-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_generative.py
163 lines (115 loc) · 6.37 KB
/
train_generative.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
from __future__ import print_function
import os
import sys
sys.dont_write_bytecode=True
import warnings
warnings.filterwarnings("ignore")
import glob
import time
import torch
import random
from torch.utils.data import DataLoader
from arguments import parse_arguments
from model import TrajectoryGenerator, TrajectoryDiscriminator
from data import dataset, collate_function
from generative_utils import discriminator_step, generator_step, check_accuracy
from utils import *
args = parse_arguments()
print(args.__dict__)
seed = 10
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)
torch.initial_seed()
torch.set_printoptions(precision=5)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if torch.cuda.is_available():
gpu_id = get_free_gpu().item()
torch.cuda.set_device(gpu_id)
if not args.test_only:
print("TRAINING DATA")
traindataset = dataset(glob.glob(f"data/{args.dset_name}/train/*.txt"), args)
print(f"Number of Training Samples: {len(traindataset)}")
print("VALIDATION DATA")
valdataset = dataset(glob.glob(f"data/{args.dset_name}/val/*.txt"), args)
print(f" Number of Validation Samples: {len(valdataset)}")
print("TEST DATA")
testdataset = dataset(glob.glob(f"data/{args.dset_name}/test/*.txt"), args)
print(f"Number of Test Samples: {len(testdataset)}")
print("-"*100)
if not args.test_only:
trainloader = DataLoader(traindataset, batch_size=args.batch_size, collate_fn=collate_function(), shuffle=True)
validloader = DataLoader(valdataset, batch_size=args.eval_batch_size if not args.eval_batch_size is None else len(valdataset), collate_fn=collate_function(), shuffle=False)
testloader = DataLoader(testdataset, batch_size=args.eval_batch_size if not args.eval_batch_size is None else len(testdataset), collate_fn=collate_function(), shuffle=False)
generator = TrajectoryGenerator(model_type=args.model_type, obs_len=args.obs_len, pred_len=args.pred_len, feature_dim=2, embedding_dim=args.embedding_dim, encoder_dim=args.encoder_dim, decoder_dim=args.decoder_dim, attention_dim=args.attention_dim, domain_parameter=args.domain_parameter, delta_bearing=30, delta_heading=30, pretrained_scene="resnet18", device=device, noise_dim=args.noise_dim, noise_type=args.noise_type, noise_mix_type=args.noise_mix_type).float().to(device)
discriminator=TrajectoryDiscriminator(args.d_model_type,seq_len=(args.obs_len+args.pred_len) if args.d_type=='global' else args.pred_len, feature_dim=2, embedding_dim=args.embedding_dim, hidden_size=args.encoder_dim, mlp_dim=1024, attention_dim=args.attention_dim, delta_bearing=args.delta_bearing, delta_heading=args.delta_heading, domain_parameter=args.domain_parameter)
discriminator=discriminator.float().to(device)
if not args.test_only:
opt_g = torch.optim.Adam(generator.parameters(), lr=args.lr_g)
opt_d = torch.optim.Adam(discriminator.parameters(), lr=args.lr_d)
if args.scheduler:
sch_g = torch.optim.lr_scheduler.ReduceLROnPlateau(opt_g, threshold=0.01, patience=10, factor=0.5, verbose=True, min_lr=1e-04)
sch_d = torch.optim.lr_scheduler.ReduceLROnPlateau(opt_d, threshold=0.01, patience=10, factor=0.5, verbose=True, min_lr=1e-04)
best_loss=float(1000)
generator.apply(init_weights)
discriminator.apply(init_weights)
model_file = f"trained-models/generative_spatial_temporal/{args.dset_name}/{args.best_k}V-{args.l}"
if not os.path.exists(f"./trained-models/{args.model_type}/{args.dset_name}"):
print(f"Creating directory ./trained-models/{args.model_type}/{args.dset_name}")
os.makedirs(f"./trained-models/{args.model_type}/{args.dset_name}")
g_file = f"{model_file}_g"
d_file=f"{model_file}_d"
if args.train_saved:
generator.load_state_dict(torch.load(f"{g_file}.pt"))
discriminator.load_state_dict(torch.load(f"{d_file}.pt"))
if args.test_only:
print("Evaluating Trained Model")
generator.load_state_dict(torch.load(f"{g_file}.pt"))
discriminator.load_state_dict(torch.load(f"{d_file}.pt"))
testloader = DataLoader(testdataset, batch_size=1, collate_fn=collate_function(), shuffle=False)
test_ade, test_fde = check_accuracy(testloader, generator, discriminator, args.num_traj)
print(f"Test ADE: {test_ade.item():.3f}")
print(f"Test FDE: {test_fde.item():.3f}")
exit()
print("---- TRAINING ---->")
for epoch in range(args.num_epochs):
print("*"*20)
epoch_ade = float(0)
d_loss_ =float(0)
g_loss_ = float(0)
generator.train()
discriminator.train()
epoch_time=time.time()
for b, batch in enumerate(trainloader):
loss_d = discriminator_step(b, batch, generator, discriminator, opt_d, d_spatial=args.d_spatial,d_type=args.d_type, d_domain=args.d_domain)
d_loss_+=loss_d.item()
loss_g, ade, fde, generator_pred, pedestrians,_ = generator_step(b, batch, generator, discriminator=discriminator, optimizer_g=opt_g, best_k=args.best_k, l=args.l, train=True, d_spatial=args.d_spatial, l2_loss_weight=args.l2_loss_weight, clip=args.clip,d_type=args.d_type, d_domain=args.d_domain)
epoch_ade+=ade.item()
g_loss_+=loss_g.item()
epoch_time=time.time()-epoch_time
d_loss_/=(b+1)
g_loss_/=(b+1)
epoch_ade/=(b+1)
print(f"[Epoch: {epoch+1}/{args.num_epochs}] Train ADE: (Min over {args.best_k}): {epoch_ade:.3f} Loss_G: {g_loss_:.3f} Loss_D: {d_loss_:.3f} --- Time Per Epoch: {epoch_time:.3f}")
generator.eval()
discriminator.eval()
val_ade, valid_fde, val_ade_ = check_accuracy(validloader, generator, discriminator, args.num_traj)
print(f"[Epoch: {epoch+1}/{args.num_epochs}] Valid ADE (Min over {args.num_traj}): {val_ade:.3f} (Min over 1): {val_ade_:.3f}")
if args.scheduler:
sch_g.step(val_ade)
sch_d.step(val_ade)
if (val_ade<best_loss):
best_loss=val_ade
test_ade, test_fde, test_ade_ = check_accuracy(testloader, generator, discriminator, args.num_traj)
torch.save(generator.state_dict(), f"{g_file}.pt")
torch.save(discriminator.state_dict(), f"{d_file}.pt")
print(f"[Epoch: {epoch+1}/{args.num_epochs}] Test ADE (Min over {args.num_traj}): {test_ade.item():.3f} (Min over 1): {test_ade_:.3f} Test FDE: {test_fde.item():.3f}")
print("Finished Training")
print("Evaluating Trained Model")
generator.load_state_dict(torch.load(f"{g_file}.pt"))
discriminator.load_state_dict(torch.load(f"{d_file}.pt"))
testloader = DataLoader(testdataset, batch_size=1, collate_fn=collate_function(), shuffle=False)
test_ade, test_fde = check_accuracy(testloader, generator, discriminator, args.num_traj)
print(f"Test ADE: {test_ade.item():.3f}")
print(f"Test FDE: {test_fde.item():.3f}")