-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
199 lines (180 loc) · 8.69 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
import argparse
import tensorflow as tf
import numpy as np
import importlib
from tqdm import tqdm
import os
import sys
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(BASE_DIR)
sys.path.append(os.path.join(BASE_DIR, 'utils'))
sys.path.append(os.path.join(BASE_DIR, 'data_loader'))
from data_generator import *
from deeplidarflow import DeepLiDARFlowNet, get_loss_KITTI, get_loss_FT3D, get_eval
from checkpointsaver import BestCheckpointSaver, get_best_checkpoint
from distutils import util
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0, help='GPU to use [default: GPU 0]')
parser.add_argument('--dataset', default='FT3D', help='Dataset type [default: FT3D]')
parser.add_argument('--data_path', default='./FlyingThings3D', help='Dataset directory [default: ./FlyingThings3D]')
parser.add_argument('--batch_size', type=int, default=2, help='batch size [default: 2]')
parser.add_argument('--max_epoch', type=int, default=650, help='max epoch [default: 650]')
parser.add_argument('--learning_rate', type=float, default=0.0001, help='learning rate [default: 0.0001]')
parser.add_argument('--best_checkpoint', type=util.strtobool, default=False, help='model best checkpoint [default: False]')
parser.add_argument('--pretrained_model', default=None, help='pretrained model to fine tune [default: None]')
FLAGS = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(FLAGS.gpu)
if FLAGS.dataset == 'KITTI':
STEP_SIZE = 1
else:
STEP_SIZE = 5
def eval_one_step(dataset, BATCH_SIZE, ops, sess, handle_data, writer, epoch, step):
data_size = len(dataset)
total_loss = 0.
total_sf_keo = 0.
total_d0_keo = 0.
total_d1_keo = 0.
total_fl_keo = 0.
total_sf_epe = 0.
count = 0.
feed_dict = {ops['handle_pl']: handle_data}
for _ in tqdm(range(data_size // BATCH_SIZE // STEP_SIZE)):
summary, batch_loss, eval_kitti = sess.run([ops['summary_pl'], ops['loss_pl'], ops['eval_kitti_pl']], feed_dict=feed_dict)
writer.add_summary(summary, step)
writer.flush()
total_loss += batch_loss
total_d0_keo += eval_kitti[0]
total_d1_keo += eval_kitti[1]
total_fl_keo += eval_kitti[2]
total_sf_keo += eval_kitti[3]
total_sf_epe += eval_kitti[4]
count += 1.
mean_loss = total_loss / count
mean_sf_keo = total_sf_keo / count
mean_sf_epe = total_sf_epe / count
tqdm.write('\n Valid: epoch/step: {:d}/{:d}, loss:{:.2f}, koe:{:.2f}, epe:{:.2f}'.format(epoch, step, mean_loss, mean_sf_keo, mean_sf_epe))
return mean_sf_keo
def train_one_step(dataset, BATCH_SIZE, ops, sess, handle_data, writer, epoch, step):
data_size = len(dataset)
total_loss = 0.
total_sf_keo = 0.
total_d0_keo = 0.
total_d1_keo = 0.
total_fl_keo = 0.
total_sf_epe = 0.
count = 0.
feed_dict = {ops['handle_pl']: handle_data}
for _ in tqdm(range(data_size // BATCH_SIZE // STEP_SIZE)):
_,summary, batch_loss, eval_kitti = sess.run([ops['train_op'], ops['summary_pl'], ops['loss_pl'], ops['eval_kitti_pl']], feed_dict=feed_dict)
writer.add_summary(summary, step)
writer.flush()
total_loss += batch_loss
total_d0_keo += eval_kitti[0]
total_d1_keo += eval_kitti[1]
total_fl_keo += eval_kitti[2]
total_sf_keo += eval_kitti[3]
total_sf_epe += eval_kitti[4]
count += 1.
mean_loss = total_loss / count
mean_sf_keo = total_sf_keo / count
mean_sf_epe = total_sf_epe / count
tqdm.write('\n Train: epoch/step: {:d}/{:d}, loss:{:.2f}, koe:{:.2f}, epe:{:.2f}'.format(epoch, step, mean_loss, mean_sf_keo, mean_sf_epe))
return mean_sf_keo
def load_model(sess, weights_dirctory = False):
saver = tf.train.Saver(name='saver', max_to_keep=100)
if weights_dirctory:
print('load best model...')
