-
Notifications
You must be signed in to change notification settings - Fork 3
/
training.py
225 lines (179 loc) · 8.19 KB
/
training.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
213
214
215
216
217
218
219
220
221
222
223
224
225
import csv
import argparse
import time
import sys
import torch
from torch import optim
import torch.distributed as dist
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader
from waymo_open_dataset.protos import occupancy_flow_metrics_pb2
from google.protobuf import text_format
from model.InteractPlanner import InteractPlanner
from utils.net_utils import *
from metric import TrainingMetrics, ValidationMetrics
# define model training epoch
def training_epoch(train_data, optimizer, epoch, scheduler):
model.train()
current = 0
start_time = time.time()
size = len(train_data)
epoch_loss = []
train_metric = TrainingMetrics()
i = 0
for batch in train_data:
# prepare data
inputs, target = batch_to_dict(batch, local_rank , use_flow=use_flow)
optimizer.zero_grad()
# query the model
bev_pred, traj, score = model(inputs)
actor_loss, occ_loss, flow_loss = occupancy_loss(bev_pred, target, use_flow=use_flow)
il_loss, _, gt_modes = imitation_loss(traj, score, target, args.use_planning)
loss = il_loss + 100*(actor_loss + occ_loss)
if use_flow:
loss += flow_loss
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
optimizer.step()
current += args.batch_size
epoch_loss.append(loss.item())
if isinstance(traj, list):
traj, score = traj[-1], score[-1]
ade, fde, l_il, l_ogm = train_metric.update(traj, score, gt_modes, target, il_loss, actor_loss + occ_loss, bev_pred)
if dist.get_rank() == 0:
sys.stdout.write(f"\rTrain: [{current:>6d}/{size*args.batch_size:>6d}]|Loss: {np.mean(epoch_loss):>.4f}-{l_il:>.4f}-{l_ogm:>.4f}|ADE:{ade:>.4f}-FDE:{fde:>.4f}|{(time.time()-start_time)/current:>.4f}s/sample")
sys.stdout.flush()
scheduler.step(epoch + i/size)
i += 1
results = train_metric.result()
return np.mean(epoch_loss), results
# define model validation epoch
def validation_epoch(valid_data,epoch):
epoch_metrics = ValidationMetrics()
model.eval()
current = 0
start_time = time.time()
size = len(valid_data)
epoch_loss = []
print(f'Validation...Epoch{epoch+1}')
for batch in valid_data:
# prepare data
inputs, target = batch_to_dict(batch, local_rank, use_flow=use_flow)
# query the model
with torch.no_grad():
bev_pred, traj, score = model(inputs)
actor_loss, occ_loss, flow_loss = occupancy_loss(bev_pred, target, use_flow=use_flow)
il_loss, _, gt_modes = imitation_loss(traj, score, target, args.use_planning)
loss = il_loss + 100*(actor_loss + occ_loss)
if use_flow:
loss += flow_loss
# compute metrics
epoch_loss.append(loss.item())
if isinstance(traj, list):
traj, score = traj[-1], score[-1]
ade, fde, l_il, l_ogm, ogm_auc,_, occ_auc,_ = epoch_metrics.update(traj, score, bev_pred,
gt_modes, target, il_loss, actor_loss + occ_loss)
current += args.batch_size
if dist.get_rank() == 0:
sys.stdout.write(f"\r\Val: [{current:>6d}/{size*args.batch_size:>6d}]|Loss: {np.mean(epoch_loss):>.4f}-{l_il:>.4f}-{l_ogm:>.4f}|ADE:{ade:>.4f}-FDE:{fde:>.4f}{(time.time()-start_time)/current:>.4f}s/sample")
sys.stdout.flush()
# process metrics
epoch_metrics = epoch_metrics.result()
return epoch_metrics,np.mean(epoch_loss)
# Define model training process
def model_training(train_data, valid_data, epochs, save_dir):
# define optimizer and loss function
optimizer = optim.AdamW(model.parameters(), lr=args.lr)
scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=2, eta_min=1e-6)
for epoch in range(epochs):
