Newer
Older
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
import argparse
import json
import os
from tqdm import tqdm
import numpy as np
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import torch.nn.functional as F
torch.backends.cudnn.benchmark = True
from model import CILRS
from data import CARLA_Data
from config import GlobalConfig
parser = argparse.ArgumentParser()
parser.add_argument('--id', type=str, default='cilrs', help='Unique experiment identifier.')
parser.add_argument('--device', type=str, default='cuda', help='Device to use')
parser.add_argument('--epochs', type=int, default=201, help='Number of train epochs.')
parser.add_argument('--val_every', type=int, default=5, help='Validation frequency (epochs).')
parser.add_argument('--batch_size', type=int, default=24, help='Batch size')
parser.add_argument('--logdir', type=str, default='log', help='Directory to log data to.')
args = parser.parse_args()
args.logdir = os.path.join(args.logdir, args.id)
writer = SummaryWriter(log_dir=args.logdir)
class Engine(object):
"""Engine that runs training and inference.
Args
- cur_epoch (int): Current epoch.
- print_every (int): How frequently (# batches) to print loss.
- validate_every (int): How frequently (# epochs) to run validation.
"""
def __init__(self, cur_epoch=0, bestval_epoch=0):
self.cur_epoch = cur_epoch
self.bestval_epoch = bestval_epoch
self.train_loss = []
self.val_loss = []
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
def train(self):
loss_epoch = 0.
num_batches = 0
model.train()
# Train loop
for data in tqdm(dataloader_train):
# efficiently zero gradients
for p in model.parameters():
p.grad = None
# create batch and move to GPU
fronts_in = data['fronts']
lefts_in = data['lefts']
rights_in = data['rights']
rears_in = data['rears']
fronts = []
lefts = []
rights = []
rears = []
for i in range(config.seq_len):
fronts.append(fronts_in[i].to(args.device, dtype=torch.float32))
if not config.ignore_sides:
lefts.append(lefts_in[i].to(args.device, dtype=torch.float32))
rights.append(rights_in[i].to(args.device, dtype=torch.float32))
if not config.ignore_rear:
rears.append(rears_in[i].to(args.device, dtype=torch.float32))
# driving labels
command = data['command'].to(args.device)
gt_velocity = data['velocity'].to(args.device, dtype=torch.float32)
gt_steer = data['steer'].to(args.device, dtype=torch.float32)
gt_throttle = data['throttle'].to(args.device, dtype=torch.float32)
gt_brake = data['brake'].to(args.device, dtype=torch.float32)
# inference
encoding = [model.encoder(fronts)]
if not config.ignore_sides:
encoding.append(model.encoder(lefts))
encoding.append(model.encoder(rights))
if not config.ignore_rear:
encoding.append(model.encoder(rears))
steer, throttle, brake, velocity = model(encoding, gt_velocity, command)
# losses
loss = 0.05 * F.l1_loss(velocity.squeeze(), gt_velocity).mean()
loss += F.l1_loss(steer.squeeze(), gt_steer.squeeze()).mean()
loss += F.l1_loss(throttle.squeeze(), gt_throttle.squeeze()).mean()
loss += F.l1_loss(brake.squeeze(), gt_brake.squeeze()).mean()
loss.backward()
loss_epoch += float(loss.item())
num_batches += 1
optimizer.step()
loss_epoch = loss_epoch / num_batches
self.train_loss.append(loss_epoch)
self.cur_epoch += 1
def validate(self):
model.eval()
with torch.no_grad():
num_batches = 0
vel_epoch = 0.
steer_epoch = 0.
throttle_epoch = 0.
brake_epoch = 0.
