File size: 6,268 Bytes
0b70b07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import time
import argparse
import numpy as np
from tqdm import tqdm
from glob import glob

import torch
from torch.utils.data import DataLoader

from video_module.dataset import Water_Image_Train_DS
from video_module.model import AFB_URR, FeatureBank
import myutils

# Enable CUDA launch blocking for debugging; set to '0' to disable.
os.environ['CUDA_LAUNCH_BLOCKING'] = '0'

def get_args():
    parser = argparse.ArgumentParser(description='Train V-BeachNet')
    parser.add_argument('--gpu', type=int, default=0, help='GPU card id.')
    parser.add_argument('--dataset', type=str, required=True, help='Dataset folder.')
    parser.add_argument('--seed', type=int, default=-1, help='Random seed.')
    parser.add_argument('--log', action='store_true', help='Save the training results.')
    parser.add_argument('--lr', type=float, default=1e-5, help='Learning rate (default: 1e-5).')
    parser.add_argument('--lu', type=float, default=0.5, help='Regularization factor (default: 0.5).')
    parser.add_argument('--resume', type=str, help='Path to checkpoint (default: none).')
    parser.add_argument('--new', action='store_true', help='Train the model from scratch.')
    parser.add_argument('--scheduler-step', type=int, default=25, help='Scheduler step size (default: 25).')
    parser.add_argument('--total-epochs', type=int, default=100, help='Total number of epochs (default: 100).')
    parser.add_argument('--budget', type=int, default=300000, help='Maximum number of features in the feature bank (default: 300000).')
    parser.add_argument('--obj-n', type=int, default=2, help='Maximum number of objects trained simultaneously.')
    parser.add_argument('--clip-n', type=int, default=6, help='Maximum number of frames in a batch.')

    return parser.parse_args()

def train_model(model, dataloader, criterion, optimizer):
    stats = myutils.AvgMeter()
    uncertainty_stats = myutils.AvgMeter()
    progress_bar = tqdm(dataloader)
    
    for _, sample in enumerate(progress_bar):
        frames, masks, obj_n, info = sample

        if obj_n.item() == 1:
            continue

        frames, masks = frames[0].to(device), masks[0].to(device)
        fb_global = FeatureBank(obj_n.item(), args.budget, device)
        k4_list, v4_list = model.memorize(frames[:1], masks[:1])
        fb_global.init_bank(k4_list, v4_list)

        scores, uncertainty = model.segment(frames[1:], fb_global)
        label = torch.argmax(masks[1:], dim=1).long()

        optimizer.zero_grad()
        loss = criterion(scores, label) + args.lu * uncertainty
        loss.backward()
        optimizer.step()

        uncertainty_stats.update(uncertainty.item())
        stats.update(loss.item())
        progress_bar.set_postfix(loss=f'{loss.item():.5f} (Avg: {stats.avg:.5f}, Uncertainty Avg: {uncertainty_stats.avg:.5f})')

    return stats.avg

def main():
    dataset = Water_Image_Train_DS(root=args.dataset, output_size=400, clip_n=args.clip_n, max_obj_n=args.obj_n)
    dataloader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2, pin_memory=True)
    print(myutils.gct(), f'Dataset with {len(dataset)} training cases.')

    model = AFB_URR(device, update_bank=False, load_imagenet_params=True).to(device)
    model.train()
    model.apply(myutils.set_bn_eval)

    optimizer = torch.optim.AdamW(filter(lambda x: x.requires_grad, model.parameters()), args.lr)

    start_epoch, best_loss = 0, float('inf')
    if args.resume:
        if os.path.isfile(args.resume):
            checkpoint = torch.load(args.resume)
            model.load_state_dict(checkpoint['model'], strict=False)
            seed = checkpoint.get('seed', int(time.time()))

            if not args.new:
                start_epoch = checkpoint['epoch'] + 1
                optimizer.load_state_dict(checkpoint['optimizer'])
                best_loss = checkpoint['loss']
                print(myutils.gct(), f'Resumed from checkpoint {args.resume} (Epoch: {start_epoch-1}, Best Loss: {best_loss}).')
            else:
                print(myutils.gct(), f'Loaded checkpoint {args.resume}. Training from scratch.')
        else:
            raise FileNotFoundError(f'No checkpoint found at {args.resume}')
    else:
        seed = args.seed if args.seed >= 0 else int(time.time())

    print(myutils.gct(), 'Random seed:', seed)
    torch.manual_seed(seed)
    np.random.seed(seed)

    criterion = torch.nn.CrossEntropyLoss().to(device)
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.scheduler_step, gamma=0.5, last_epoch=start_epoch-1)

    for epoch in range(start_epoch, args.total_epochs):
        print(f'\n{myutils.gct()} Epoch: {epoch}, Learning Rate: {scheduler.get_last_lr()[0]:.6f}')
        loss = train_model(model, dataloader, criterion, optimizer)

        if args.log:
            checkpoint = {
                'epoch': epoch,
                'model': model.state_dict(),
                'optimizer': optimizer.state_dict(),
                'loss': loss,
                'seed': seed
            }
            torch.save(checkpoint, os.path.join(model_path, 'final.pth'))
            if loss < best_loss:
                best_loss = loss
                torch.save(checkpoint, os.path.join(model_path, 'best.pth'))
                print('Updated best model.')

        scheduler.step()

if __name__ == '__main__':
    args = get_args()
    print(myutils.gct(), f'Arguments: {args}')

    if args.gpu >= 0 and torch.cuda.is_available():
        device = torch.device('cuda', args.gpu)
    else:
        raise ValueError('CUDA is required. Ensure --gpu is set to >= 0.')

    if args.log:
        log_dir = os.path.join('logs', time.strftime('%Y%m%d-%H%M%S'))
        model_path = os.path.join(log_dir, 'model')
        os.makedirs(model_path, exist_ok=True)
        myutils.save_scripts(log_dir, scripts_to_save=glob('*.*'))
        myutils.save_scripts(log_dir, scripts_to_save=glob('dataset/*.py', recursive=True))
        myutils.save_scripts(log_dir, scripts_to_save=glob('model/*.py', recursive=True))
        myutils.save_scripts(log_dir, scripts_to_save=glob('myutils/*.py', recursive=True))
        print(myutils.gct(), f'Created log directory: {log_dir}')

    main()
    print(myutils.gct(), 'Training completed.')