if dist.get_rank() == 0:
print(f"Epoch {epoch+1}/{epochs}")
if epoch<=continue_ep and continue_ep!=0:
scheduler.step()
continue
train_data.sampler.set_epoch(epoch)
valid_data.sampler.set_epoch(epoch)
train_loss,train_res = training_epoch(train_data, optimizer, epoch, scheduler)
valid_metrics,val_loss = validation_epoch(valid_data,epoch)
# save to training log
log = {'epoch': epoch+1, 'loss': train_loss, 'lr': optimizer.param_groups[0]['lr']}
log.update(valid_metrics)
if dist.get_rank() == 0:
if epoch == 0:
with open(save_dir + f'train_log.csv', 'a') as csv_file:
writer = csv.writer(csv_file)
writer.writerow(log.keys())
writer.writerow(log.values())
else:
with open(save_dir + f'train_log.csv', 'a') as csv_file:
writer = csv.writer(csv_file)
writer.writerow(log.values())
# adjust learning rate
scheduler.step()
# save model at the end of epoch
if dist.get_rank() == 0:
torch.save(model.state_dict(), save_dir+f'model_{epoch+1}_{train_loss:4f}_{val_loss:4f}.pth')
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int)
parser.add_argument("--batch_size", type=int, default=8)
parser.add_argument("--dim", type=int, default=256)
parser.add_argument("--lr", type=float, default=1e-4)
parser.add_argument("--epochs", type=int, default=30)
parser.add_argument("--use_flow", type=bool, action='store_true', default=True,
help='whether to use flow warp')
parser.add_argument("--save_dir", type=str, default='',help='path to save logs')
parser.add_argument("--data_dir", type=str, default='',
help='path to load preprocessed train & val sets')
parser.add_argument("--model_dir", type=str, default='',
help='path to load models for continue training')
args = parser.parse_args()
local_rank = args.local_rank
use_flow = args.use_flow
config = occupancy_flow_metrics_pb2.OccupancyFlowTaskConfig()
config_text = f"""
num_past_steps: {10}
num_future_steps: {50}
num_waypoints: {5}
cumulative_waypoints: {'false'}
normalize_sdc_yaw: true
grid_height_cells: {128}
grid_width_cells: {128}
sdc_y_in_grid: {int(128*0.75)}
sdc_x_in_grid: {64}
pixels_per_meter: {1.6}
agent_points_per_side_length: 48
agent_points_per_side_width: 16
"""
text_format.Parse(config_text, config)
model = InteractPlanner(config, dim=args.dim, enc_layer=2, heads=8, dropout=0.1,
timestep=5, decoder_dim=384, fpn_len=2, flow_pred=use_flow)
save_dir = args.save_dir + f"models/"
os.makedirs(save_dir,exist_ok=True)
torch.cuda.set_device(local_rank)
dist.init_process_group(backend='nccl')
model = model.to(local_rank)
if args.model_dir!= '':
kw_dict = {}
for k,v in torch.load(save_dir + args.load_dir,map_location='cpu').items():
kw_dict[k[7:]] = v
model.load_state_dict(kw_dict)
continue_ep = int(args.load_dir.split('_')[-3]) - 1
print(f'model loaded!:epoch {continue_ep + 1}')
else:
continue_ep = 0
model = DDP(model, device_ids=[local_rank], output_device=local_rank)
train_dataset = DrivingData(args.data_dir + f'train/*.npz',use_flow=use_flow)
valid_dataset = DrivingData(args.data_dir + f'valid/*.npz',use_flow=use_flow)
training_size = len(train_dataset)
valid_size = len(valid_dataset)
if dist.get_rank() == 0:
print(f'Length train: {training_size}; Valid: {valid_size}')
train_sampler = DistributedSampler(train_dataset)
valid_sampler = DistributedSampler(valid_dataset, shuffle=False)
train_data = DataLoader(train_dataset, batch_size=args.batch_size,
sampler=train_sampler, num_workers=16)
valid_data = DataLoader(valid_dataset, batch_size=args.batch_size,
sampler=valid_sampler, num_workers=4)
model_training(train_data, valid_data, args.epochs, save_dir)