forked from OpenDriveLab/Openpilot-Deepdive
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_pl.py
249 lines (199 loc) · 10.9 KB
/
main_pl.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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
import os
import torch
from torch import nn
import torch.nn.functional as F
from torch import optim
from torch.utils.data import DataLoader
from argparse import ArgumentParser
from data import PlanningDataset, SequencePlanningDataset, Comma2k19SequenceDataset
from model import PlaningNetwork, MultipleTrajectoryPredictionLoss, SequencePlanningNetwork
from utils import draw_trajectory_on_ax, get_val_metric
import pytorch_lightning as pl
import matplotlib.pyplot as plt
from pytorch_lightning.callbacks import LearningRateMonitor
from tqdm import tqdm
class PlanningBaselineV0(pl.LightningModule):
def __init__(self, M, num_pts, mtp_alpha, lr) -> None:
super().__init__()
self.M = M
self.num_pts = num_pts
self.mtp_alpha = mtp_alpha
self.lr = lr
self.net = PlaningNetwork(M, num_pts)
self.mtp_loss = MultipleTrajectoryPredictionLoss(mtp_alpha, M, num_pts, distance_type='angle')
@staticmethod
def add_model_specific_args(parent_parser):
parser = parent_parser.add_argument_group('PlanningBaselineV0')
parser.add_argument('--M', type=int, default=3)
parser.add_argument('--num_pts', type=int, default=20)
parser.add_argument('--mtp_alpha', type=float, default=1.0)
return parent_parser
def forward(self, x):
# in lightning, forward defines the prediction/inference actions
return self.net(x)
def training_step(self, batch, batch_idx):
# training_step defines the train loop. It is independent of forward
inputs, labels = batch['input_img'], batch['future_poses']
pred_cls, pred_trajectory = self.net(inputs)
cls_loss, reg_loss = self.mtp_loss(pred_cls, pred_trajectory, labels)
self.log('loss/cls', cls_loss)
self.log('loss/reg', reg_loss.mean())
self.log('loss/reg_x', reg_loss[0])
self.log('loss/reg_y', reg_loss[1])
self.log('loss/reg_z', reg_loss[2])
if batch_idx % 10 == 0:
trajectories = list(pred_trajectory[0].detach().cpu().numpy().reshape(self.M, self.num_pts, 3)) # M, num_pts, 3
trajectories.append(labels[0].detach().cpu().numpy())
confs = list(F.softmax(pred_cls[0].detach().cpu(), dim=-1).numpy()) + [1, ] # M,
fig, ax = plt.subplots()
ax = draw_trajectory_on_ax(ax, trajectories, confs)
plt.tight_layout()
self.logger.experiment.add_figure('train_vis', fig, self.global_step)
plt.close(fig)
return cls_loss + self.mtp_alpha * reg_loss.mean()
def configure_optimizers(self):
optimizer = optim.SGD(self.parameters(), lr=self.lr, momentum=0.9, weight_decay=0.01)
# optimizer = optim.Adam(self.parameters(), lr=self.lr, weight_decay=0.01)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, 20, 0.9)
return [optimizer], [lr_scheduler]
def validation_step(self, batch, batch_idx):
inputs, labels = batch['input_img'], batch['future_poses']
pred_cls, pred_trajectory = self.net(inputs)
metrics = get_val_metric(pred_cls, pred_trajectory.view(-1, self.M, self.num_pts, 3), labels)
self.log_dict(metrics)
# Pytorch-lightning will collect those binary values and calculate the mean
class SequencePlanningBaselineV0(pl.LightningModule):
def __init__(self, M, num_pts, mtp_alpha, lr, optimizer) -> None:
super().__init__()
self.M = M
self.num_pts = num_pts
self.mtp_alpha = mtp_alpha
self.lr = lr
self.optimizer = optimizer
self.net = SequencePlanningNetwork(M, num_pts)
self.mtp_loss = MultipleTrajectoryPredictionLoss(mtp_alpha, M, num_pts, distance_type='angle')
@staticmethod
def add_model_specific_args(parent_parser):
parser = parent_parser.add_argument_group('SequencePlanningBaselineV0')
parser.add_argument('--M', type=int, default=3)
parser.add_argument('--num_pts', type=int, default=20)
parser.add_argument('--mtp_alpha', type=float, default=1.0)
parser.add_argument('--optimizer', type=str, default='sgd')
return parent_parser
def forward(self, x, hidden=None):
if hidden is None:
hidden = torch.zeros((2, x.size(0), 512)).to(self.device)
# in lightning, forward defines the prediction/inference actions
return self.net(x, hidden)
def training_step(self, batch, batch_idx):
# training_step defines the train loop. It is independent of forward
seq_inputs, seq_labels = batch['seq_input_img'], batch['seq_future_poses']
bs = seq_labels.size(0)
seq_length = seq_labels.size(1)
cls_loss_total, reg_loss_total = 0, 0
hidden = torch.zeros((2, bs, 512)).to(self.device)
for t in range(seq_length):
inputs, labels = seq_inputs[:, t, :, :, :], seq_labels[:, t, :, :]
pred_cls, pred_trajectory, hidden = self.net(inputs, hidden)
cls_loss, reg_loss = self.mtp_loss(pred_cls, pred_trajectory, labels)
cls_loss_total += cls_loss
reg_loss_total += reg_loss
