.class_names[int(out_class)], out_score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
label = label.encode('utf-8')
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
draw.rectangle(coordinate[:2].tolist() + coordinate[2:].tolist(),
outline=cfg.rect_color, width=int(2 * cfg.thickness))
draw.text(text_origin, str(label, 'UTF-8'),
fill=cfg.font_color, font=font)
del draw
image = np.array(image)
segmentation = segmentation.argmax(axis=-1)
for i in range(cfg.segmentation_class_names.__len__()):
if not i:
continue
image[np.equal(segmentation, i)] = cfg.segmentation_colors[i]
image = Image.fromarray(image)
image.save(sample_path.format(batch), quality=95, subsampling=0)