-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
264 lines (220 loc) · 8.79 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
253
254
255
256
257
258
259
260
261
262
263
264
import os
import torch
from tqdm import tqdm_notebook
from matplotlib import pyplot as plt
from itertools import product
import json
from collections import defaultdict
from torch import nn
from torch.autograd import Variable
from torch.functional import F
from torchvision import models
import torchvision
from torch.utils.data import Dataset, DataLoader
import torch.optim as optim
import cv2
import numpy as np
# from fastai import transforms, model, dataset, conv_learner
from PIL import ImageDraw, ImageFont
from matplotlib import patches, patheffects
from matplotlib import pyplot as plt
import numpy as np
import pandas as pd
from augmentation import SSDAugmentation, SSD_Val_Augmentation
from Config import Config
from model import get_SSD_model, lr_find
from VOC_data import VOC_dataset
from loss import *
from mAP import mAP
torch.set_printoptions(precision=3)
# PATH = 'C:\\datasets\\pascal\\'
# anno_path = f'{PATH}PASCAL_VOC\\pascal_train2007.json'
# train_dataset = VOC_dataset(PATH, anno_path)
# batch_size = 16
# learning_rate = 5e-4
# vgg_weight_path = 'C:\\Users\\ruifr\\.torch\\models\\vgg16-397923af.pth'
def detection_collate_fn(batch):
imgs, bboxes, labels, img_id, ignore, img_scale = [], [], [], [], [], []
for i, b, l, id, ig, s in batch:
imgs.append(i)
bboxes.append(b)
labels.append(l)
img_id.append(id)
ignore.append(ig)
img_scale.append(s)
return torch.stack(imgs), bboxes, labels, img_id, ignore, img_scale
def adjust_lr(epoch):
lr = 1
if epoch > 80:
lr /= 10
if epoch > 100:
lr /= 10
return lr
def train_ssd():
trn_id_fname, trn_id_annotation, trn_id_single_anno, idx_category, category_idx, imgs, imgs_id, imgs_bbox, imgs_class = (
get_anno_data()
)
config = Config("remote")
test_dataset = VOC_dataset(
config.voc2007_root, config.voc2012_root, config.voc2007_test_anno, "test"
)
trn_dataset = VOC_dataset(
config.voc2007_root, config.voc2012_root, config.anno_path, "trn"
)
test_dataloader = DataLoader(
test_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
collate_fn=detection_collate_fn,
)
trn_dataloader = DataLoader(
trn_dataset,
batch_size=32,
shuffle=True,
num_workers=8,
collate_fn=detection_collate_fn,
)
ssd_model = get_SSD_model(1, config.vgg_weight_path, config.vgg_reduced_weight_path)
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
ssd_model = ssd_model.to(device)
prior_box = get_prior_box()
# prior_box = prior_box.to(device)
loss_array = []
val_array = []
print("success build ssd model to train")
optimizer = torch.optim.SGD(
ssd_model.parameters(), lr=1e-3, momentum=0.9, weight_decay=5 * 1e-4
)
lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, adjust_lr)
for epoch in range(120):
lr_scheduler.step()
# train
for i, batch in enumerate(trn_dataloader):
imgs, bboxes, labels, img_id, ignores, img_scale = batch
# bboxes, labels, img_id, ignores, img_scale = bboxes[0], labels[0], img_id[0], ignores[0], img_scale[0]
imgs = imgs.to(device)
# bboxes = bboxes.to(device)
cls_preds, loc_preds = ssd_model(imgs)
ssd_model.zero_grad()
total_loss = 0
total_loc_loss, total_cls_loss = 0, 0
for idx in range(imgs.shape[0]):
img, bbox, label = imgs[idx], bboxes[idx], labels[idx]