cls_loss_total = cls_loss_total / (seq_length + 1)
reg_loss_total = reg_loss_total / (seq_length + 1)
self.log('loss/cls', cls_loss_total)
self.log('loss/reg', reg_loss_total.mean())
self.log('loss/reg_x', reg_loss_total[0])
self.log('loss/reg_y', reg_loss_total[1])
self.log('loss/reg_z', reg_loss_total[2])
# if batch_idx % 10 == 0:
# trajectories = list(pred_trajectory[0].detach().cpu().numpy().reshape(self.M, self.num_pts, 3)) # M, num_pts, 3
# trajectories.append(labels[0].detach().cpu().numpy())
# confs = list(F.softmax(pred_cls[0].detach().cpu(), dim=-1).numpy()) + [1, ] # M,
# fig, ax = plt.subplots()
# ax = draw_trajectory_on_ax(ax, trajectories, confs)
# plt.tight_layout()
# self.logger.experiment.add_figure('train_vis', fig, self.global_step)
# plt.close(fig)
return cls_loss_total + self.mtp_alpha * reg_loss_total.mean()
def configure_optimizers(self):
if self.optimizer == 'sgd':
optimizer = optim.SGD(self.parameters(), lr=self.lr, momentum=0.9, weight_decay=0.01)
elif self.optimizer == 'adam':
optimizer = optim.Adam(self.parameters(), lr=self.lr, weight_decay=0.01)
else:
raise NotImplementedError
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, 20, 0.9)
return [optimizer], [lr_scheduler]
def validation_step(self, batch, batch_idx):
seq_inputs, seq_labels = batch['seq_input_img'], batch['seq_future_poses']
bs = seq_labels.size(0)
seq_length = seq_labels.size(1)
hidden = torch.zeros((2, bs, 512)).to(self.device)
for t in range(seq_length):
inputs, labels = seq_inputs[:, t, :, :, :], seq_labels[:, t, :, :]
pred_cls, pred_trajectory, hidden = self.net(inputs, hidden)
metrics = get_val_metric(pred_cls, pred_trajectory.view(-1, self.M, self.num_pts, 3), labels)
self.log_dict(metrics)
class SequenceBaselineV1(SequencePlanningBaselineV0):
def __init__(self, M, num_pts, mtp_alpha, lr, optimizer) -> None:
super().__init__(M, num_pts, mtp_alpha, lr, optimizer)
self.automatic_optimization = False
self.optimize_per_n_step = 40
def training_step(self, batch, batch_idx):
# manual backward
opt = self.optimizers()
seq_inputs, seq_labels = batch['seq_input_img'], batch['seq_future_poses']
bs = seq_labels.size(0)
seq_length = seq_labels.size(1)
hidden = torch.zeros((2, bs, 512)).to(self.device)
total_loss = 0
for t in tqdm(range(seq_length), leave=False):
inputs, labels = seq_inputs[:, t, :, :, :], seq_labels[:, t, :, :]
pred_cls, pred_trajectory, hidden = self.net(inputs, hidden)
cls_loss, reg_loss = self.mtp_loss(pred_cls, pred_trajectory, labels)
total_loss += (cls_loss + self.mtp_alpha * reg_loss.mean()) / self.optimize_per_n_step
self.log('loss/cls', cls_loss)
self.log('loss/reg', reg_loss.mean())
self.log('loss/reg_x', reg_loss[0])
self.log('loss/reg_y', reg_loss[1])
self.log('loss/reg_z', reg_loss[2])
if (t + 1) % self.optimize_per_n_step == 0:
opt.zero_grad()
self.manual_backward(total_loss)
opt.step()
hidden = hidden.clone().detach()
total_loss = 0
if not isinstance(total_loss, int):
opt.zero_grad()
self.manual_backward(total_loss)
opt.step()
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument('--batch_size', type=int, default=32 * 4)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--n_workers', type=int, default=8)
parser = SequencePlanningBaselineV0.add_model_specific_args(parser)
parser = pl.Trainer.add_argparse_args(parser)
args = parser.parse_args()
# data_p = {10: 'p3_10pts_%s.json', 20: 'p3_%s.json'}[args.num_pts]
# data_p = 'p3_10pts_can_bus_%s_temporal.json'
# train = SequencePlanningDataset(split='train', json_path_pattern=data_p)
# val = SequencePlanningDataset(split='val', json_path_pattern=data_p)
train = Comma2k19SequenceDataset('data/comma2k19_train_non_overlap.txt', 'data/comma2k19/','train', use_memcache=False)
val = Comma2k19SequenceDataset('data/comma2k19_val_non_overlap.txt', 'data/comma2k19/','val', use_memcache=False)
train_loader = DataLoader(train, args.batch_size, shuffle=True, num_workers=args.n_workers, persistent_workers=True, prefetch_factor=2, pin_memory=True)
val_loader = DataLoader(val, args.batch_size, num_workers=args.n_workers, persistent_workers=True, prefetch_factor=2, pin_memory=True)
planning_v0 = SequenceBaselineV1(args.M, args.num_pts, args.mtp_alpha, args.lr, args.optimizer)
lr_monitor = LearningRateMonitor(logging_interval='step')
trainer = pl.Trainer.from_argparse_args(args,
accelerator='ddp' if args.gpus > 1 else None,
profiler='simple',
benchmark=True,
log_every_n_steps=1,
flush_logs_every_n_steps=10,
callbacks=[lr_monitor],
check_val_every_n_epoch=10, # val every 10 epoch to speed up train process
# val_check_interval=0.0, # Disable in-batch val
progress_bar_refresh_rate=1, # for slurm env
)
trainer.fit(planning_v0, train_loader, val_loader)