-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_coda_small_kitti.py
193 lines (153 loc) · 8.35 KB
/
train_coda_small_kitti.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
# -*- coding:utf-8 -*-
# author: Xinge
# @file: train_cylinder_asym.py
import os
import time
import argparse
import sys
import numpy as np
import torch
import torch.optim as optim
from tqdm import tqdm
from utils.metric_util import per_class_iu, fast_hist_crop
from dataloader.pc_dataset import get_SemKITTI_label_name
from builder import data_builder, model_builder, loss_builder
from config.config import load_config_data
from utils.load_save_util import load_remapped_checkpoint
import warnings
warnings.filterwarnings("ignore")
def main(args):
pytorch_device = torch.device('cuda:0')
config_path = args.config_path
configs = load_config_data(config_path)
dataset_config = configs['dataset_params'] # Use coda_kitti_test_subset.yaml
train_dataloader_config = configs['train_data_loader']
val_dataloader_config = configs['val_data_loader']
val_batch_size = val_dataloader_config['batch_size']
train_batch_size = train_dataloader_config['batch_size']
model_config = configs['model_params']
train_hypers = configs['train_params']
grid_size = model_config['output_shape']
num_class = model_config['num_class']
ignore_label = dataset_config['ignore_label']
model_load_path = train_hypers['model_load_path']
model_save_path = train_hypers['model_save_path']
wd = train_hypers['weight_decay'] # weight decay
amp = train_hypers['mixed_fp16']
SemKITTI_label_name = get_SemKITTI_label_name(dataset_config["label_mapping"])
unique_label = np.asarray(sorted(list(SemKITTI_label_name.keys())))[1:] - 1
unique_label_str = [SemKITTI_label_name[x] for x in unique_label + 1]
print(unique_label_str)
my_model = model_builder.build(model_config)
if os.path.exists(model_load_path):
my_model = load_remapped_checkpoint(model_load_path, my_model)
my_model.to(pytorch_device)
# Mixed Precision
amp_scaler = torch.cuda.amp.GradScaler(enabled=amp)
# EPS is a fixed value to prevent MixedPrecision training errors.
optimizer = optim.AdamW(my_model.parameters(), lr=train_hypers["learning_rate"], eps=1e-4, weight_decay=wd)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, train_hypers['max_num_epochs'])
loss_func, lovasz_softmax = loss_builder.build(wce=True, lovasz=True,
num_class=num_class, ignore_label=ignore_label)
train_dataset_loader, val_dataset_loader = data_builder.build(dataset_config,
train_dataloader_config,
val_dataloader_config,
grid_size=grid_size)
# training
epoch = 0
best_val_miou = 0
my_model.train()
global_iter = 0
check_iter = train_hypers['eval_every_n_steps']
while epoch < train_hypers['max_num_epochs']:
loss_list = []
pbar = tqdm(total=len(train_dataset_loader))
# time.sleep(10)
# lr_scheduler.step(epoch)
for i_iter, (_, train_vox_label, train_grid, _, train_pt_fea) in enumerate(train_dataset_loader):
if global_iter % check_iter == 0:
# Evaluation set
my_model.eval()
hist_list = []
val_loss_list = []
with torch.no_grad():
for i_iter_val, (_, val_vox_label, val_grid, val_pt_labs, val_pt_fea) in enumerate(
val_dataset_loader):
val_pt_fea_ten = [torch.from_numpy(i).type(torch.FloatTensor).to(pytorch_device) for i in
val_pt_fea]
val_grid_ten = [torch.from_numpy(i).to(pytorch_device) for i in val_grid]