# make sure to include it in the checpoints/best_checkpoints JSON file!!
if not os.path.exists('best_checkpoints'):
raise(ValueError("No best check points available \n"))
saver.restore(sess, get_best_checkpoint('model', select_maximum_value=False))
elif FLAGS.pretrained_model is not None:
if not os.path.exists(FLAGS.pretrained_model + '.meta'):
raise(ValueError("No pretrained model available \n"))
print('load pretrained model...')
saver.restore(sess, FLAGS.pretrained_model)
else:
print('Initializing new model...')
sess.run(tf.global_variables_initializer())
return True
def train(data_type, dataset_train, dataset_val, BATCH_SIZE, MAX_EPOCH, SAMPLES, LEARNING_RATE):
with tf.device('/gpu:'+str(FLAGS.gpu)):
with tf.Graph().as_default():
handle, inputs, one_shot_train = init_input_pipeline(data_type, dataset_train, BATCH_SIZE, SAMPLES, True)
_, _, one_shot_val = init_input_pipeline(data_type, dataset_val, BATCH_SIZE, SAMPLES)
images = inputs[0]
gt_sf = inputs[1]
gt_shape = inputs[2][0]
if len(inputs) == 4:
interp_shape = inputs[3][0]
else:
interp_shape = None
out_sf_list, out_sf = DeepLiDARFlowNet(images, gt_shape, interp_shape)
if FLAGS.dataset == 'KITTI':
loss = get_loss_KITTI(out_sf_list, gt_sf, gt_shape)
else:
loss = get_loss_FT3D(out_sf_list, gt_sf)
eval_kitti = get_eval(out_sf, gt_sf)
# preparing the training params
with tf.variable_scope('train'):
reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
loss_with_l2 = tf.add_n([loss] + reg_losses, name='loss_with_L2')
global_step = tf.Variable(-1, dtype=tf.int32, trainable=False, name='global_step')
increment_global_step = tf.assign_add(global_step, 1, name='increment_global_step')
optimizer = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, name='Adam_optimizer')
train_op = optimizer.minimize(loss_with_l2, name='train_op')
# preparing the summary objects
with tf.variable_scope('summaries'):
tf.summary.scalar('loss', loss)
tf.summary.scalar('outliers', eval_kitti[3])
tf.summary.scalar('epe', eval_kitti[4])
summary = tf.summary.merge_all()
saver = tf.train.Saver(name='saver',max_to_keep=100)
# Create a session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.log_device_placement = False
sess = tf.Session(config=config)
handle_train = sess.run(one_shot_train.string_handle())
handle_val = sess.run(one_shot_val.string_handle())
# summary writers
writer_train = tf.summary.FileWriter(os.path.join('summary', 'train'), sess.graph)
writer_valid = tf.summary.FileWriter(os.path.join('summary', 'valid'), sess.graph)
load_model(sess, weights_dirctory=FLAGS.best_checkpoint)
best_ckpt_saver = BestCheckpointSaver(save_dir='model', num_to_keep=5, maximize=False, saver=saver)
ops = {'handle_pl': handle,
'images_pl': images,
'gt_sf_pl': gt_sf,
'gt_shape_pl': gt_shape,
'train_op':train_op,
'out_sf_list': out_sf_list,
'out_sf_pl': out_sf,
'loss_pl': loss,
'eval_kitti_pl': eval_kitti,
'summary_pl': summary}
for epoch in range(MAX_EPOCH):
count = 0
name = 'DeepLiDARFlow-' + FLAGS.dataset + '-epoch-' + str(epoch) + '-step-'
while count < STEP_SIZE:
step = sess.run(increment_global_step)
train_one_step(dataset_train, BATCH_SIZE, ops, sess, handle_train, writer_train, epoch, step)
sf_keo = eval_one_step(dataset_val, BATCH_SIZE, ops, sess, handle_val, writer_valid, epoch, step)
# Save the variables to disk.
best_ckpt_saver.handle(sf_keo, sess, step, name)
count += 1
if __name__ == '__main__':
DATAPATH = FLAGS.data_path
DATASET = importlib.import_module(FLAGS.dataset)
SAMPLES = np.array([100, 500, 1000, 5000, 10000])
dataset_train = DATASET.SceneFlow(
root= DATAPATH,
mode='TRAIN')
dataset_val = DATASET.SceneFlow(
root= DATAPATH,
mode='VAL')
# Hyperparameters
BATCH_SIZE = FLAGS.batch_size
MAX_EPOCH = FLAGS.max_epoch
DATA_TYPE = FLAGS.dataset
LEARNING_RATE = FLAGS.learning_rate
train(DATA_TYPE, dataset_train, dataset_val, BATCH_SIZE, MAX_EPOCH, SAMPLES, LEARNING_RATE)