# Validation loop
for batch_num, data in enumerate(tqdm(dataloader_val), 0):
# create batch and move to GPU
fronts_in = data['fronts']
lefts_in = data['lefts']
rights_in = data['rights']
rears_in = data['rears']
fronts = []
lefts = []
rights = []
rears = []
for i in range(config.seq_len):
fronts.append(fronts_in[i].to(args.device, dtype=torch.float32))
if not config.ignore_sides:
lefts.append(lefts_in[i].to(args.device, dtype=torch.float32))
rights.append(rights_in[i].to(args.device, dtype=torch.float32))
if not config.ignore_rear:
rears.append(rears_in[i].to(args.device, dtype=torch.float32))
# driving labels
command = data['command'].to(args.device)
gt_velocity = data['velocity'].to(args.device, dtype=torch.float32)
gt_steer = data['steer'].to(args.device, dtype=torch.float32)
gt_throttle = data['throttle'].to(args.device, dtype=torch.float32)
gt_brake = data['brake'].to(args.device, dtype=torch.float32)
# inference
encoding = [model.encoder(fronts)]
encoding.append(model.encoder(lefts))
encoding.append(model.encoder(rights))
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
encoding.append(model.encoder(rears))
steer, throttle, brake, velocity = model(encoding, gt_velocity, command)
# losses
vel_epoch += float(F.l1_loss(velocity.squeeze(), gt_velocity.squeeze()).mean())
steer_epoch += float(F.l1_loss(steer.squeeze(), gt_steer.squeeze()).mean())
throttle_epoch += float(F.l1_loss(throttle.squeeze(), gt_throttle.squeeze()).mean())
brake_epoch += float(F.l1_loss(brake.squeeze(), gt_brake.squeeze()).mean())
num_batches += 1
vel_loss = vel_epoch / float(num_batches)
steer_loss = steer_epoch / float(num_batches)
throttle_loss = throttle_epoch / float(num_batches)
brake_loss = brake_epoch / float(num_batches)
tqdm.write(f'Epoch {self.cur_epoch:03d}, Batch {batch_num:03d}:' +
f' Vel: {vel_loss:3.3f} Str: {steer_loss:3.3f} Thr: {throttle_loss:3.3f} Brk: {brake_loss:3.3f}')
self.val_loss.append(0.05 * vel_loss + steer_loss + throttle_loss + brake_loss)
def save(self):
save_best = False
if self.val_loss[-1] <= self.bestval:
self.bestval = self.val_loss[-1]
self.bestval_epoch = self.cur_epoch
save_best = True
# Create a dictionary of all data to save
log_table = {
'epoch': self.cur_epoch,
'bestval': self.bestval,
'bestval_epoch': self.bestval_epoch,
'train_loss': self.train_loss,
'val_loss': self.val_loss,
}
# Save the recent model/optimizer states
torch.save(model.state_dict(), os.path.join(args.logdir, 'model.pth'))
torch.save(optimizer.state_dict(), os.path.join(args.logdir, 'recent_optim.pth'))
# Log other data corresponding to the recent model
with open(os.path.join(args.logdir, 'recent.log'), 'w') as f:
f.write(json.dumps(log_table))
tqdm.write('====== Saved recent model ======>')
if save_best:
torch.save(model.state_dict(), os.path.join(args.logdir, 'best_model.pth'))
torch.save(optimizer.state_dict(), os.path.join(args.logdir, 'best_optim.pth'))
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
tqdm.write('====== Overwrote best model ======>')
# Config
config = GlobalConfig()
# Data
train_set = CARLA_Data(root=config.train_data, config=config)
val_set = CARLA_Data(root=config.val_data, config=config)
dataloader_train = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True)
dataloader_val = DataLoader(val_set, batch_size=args.batch_size, shuffle=False, num_workers=8, pin_memory=True)
# Model
model = CILRS(config, args.device)
optimizer = optim.AdamW(model.parameters(), lr=config.lr)
trainer = Engine()
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print ('Total trainable parameters: ', params)
# Create logdir
if not os.path.isdir(args.logdir):
os.makedirs(args.logdir)
print ('Created dir:', args.logdir)
elif os.path.isfile(os.path.join(args.logdir, 'recent.log')):
print ('Loading checkpoint from ' + args.logdir)
with open(os.path.join(args.logdir, 'recent.log'), 'r') as f:
log_table = json.load(f)
# Load variables
trainer.cur_epoch = log_table['epoch']
trainer.bestval = log_table['bestval']
trainer.bestval_epoch = log_table['bestval_epoch']
trainer.train_loss = log_table['train_loss']
trainer.val_loss = log_table['val_loss']
# Load checkpoint
model.load_state_dict(torch.load(os.path.join(args.logdir, 'model.pth')))
optimizer.load_state_dict(torch.load(os.path.join(args.logdir, 'recent_optim.pth')))
# Log args
with open(os.path.join(args.logdir, 'args.txt'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
for epoch in range(trainer.cur_epoch, args.epochs):
trainer.train()
if epoch % args.val_every == 0:
trainer.validate()
trainer.save()