val_label_tensor = val_vox_label.type(torch.LongTensor).to(pytorch_device)
predict_labels = my_model(val_pt_fea_ten, val_grid_ten, val_batch_size)
# remap outputs
# output_remap = {0: 0, 1: 0, 2: 0, 3: 0, 4: 0, 5: 0, 6: 0, 7: 0, 8: 0, 9: 11, 10: 0, 11: 12, 12: 4, 13: 0, 14: 0, 15: 13, 16: 0, 17: 10, 18: 0, 19: 0, 20: 1, 21: 2, 22: 3, 23: 5, 24: 6, 25: 7, 26: 8, 27: 9, 28: 14}
# predict_labels.apply_(output_remap.get)
# aux_loss = loss_fun(aux_outputs, point_label_tensor)
loss = lovasz_softmax(torch.nn.functional.softmax(predict_labels).detach(), val_label_tensor,
ignore=0) + loss_func(predict_labels.detach(), val_label_tensor)
predict_labels = torch.argmax(predict_labels, dim=1)
predict_labels = predict_labels.cpu().detach().numpy()
for count, i_val_grid in enumerate(val_grid):
hist_list.append(fast_hist_crop(predict_labels[
count, val_grid[count][:, 0], val_grid[count][:, 1],
val_grid[count][:, 2]], val_pt_labs[count],
unique_label))
val_loss_list.append(loss.detach().cpu().numpy())
my_model.train()
iou = per_class_iu(sum(hist_list))
print('Validation per class iou: ')
for class_name, class_iou in zip(unique_label_str, iou):
print('%s : %.2f%%' % (class_name, class_iou * 100))
val_miou = np.nanmean(iou) * 100
del val_vox_label, val_grid, val_pt_fea, val_grid_ten
# save model if performance is improved
if best_val_miou < val_miou:
best_val_miou = val_miou
torch.save(my_model.state_dict(), model_save_path)
print('Current val miou is %.3f while the best val miou is %.3f' %
(val_miou, best_val_miou))
print('Current val loss is %.3f' %
(np.mean(val_loss_list)))
# Training set
train_pt_fea_ten = [torch.from_numpy(i).type(torch.FloatTensor).to(pytorch_device) for i in train_pt_fea]
# train_grid_ten = [torch.from_numpy(i[:,:2]).to(pytorch_device) for i in train_grid]
train_vox_ten = [torch.from_numpy(i).to(pytorch_device) for i in train_grid]
point_label_tensor = train_vox_label.type(torch.LongTensor).to(pytorch_device)
# forward + backward + optimize
with torch.cuda.amp.autocast(enabled=amp):
outputs = my_model(train_pt_fea_ten, train_vox_ten, train_batch_size)
# remap outputs
# output_remap = {0: 0, 1: 0, 2: 0, 3: 0, 4: 0, 5: 0, 6: 0, 7: 0, 8: 0, 9: 11, 10: 0, 11: 12, 12: 4, 13: 0, 14: 0, 15: 13, 16: 0, 17: 10, 18: 0, 19: 0, 20: 1, 21: 2, 22: 3, 23: 5, 24: 6, 25: 7, 26: 8, 27: 9, 28: 14}
# outputs.apply_(output_remap.get)
loss = lovasz_softmax(torch.nn.functional.softmax(outputs), point_label_tensor, ignore=0) + loss_func(
outputs, point_label_tensor)
amp_scaler.scale(loss).backward()
amp_scaler.step(optimizer)
amp_scaler.update()
# optimizer.step()
loss_list.append(loss.item())
if global_iter % 1000 == 0:
if len(loss_list) > 0:
print('epoch %d iter %5d, loss: %.3f\n' %
(epoch, i_iter, np.mean(loss_list)))
else:
print('loss error')
optimizer.zero_grad()
pbar.update(1)
global_iter += 1
if global_iter % check_iter == 0:
if len(loss_list) > 0:
print('epoch %d iter %5d, loss: %.3f\n' %
(epoch, i_iter, np.mean(loss_list)))
else:
print('loss error')
scheduler.step() # Updates cosine annealing
pbar.close()
epoch += 1
if __name__ == '__main__':
# Training settings
parser = argparse.ArgumentParser(description='')
parser.add_argument('-y', '--config_path', default='config/coda_kitti_test_subset.yaml')
args = parser.parse_args()
print(' '.join(sys.argv))
print(args)
main(args)