-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_4ch_vit.py
212 lines (155 loc) · 7.44 KB
/
train_4ch_vit.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
import os
import shutil
import json
import time
from apex import amp
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import LambdaLR
from torch.utils.data import DataLoader
from toolbox import get_dataset # loss
from toolbox.optim.Ranger import Ranger
from toolbox import get_logger
from toolbox import get_model
from toolbox import averageMeter, runningScore
from toolbox import save_ckpt
from toolbox.datasets.irseg import IRSeg
from toolbox.datasets.pst900 import PSTSeg
from toolbox.losses import lovasz_softmax
def KD_KLDivLoss(Stu_output, Tea_output, temperature):
T = temperature
KD_loss = nn.KLDivLoss()(F.log_softmax(Stu_output/T, dim=1), F.softmax(Tea_output/T, dim=1)).cuda()
KD_loss = KD_loss * T * T
return KD_loss
def loss_fn(x, y):
x = F.normalize(x, dim=1, p=2)
y = F.normalize(y, dim=1, p=2)
return 2 - 2 * (x * y).sum(dim=1)
class eeemodelLoss(nn.Module):
def __init__(self, class_weight=None, ignore_index=-100, reduction='mean'):
super(eeemodelLoss, self).__init__()
self.class_weight_semantic = torch.from_numpy(np.array(
[1.5105, 16.6591, 29.4238, 34.6315, 40.0845, 41.4357, 47.9794, 45.3725, 44.9000])).float()
self.class_weight_binary = torch.from_numpy(np.array([1.5121, 10.2388])).float()
self.class_weight_boundary = torch.from_numpy(np.array([1.4459, 23.7228])).float()
self.class_weight = class_weight
# self.LovaszSoftmax = lovasz_softmax()
self.cross_entropy = nn.CrossEntropyLoss()
self.mse = nn.MSELoss()
self.semantic_loss = nn.CrossEntropyLoss(weight=self.class_weight_semantic)
self.binary_loss = nn.CrossEntropyLoss(weight=self.class_weight_binary)
self.boundary_loss = nn.CrossEntropyLoss(weight=self.class_weight_boundary)
def forward(self, inputs, targets):
semantic_gt, binary_gt, boundary_gt = targets
semantic_out,semantic_out_2 = inputs
loss1 = self.semantic_loss(semantic_out, semantic_gt)
loss2 = lovasz_softmax(F.softmax(semantic_out, dim=1), semantic_gt, ignore=255)
loss3 = self.semantic_loss(semantic_out_2, semantic_gt)
loss = loss1 +loss2+loss3
return loss
def load_state_dict_from_file(file: str, only_state_dict=True):
file = os.path.realpath(os.path.expanduser(file))
checkpoint = torch.load(file, map_location="cpu")
if only_state_dict and "state_dict" in checkpoint:
checkpoint = checkpoint["state_dict"]
return checkpoint
def run(args):
torch.cuda.set_device(args.cuda)
#torch.cuda.set_device()
with open(args.config, 'r') as fp:
cfg = json.load(fp)
#logdir = f'run/{time.strftime("%Y-%m-%d-%H-%M")}-{cfg["dataset"]}-{cfg["model_name"]}-'
logdir = f'run/{cfg["dataset"]}-{cfg["model_name"]}-{cfg["ablation"]}'
if not os.path.exists(logdir):
os.makedirs(logdir)
shutil.copy(args.config, logdir)
logger = get_logger(logdir)
logger.info(f'Conf | use logdir {logdir}')
model = get_model(cfg)
device = torch.device(f'cuda:{args.cuda}')
model.to(device)
# weight = load_state_dict_from_file('./checkpoints/b1-r288.pt')
# #model.backbone1.load_state_dict(weight,strict=False)
# model.load_state_dict(weight,strict=False)
trainset, _, testset = get_dataset(cfg)
train_loader = DataLoader(trainset, batch_size=cfg['ims_per_gpu'], shuffle=True, num_workers=cfg['num_workers'],
pin_memory=True)
test_loader = DataLoader(testset, batch_size=cfg['ims_per_gpu'], shuffle=False, num_workers=cfg['num_workers'],
pin_memory=True)
params_list = model.parameters()
optimizer = Ranger(params_list, lr=cfg['lr_start'], weight_decay=cfg['weight_decay'])
scheduler = LambdaLR(optimizer, lr_lambda=lambda ep: (1 - ep / cfg['epochs']) ** 0.9)
train_criterion = eeemodelLoss().to(device)
criterion = nn.CrossEntropyLoss().to(device)
train_loss_meter = averageMeter()
test_loss_meter = averageMeter()
running_metrics_test = runningScore(cfg['n_classes'], ignore_index=cfg['id_unlabel'])
best_test = 0
best_iou = 0
#amp.register_float_function(torch, 'sigmoid')
#model, optimizer = amp.initialize(model, optimizer, opt_level=args.opt_level)
#model, optimizer = torch.cuda.amp.initialize(model, optimizer, opt_level=args.opt_level)
for ep in range(cfg['epochs']):
# training
model.train()
train_loss_meter.reset()
for i, sample in enumerate(train_loader):
optimizer.zero_grad()
image = sample['image'].to(device)
depth = sample['depth'].to(device)
label = sample['label'].to(device)
bound = sample['bound'].to(device)
binary_label = sample['binary_label'].to(device)
targets = [label, binary_label, bound]
predict = model(image, depth)
loss = train_criterion(predict, targets)
####################################################
#with amp.scale_loss(loss, optimizer) as scaled_loss:
#scaled_loss.backward()
loss.backward()
optimizer.step()
train_loss_meter.update(loss.item())
scheduler.step(ep)
# test
with torch.no_grad():
model.eval()
running_metrics_test.reset()
test_loss_meter.reset()
for i, sample in enumerate(test_loader):
image = sample['image'].to(device)
# Here, depth is TIR.
depth = sample['depth'].to(device)
label = sample['label'].to(device)
predict = model(image, depth)[0]
loss = criterion(predict, label)
test_loss_meter.update(loss.item())
predict = predict.max(1)[1].cpu().numpy() # [1, h, w]
label = label.cpu().numpy()
running_metrics_test.update(label, predict)
train_loss = train_loss_meter.avg
test_loss = test_loss_meter.avg
test_macc = running_metrics_test.get_scores()[0]["class_acc: "]
test_miou = running_metrics_test.get_scores()[0]["mIou: "]
test_avg = (test_macc + test_miou) / 2
logger.info(
f'Iter | [{ep + 1:3d}/{cfg["epochs"]}] loss={train_loss:.3f}/{test_loss:.3f}, mPA={test_macc:.3f}, miou={test_miou:.3f}, avg={test_avg:.3f}')
if test_avg > best_test:
best_test = test_avg
#best_iou = test_miou
save_ckpt(logdir, model,ep+1)
logger.info(
f'Save Iter = [{ep + 1:3d}], mPA={test_macc:.3f}, miou={test_miou:.3f}, avg={test_avg:.3f}')
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description="config")
parser.add_argument("--config", type=str, default="configs/base_vit.json", help="Configuration file to use")
parser.add_argument("--opt_level", type=str, default='O1')
parser.add_argument("--inputs", type=str.lower, default='rgb', choices=['rgb', 'rgbd'])
parser.add_argument("--resume", type=str, default='',
help="use this file to load last checkpoint for continuing training")
parser.add_argument("--cuda", type=int, default='0', help="set cuda device id")
args = parser.parse_args()
print("Starting Training!")
run(args)