forked from Z-Zheng/xview2_4th_solution
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
144 lines (111 loc) · 4.79 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
from simplecv import apex_ddp_train as train
from data import xview2_loader
from module import register_model
import torch
import simplecv as sc
import time
from simplecv.util.logger import eval_progress, speed
import os
from concurrent.futures import ProcessPoolExecutor
import numpy as np
from simplecv.core.transform_base import ParallelTestTransform
def tta(model, image, tta_config):
trans = ParallelTestTransform(
*tta_config
)
images = trans.transform(image)
with torch.no_grad():
outs = [model(im) for im in images]
outs = trans.inv_transform(outs)
out = sum(outs) / len(outs)
return out
def evaluate_cls_fn(self, test_dataloader, config=None):
vis_dir = os.path.join(self.model_dir, 'vis-{}'.format(self.checkpoint.global_step))
torch.cuda.empty_cache()
self.model.eval()
total_time = 0.
ppe = ProcessPoolExecutor(max_workers=4)
metric_op = sc.metric.NPPixelMertic(max(self.model.module.config.head.num_classes, 2), self.model_dir)
# 0 - bg, 1 - building
viz_op = sc.viz.VisualizeSegmm(vis_dir, [0, 0, 0, 0, 0, 255])
with torch.no_grad():
for idx, (ret, ret_gt) in enumerate(test_dataloader):
start = time.time()
ret = ret.to(torch.device('cuda'))
if config is not None and 'tta' in config:
y = tta(self.model, ret, config['tta'])
else:
y = self.model(ret)
cls = (y > 0.5).cpu()
cls = cls.numpy()
cls_gt = ret_gt['cls']
cls_gt = cls_gt.numpy()
y_true = cls_gt.ravel()
y_pred = cls.ravel()
y_true = np.where(y_true > 0, np.ones_like(y_true), np.zeros_like(y_true))
metric_op.forward(y_true, y_pred)
time_cost = round(time.time() - start, 3)
total_time += time_cost
filename = ret_gt['image_filename']
ppe.submit(viz_op, cls, filename[0].replace('jpg', 'png'))
speed(self._logger, time_cost, 'batch')
eval_progress(self._logger, idx + 1, len(test_dataloader))
torch.cuda.empty_cache()
speed(self._logger, round(total_time / len(test_dataloader), 3), 'batch (avg)')
metric_op.summary_all()
torch.cuda.empty_cache()
def evaluate_loc_cls_fn(self, test_dataloader, config=None):
# vis_dir = os.path.join(self.model_dir, 'vis-{}'.format(self.checkpoint.global_step))
torch.cuda.empty_cache()
self.model.eval()
total_time = 0.
# ppe = ProcessPoolExecutor(max_workers=4)
loc_metric_op = sc.metric.NPPixelMertic(2, self.model_dir)
damage_metric_op = sc.metric.NPPixelMertic(max(self.model.module.config.head.num_classes, 2), self.model_dir)
# # 0 - bg, 1 - building
# viz_op = sc.viz.VisualizeSegmm(vis_dir, [0, 0, 0, 0, 0, 255])
with torch.no_grad():
for idx, (ret, ret_gt) in enumerate(test_dataloader):
start = time.time()
ret = ret.to(torch.device('cuda'))
if config is not None and 'tta' in config:
y = tta(self.model, ret, config['tta'])
else:
y = self.model(ret)
loc_prob = y[:, :1, :, :]
loc_pred = (loc_prob > 0.5).cpu().numpy().ravel()
gt = ret_gt['cls']
loc_true = gt[:, :, :, 0].numpy().ravel()
loc_true = np.where(loc_true > 0, np.ones_like(loc_true), np.zeros_like(loc_true))
loc_metric_op.forward(loc_true, loc_pred)
dam_prob = y[:, 1:, :, :]
dam_pred = dam_prob.argmax(dim=1).cpu().numpy().ravel()
dam_true = gt[:, :, :, 1].numpy().ravel()
valid_inds = np.where(dam_true != self.model.module.config.loss.ignore_index)[0]
dam_true = dam_true[valid_inds]
dam_pred = dam_pred[valid_inds]
damage_metric_op.forward(dam_true, dam_pred)
time_cost = round(time.time() - start, 3)
total_time += time_cost
speed(self._logger, time_cost, 'batch')
eval_progress(self._logger, idx + 1, len(test_dataloader))
torch.cuda.empty_cache()
speed(self._logger, round(total_time / len(test_dataloader), 3), 'batch (avg)')
loc_metric_op.summary_all()
damage_metric_op.summary_all()
torch.cuda.empty_cache()
def register_evaluate_fn(launcher):
launcher.override_evaluate(evaluate_cls_fn)
if __name__ == '__main__':
torch.backends.cudnn.benchmark = True
SEED = 2333
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
args = train.parser.parse_args()
train.run(local_rank=args.local_rank,
config_path=args.config_path,
model_dir=args.model_dir,
opt_level=args.opt_level,
cpu_mode=args.cpu,
after_construct_launcher_callbacks=[register_evaluate_fn],
opts=args.opts)