cls_pred, loc_pred = cls_preds[idx], loc_preds[idx]
iou = get_iou(bbox, prior_box)
pos_mask, cls_target, bbox_target = get_target(
iou, prior_box, img, bbox, label
)
pos_mask, cls_target, bbox_target = (
pos_mask.to(device),
cls_target.to(device),
bbox_target.to(device),
)
loss_loc, loss_cls = loss(
cls_pred, loc_pred, pos_mask, cls_target, bbox_target
)
total_loc_loss += loss_loc
total_cls_loss += loss_cls
total_loss += loss_cls + loss_loc
total_loss /= float(imgs.shape[0])
total_cls_loss /= float(imgs.shape[0])
total_loc_loss /= float(imgs.shape[0])
total_loss.backward()
optimizer.step()
cls_loss = round(float(total_cls_loss), 3)
loc_loss = round(float(total_loc_loss), 3)
t_loss = round(float(total_loss), 3)
if i % 5 == 0:
print(
epoch * 515 + i,
"cls_loss: {}, loc_loss: {}, loss: {}".format(
cls_loss, loc_loss, t_loss
),
)
loss_array.append(t_loss)
# val and save every 5 epoch
if epoch % 5 == 0:
torch.save(ssd_model.state_dict(), "f_trained_{}_epoch".format(i))
print("val--------------------------")
for val_i, batch in enumerate(test_dataloader):
imgs, bboxes, labels, img_id, ignores, img_scale = batch
imgs = imgs.to(device)
cls_preds, loc_preds = ssd_model(imgs)
total_loss = 0
total_loc_loss, total_cls_loss = 0, 0
for idx in range(imgs.shape[0]):
img, bbox, label = imgs[idx], bboxes[idx], labels[idx]
cls_pred, loc_pred = cls_preds[idx], loc_preds[idx]
iou = get_iou(bbox, prior_box)
pos_mask, cls_target, bbox_target = get_target(
iou, prior_box, img, bbox, label
)
pos_mask, cls_target, bbox_target = (
pos_mask.to(device),
cls_target.to(device),
bbox_target.to(device),
)
loss_loc, loss_cls = loss(
cls_pred, loc_pred, pos_mask, cls_target, bbox_target
)
total_loc_loss += loss_loc
total_cls_loss += loss_cls
total_loss += loss_cls + loss_loc
total_loss /= float(imgs.shape[0])
total_cls_loss /= float(imgs.shape[0])
total_loc_loss /= float(imgs.shape[0])
cls_loss = round(float(total_cls_loss), 3)
loc_loss = round(float(total_loc_loss), 3)
t_loss = round(float(total_loss), 3)
if val_i % 100 == 0:
print(
val_i,
"cls_loss: {}, loc_loss: {}, loss: {}".format(
cls_loss, loc_loss, t_loss
),
)
val_array.append(t_loss)
# valdiate the mAP
print("mAP")
mean_average_precision = mAP(config.voc2007_test_anno)
for i, batch in tqdm_notebook(enumerate(test_dataloader)):
imgs, bboxes, labels, img_id, ignores, img_scale = batch
bboxes, labels, img_id, ignores, img_scale = (
bboxes[0],
labels[0],
img_id[0],
ignores[0],
img_scale[0],
)
imgs = imgs.to(device)
cls_preds, loc_preds = ssd_model(imgs)
res_score, res_bbox, res_cls = mean_average_precision.nms(
cls_preds, loc_preds, prior_box, conf_threshold=0.01
)
for _ in range(len(bboxes)):
bboxes[_][1] *= 300
bboxes[_][3] *= 300
bboxes[_][0] *= 300
bboxes[_][2] *= 300
for _ in range(len(bboxes)):
mean_average_precision.add_single_gt(
img_id, bboxes[_], labels[_], ignores[_]
)
for _ in range(len(res_score)):
mean_average_precision.add_single_pred(
img_id,
res_score[_].cpu().detach().numpy(),
res_bbox[_].cpu().detach().numpy(),
res_cls[_],
)
res = mean_average_precision.calculate_mAP()
# save loss array and map to log
with open("res.json", "a") as f:
json.dump({"epoch": epoch, "loss": loss_array, "map": res}, f)
f.write("\n")
print("mAP: ", np.mean([res[k] for k in res.keys()]))
print("finish val--------------------------")
if __name__ == "__main__":
train_ssd()