-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
165 lines (143 loc) · 6.97 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
#!/usr/bin/env python
# -*- coding:utf-8 -*-
# author:Yifan Zhu
# datetime:2020/10/2 2:56
# file: train.py
# software: PyCharm
import os
import glob
import yaml
import numpy as np
import torch
import torch.optim as optim
from utils import dataset
from network.training import get_prior_z, build_network, train, input_encoder_param, joint_train
if __name__ == '__main__':
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print('device:', device)
CONFIG_PATH = 'configs/default.yaml'
with open(CONFIG_PATH, 'r') as f:
cfg = yaml.load(f, Loader=yaml.FullLoader)
# hyper-parameters
num_epochs = cfg['training']['epochs']
points_batch = cfg['training']['subsamples_each_step']
batch_size_shapenet = cfg['training']['batch_size_shapenet']
batch_size_kitti = cfg['training']['batch_size_kitti']
lr = cfg['training']['lr']
use_lr_schedule = cfg['training']['lr_schedule']
retrieve_model = cfg['training']['retrieve_model']
use_eik = cfg['model']['use_eik']
variational = cfg['model']['variational']
use_kl = cfg['model']['use_kl']
eik_weight = cfg['model']['eik_weight']
vae_weight = cfg['model']['vae_weight']
z_dim = cfg['model']['z_dim']
geo_initial = cfg['training']['geo_initial']
use_normal = cfg['training']['use_normal']
enforce_symmetry = cfg['training']['enforce_symmetry']
skip_connection = cfg['model']['skip_connection']
input_mapping = cfg['training']['input_mapping']
embedding_method = cfg['training']['embedding_method']
beta = cfg['model']['beta']
partial_input = cfg['data']['partial_input']
data_completeness = cfg['data']['data_completeness']
data_sparsity = cfg['data']['data_sparsity']
# save folder
save_fold = cfg['dir']['save_fold']
os.makedirs('models' + save_fold, exist_ok=True)
# save config file
f = open('models' + save_fold + '/config.yaml', "w")
yaml.dump(cfg, f)
f.close()
# input mapping
args = ()
if input_mapping:
args = input_encoder_param(input_mapping, embedding_method, device)
# build network
p0_z = get_prior_z(device, z_dim=z_dim)
net = build_network(*args, input_dim=3, p0_z=p0_z, z_dim=z_dim, beta=beta, skip_connection=skip_connection,
variational=variational, use_kl=use_kl, geo_initial=geo_initial)
if retrieve_model:
model_path = cfg['training']['retrieve_path']
checkpoint = cfg['training']['checkpoint']
saved_model_state = torch.load('models' + model_path + '/model_{}.pth'.format(checkpoint), map_location='cpu')
net.load_state_dict({k.replace('module.', ''): v for k, v in saved_model_state.items()})
# set multi-gpu if available
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
net = torch.nn.DataParallel(net)
net.to(device)
num_params = sum(p.numel() for p in net.parameters() if p.requires_grad)
print('The number of parameters of model is', num_params)
# create dataloader
# ShapeNet
DATA_PATH = cfg['data']['path']
fields = {
'inputs': dataset.PointCloudField(cfg['data']['pointcloud_file'])
}
category = cfg['data']['classes']
shapenet_dataset = dataset.ShapenetDataset(dataset_folder=DATA_PATH, fields=fields, categories=category,
split='train', with_normals=use_normal, points_batch=points_batch,
partial_input=partial_input, data_completeness=data_completeness,
data_sparsity=data_sparsity)
shapenet_loader = torch.utils.data.DataLoader(
shapenet_dataset, batch_size=batch_size_shapenet, num_workers=0, shuffle=True, drop_last=True, pin_memory=True)
# KITTI
kitti_dataset = dataset.KITTI360Dataset(cfg['data']['kitti_pcl_path'], 'train',
cfg['data']['kitti_class'], points_batch)
kitti_loader = torch.utils.data.DataLoader(kitti_dataset, batch_size=batch_size_kitti, num_workers=0, shuffle=True,
drop_last=True, pin_memory=True)
# create optimizer
optimizer = optim.Adam(net.parameters(), lr=lr)
print(optimizer)
print("Training!")
avg_training_loss = []
rec_training_loss = []
eik_training_loss = []
vae_training_loss = []
for epoch in range(num_epochs):
if use_lr_schedule:
if epoch % 500 == 0 and epoch >= 1000:
for param_group in optimizer.param_groups:
param_group['lr'] = lr / (2 ** (epoch // 500 - 1))
print(optimizer)
if cfg['data']['dataset'] == 'shapenet':
if epoch == 0:
print("Train on shapenet!")
print('shapenet objects:', len(shapenet_dataset), category)
avg_loss, rec_loss, eik_loss, vae_loss = train(net, shapenet_loader, optimizer, device, eik_weight,
vae_weight, use_normal, use_eik, enforce_symmetry)
elif cfg['data']['dataset'] == 'kitti':
if epoch == 0:
print("Train on kitti!")
print('kitti objects:', len(kitti_dataset), cfg['data']['kitti_class'])
avg_loss, rec_loss, eik_loss, vae_loss = train(net, kitti_loader, optimizer, device, eik_weight, vae_weight,
False, use_eik, enforce_symmetry)
else:
if epoch == 0:
print("Joint training!")
print('shapenet objects:', len(shapenet_dataset), category, 'kitti objects:', len(kitti_dataset),
cfg['data']['kitti_class'])
avg_loss, rec_loss, eik_loss, vae_loss = joint_train(net, shapenet_loader, kitti_loader, optimizer, device,
eik_weight, vae_weight, use_normal, use_eik,
enforce_symmetry, cfg['data']['kitti_weight'])
avg_training_loss.append(avg_loss)
rec_training_loss.append(rec_loss)
eik_training_loss.append(eik_loss)
vae_training_loss.append(vae_loss)
print('Epoch [%d / %d] average training loss: %f rec loss: %f eik loss: %f vae loss %f' % (
epoch + 1, num_epochs, avg_loss, rec_loss, eik_loss, vae_loss))
if epoch % 100 == 0 and epoch:
torch.save(net.state_dict(), 'models' + save_fold + '/model_{0:04d}.pth'.format(epoch))
torch.save(net.state_dict(), 'models' + save_fold + '/model_final.pth')
# plot loss
import matplotlib.pyplot as plt
fig = plt.figure()
p1, = plt.plot(avg_training_loss)
p2, = plt.plot(rec_training_loss)
p3, = plt.plot(eik_training_loss)
p4, = plt.plot(vae_training_loss)
plt.legend([p1, p2, p3, p4], ["total_loss", "rec_loss", "eik_loss", "vae_loss"])
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.savefig('models' + save_fold + '/loss.png')