-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
163 lines (121 loc) · 6.95 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
161
162
163
import os
import gc
import time
import argparse
from tqdm import tqdm
import torch
from torch.cuda import amp
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import DataLoader
from utils import util
from utils import dataset
from models.unet import UNet
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--cities', type=list, choices=['ANTWERP', 'BANGKOK', 'BARCELONA', "MOSCOW"], default=['ANTWERP'])
parser.add_argument('--train_year', type=list, choices=[2019, 2020], default=[2019])
parser.add_argument('--val_year', type=list, choices=[2019, 2020], default=[2020])
parser.add_argument('--model', type=str, choices=["UNET"], default="UNET")
parser.add_argument('--scheduler', type=str, default="StepLR")
parser.add_argument('--learning_rate', type=float, default=3e-4)
parser.add_argument('--batch_size', type=int, default=1)
parser.add_argument('--num_workers', type=int, default=0)
parser.add_argument('--num_epochs', type=int, default=6)
parser.add_argument('--L1_regularization', type=bool, default=False)
parser.add_argument('--wd', type=float, default=2e-4)
parser.add_argument('--num_file_train', type=int, default=14)
parser.add_argument('--accumulation_step', type=int, default=8)
parser.add_argument('--use_mask', type=bool, default=False)
parser.add_argument('--device', type=str, default=torch.device("cuda:0" if torch.cuda.is_available() else "cpu"))
parser.add_argument('--seed', type=int, default=14563)
return parser.parse_args()
def train(args, writer, experiment_name):
train_files, valid_files, static_map = util.get_files(args)
valid_dataset = dataset.TrainDataset(valid_files, static_map)
valid_dataloader = DataLoader(dataset=valid_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, collate_fn=dataset.collate_fn,
sampler=dataset.val_local_sampler(valid_dataset), pin_memory=True)
if args.model == "UNET":
model = UNet()
model.to(args.device)
criterian = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate, weight_decay=args.wd)
scheduler = util.get_scheduler(args.scheduler, optimizer)
scaler = amp.GradScaler()
best_val_loss = 1e10
dataset_size = 0
global_step = 0
for epoch in range(args.num_epochs):
train_dataset = dataset.TrainDataset(train_files[epoch*args.num_file_train: (epoch+1)*args.num_file_train], static_map)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, collate_fn=dataset.collate_fn, # noqa
sampler=dataset.train_local_sampler(train_dataset), pin_memory=True) # noqa
length_dataloader = len(train_dataloader)
model.train()
running_loss = 0.0
model.zero_grad()
pbar = tqdm(enumerate(train_dataloader), total=length_dataloader, desc=f"Training {epoch+1}/{args.num_epochs}")
for idx, data in pbar:
inputs = data[0].to(args.device, dtype=torch.float) # (bs, 105, 496, 448)
targets = data[1].to(args.device, dtype=torch.float) # (bs, 48, 495, 436)
with amp.autocast():
prediction = model(inputs)
if args.use_mask:
mask_city = util.create_static_mask(args, static_map)
pred = prediction[:, :, 1:, 6:-6] * mask_city # (bs, 105, 495, 436)
else:
pred = prediction[:, :, 1:, 6:-6] # (bs, 105, 495, 436)
loss = criterian(pred, targets)
if args.L1_regularization:
l1_weight = 0.001
l1 = l1_weight * sum(p.abs().sum() for p in model.parameters())
loss += l1
loss = loss / args.accumulation_step
scaler.scale(loss).backward()
if ((idx + 1) % args.accumulation_step == 0) or (idx + 1 == length_dataloader):
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
global_step += 1
dataset_size += inputs.shape[0]
running_loss += (loss.item() * inputs.shape[0] * args.accumulation_step)
epoch_loss = running_loss / dataset_size
pbar.set_postfix(train_loss=f"{epoch_loss:0.5f}")
# if scheduler is not None:
# scheduler.step(epoch_loss)
if global_step % 8 == 0:
writer.add_scalar("Training_loss/minibatches", epoch_loss, global_step)
torch.cuda.empty_cache()
gc.collect()
train_epoch_loss = epoch_loss
val_epoch_loss = util.evaluate(args, model, valid_dataloader, static_map, criterian, epoch, writer, experiment_name)
current_lr = optimizer.param_groups[0]["lr"]
if scheduler is not None:
scheduler.step()
util.log_to_tensorboard(writer, model, train_epoch_loss, val_epoch_loss, current_lr, epoch, save_histogram=True)
print(f"Epoch: {epoch+1} | {train_epoch_loss=:0.5f} | {val_epoch_loss=:0.5f} | {current_lr=:0.7f}")
if val_epoch_loss < best_val_loss:
best_val_loss = val_epoch_loss
if os.path.isdir(f"checkpoints/{experiment_name}"):
torch.save({'epoch': epoch+1, 'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()},
f"checkpoints/{experiment_name}/best_model.pth") # noqa
else:
os.makedirs(f"checkpoints/{experiment_name}")
torch.save({'epoch': epoch+1, 'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()},
f"checkpoints/{experiment_name}/best_model.pth") # noqa
torch.save({'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()},
f"checkpoints/{experiment_name}/last_model.pth")
if __name__ == "__main__":
start = time.time()
args = parse_args()
util.seed_everything(args.seed)
util.make_dir()
experiment_name = f"Training_{args.train_year[0]}_Validation_{args.val_year[0]}_{args.cities[0]}_{args.model}_unet512filter_L1_dropout"
writer = SummaryWriter(f"logs/{experiment_name}")
writer.add_text(experiment_name, f"scheduler: {args.scheduler}, wd: {args.wd}, L1: {args.L1_regularization}, lambda: 0.001, mask: {args.use_mask}, gr_acc: {args.accumulation_step}")
print(f"Experiment(training): {experiment_name}\n")
train(args, writer, experiment_name)
print("Training_time (minute): ", round((time.time() - start)/60, 3))