-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
160 lines (125 loc) · 7.32 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
import os
import sys
import torch
import argparse
from datetime import datetime
from torch.utils.data import DataLoader
from method import dataset, model, utils
def main(args):
if args.checkpoint_dir is not None:
args.checkpoint_dir = os.path.join(args.checkpoint_dir,
datetime.strftime(datetime.now(), "%Y%m%d-%H%M%S-") + args.category)
os.makedirs(args.checkpoint_dir)
txt_file = open(os.path.join(args.checkpoint_dir, "training.log"), "w")
print("Logging into", os.path.join(args.checkpoint_dir, "training.log"), flush=True)
else:
txt_file = sys.stdout
train_ds = dataset.ChangeItDataset(
pickle_roots=args.pickle_roots, single_class=args.category, annotation_root=args.annotation_root,
file_mode="unannotated", noise_adapt_weight_root=args.noise_adapt_weight_root,
noise_adapt_weight_threshold_file=args.noise_adapt_weight_threshold_file, deterministic=args.ds_no_augment)
test_ds = dataset.ChangeItDataset(
pickle_roots=args.pickle_roots, single_class=args.category, annotation_root=args.annotation_root,
file_mode="annotated", deterministic=True)
print(train_ds, test_ds, sep="\n", file=txt_file, flush=True)
train_ds_iter = DataLoader(
train_ds, batch_size=args.batch_size, shuffle=True, drop_last=True,
num_workers=6, pin_memory=False, collate_fn=dataset.changeit_collate_fn)
test_ds_iter = DataLoader(
test_ds, batch_size=1, shuffle=True, drop_last=False,
num_workers=6, pin_memory=False, collate_fn=dataset.changeit_collate_fn)
# MODEL
network = model.MultiClassMLP(layers=[2048 + 2048, 512], n_classes=1).cuda()
optimizer = torch.optim.SGD(network.parameters(), args.lr, momentum=0.9, weight_decay=args.weight_decay)
scheduler = utils.get_cosine_schedule_with_warmup(optimizer, 10 * len(train_ds_iter),
len(train_ds_iter) * args.n_epochs)
loss_function = model.LookForTheChangeLoss(args.delta, args.kappa)
loss_metric = utils.AverageMeter()
loss_norm_metric = utils.AverageMeter()
unsup_state_loss_metric = utils.AverageMeter()
unsup_action_loss_metric = utils.AverageMeter()
for epoch in range(1, args.n_epochs + 1):
network.train()
loss_metric.reset()
loss_norm_metric.reset()
unsup_state_loss_metric.reset()
unsup_action_loss_metric.reset()
for i, batch in enumerate(train_ds_iter):
optimizer.zero_grad()
features = batch["features"].cuda()
log_pred = network(features)
log_pred_state = utils.select_correct_classes(log_pred["state"], batch["classes"])
log_pred_action = utils.select_correct_classes(log_pred["action"], batch["classes"])
video_loss_weight = None
if args.video_level_temp is not None:
assert batch["video_level_scores"] is not None
video_loss_weight = batch["video_level_scores"] * (-1 / args.video_level_temp)
video_loss_weight = 1 / (1 + torch.exp(video_loss_weight))
video_loss_weight = video_loss_weight.cuda()
unsup_state_loss, unsup_action_loss = \
loss_function(log_pred_state, log_pred_action, batch["lens"].cuda(), video_loss_weight)
loss = unsup_state_loss + unsup_action_loss
loss.backward()
total_norm = torch.norm(
torch.stack([torch.norm(p.grad.detach(), 2.0) for p in network.parameters()]), 2.0)
loss_norm_metric.update(total_norm.item(), len(features))
torch.nn.utils.clip_grad_norm_(network.parameters(), args.grad_clipping)
optimizer.step()
scheduler.step()
loss_metric.update(loss.item(), len(features))
unsup_state_loss_metric.update(unsup_state_loss.item(), len(features))
unsup_action_loss_metric.update(unsup_action_loss.item(), len(features))
if epoch % args.eval_each != 0:
continue
network.eval()
joint_meter = utils.JointMeter(train_ds.n_classes)
for i, batch in enumerate(test_ds_iter):
features = batch["features"].cuda()
with torch.no_grad():
log_pred = network(features)
log_pred_state = utils.select_correct_classes(log_pred["state"], batch["classes"])
log_pred_action = utils.select_correct_classes(log_pred["action"], batch["classes"])
test_pred_state = torch.softmax(log_pred_state, -1).cpu().numpy()[0]
test_pred_action = torch.sigmoid(log_pred_action).cpu().numpy()[0, :, 0]
joint_meter.log(test_pred_action, test_pred_state, batch["annotations"][0],
category=batch["classes"][0].item())
print(f"Epoch {epoch}/{args.n_epochs} ("
f"T loss: {loss_metric.value:.3f}, "
f"T lr: {scheduler.get_last_lr()[0]:.6f}, "
f"T grad norm: {loss_norm_metric.value:.1f}, "
f"T unsup state loss: {unsup_state_loss_metric.value:.3f}, "
f"T unsup action loss: {unsup_action_loss_metric.value:.3f}, "
f"V state acc: {joint_meter.acc:.1f}%, "
f"V state prec: {joint_meter.sp:.1f}%, "
f"V state joint prec: {joint_meter.jsp:.1f}%, "
f"V action prec: {joint_meter.ap:.1f}%, "
f"V action joint prec: {joint_meter.jap:.1f}%)", file=txt_file, flush=True)
print("> {:20} {:>6} {:>6} {:>6} {:>6} {:>6}".format("CATEGORY", "SAcc", "SP", "JtSP", "AP", "JtAP"), file=txt_file)
print("\n".join([
"> {:20}{:6.1f}%{:6.1f}%{:6.1f}%{:6.1f}%{:6.1f}%".format(cls_name, *joint_meter[train_ds.classes[cls_name]])
for cls_name in sorted(train_ds.classes.keys())
]), file=txt_file, flush=True)
if args.checkpoint_dir is not None:
save_name = f"sa{joint_meter.acc:.1f}-sp{joint_meter.sp:.1f}-" \
f"sj{joint_meter.jsp:.1f}-ap{joint_meter.ap:.1f}-aj{joint_meter.jap:.1f}"
torch.save(network.state_dict(), os.path.join(args.checkpoint_dir, f"model{epoch:03d}-{save_name}.pth.tar"))
txt_file.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--pickle_roots", type=str, nargs="+", required=True)
parser.add_argument("--category", type=str, default=None)
parser.add_argument("--annotation_root", type=str, default=None)
parser.add_argument("--batch_size", type=int, default=48)
parser.add_argument("--lr", default=0.0001, type=float)
parser.add_argument("--weight_decay", type=float, default=0.001)
parser.add_argument("--grad_clipping", type=float, default=2000.)
parser.add_argument("--n_epochs", default=100, type=int)
parser.add_argument("--eval_each", default=5, type=int)
parser.add_argument("--delta", type=int, default=2)
parser.add_argument("--kappa", type=int, default=60)
parser.add_argument("--video_level_temp", type=float, default=0.001)
parser.add_argument("--checkpoint_dir", type=str, default="./logs")
parser.add_argument("--ds_no_augment", action="store_true")
parser.add_argument("--noise_adapt_weight_root", type=str, default=None)
parser.add_argument("--noise_adapt_weight_threshold_file", type=str, default=None)
main(parser.parse_args())