PointStack/train.py at main · LongerVision/PointStack · GitHub

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
import os
import torch
import argparse
import datetime
import numpy as np
import random
import shutil

from glob import glob
from tqdm import tqdm
from torch.utils.data import DataLoader

from core.builders import build_dataset, build_network, build_optimizer
from utils.runtime_utils import cfg, cfg_from_yaml_file, validate
from utils.vis_utils import visualize_numpy

def parse_config():
    parser = argparse.ArgumentParser(description='arg parser')
    parser.add_argument('--cfg_file', type=str, default=None, help='specify the config for training')
    parser.add_argument('--exp_name', type=str, default=None, help='specify experiment name for saving outputs')
    parser.add_argument('--ckpt', type=str, default=None, help='checkpoint to start from')
    parser.add_argument('--random_seed', type=int, default=0, help='random seed number')
    parser.add_argument('--val_steps', type=int, default=1, help='perform validation every n steps')
    parser.add_argument('--pretrained_ckpt', type = str, default = None, help='path to pretrained ckpt')
    args = parser.parse_args()

    cfg_from_yaml_file(args.cfg_file, cfg)
    exp_dir = cfg.ROOT_DIR / 'experiments' / cfg.DATASET.NAME / args.exp_name
    os.makedirs(exp_dir, exist_ok=True)
    shutil.copy2(args.cfg_file, exp_dir)

    return args, cfg

args, cfg = parse_config()

random_seed = cfg.RANDOM_SEED # Setup seed for reproducibility
torch.manual_seed(random_seed)
torch.cuda.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)

# Build Dataloader
train_dataset = build_dataset(cfg, split = 'train')
train_dataloader = DataLoader(train_dataset, batch_size=cfg.OPTIMIZER.BATCH_SIZE, shuffle=True, drop_last=True)

val_dataset = build_dataset(cfg, split='val')
val_dataloader = DataLoader(val_dataset, batch_size=1, shuffle=False, drop_last=False)

# Build Network and Optimizer
net = build_network(cfg)
if args.pretrained_ckpt is not None:
    pretrained_state_dict = torch.load(args.pretrained_ckpt)['model_state_dict']

    for k, v in net.state_dict().items():
        if (v.shape != pretrained_state_dict[k].shape):
            del pretrained_state_dict[k]

    net.load_state_dict(pretrained_state_dict, strict = False)

net = net.cuda()
opt, scheduler = build_optimizer(cfg, net.parameters(), len(train_dataloader))


from torch.utils.tensorboard import SummaryWriter
ckpt_dir = cfg.ROOT_DIR / 'experiments' / cfg.DATASET.NAME / args.exp_name / 'ckpt'
tensorboard_dir = cfg.ROOT_DIR / 'experiments' / cfg.DATASET.NAME / args.exp_name / 'tensorboard'

os.makedirs(ckpt_dir, exist_ok=True)
os.makedirs(tensorboard_dir, exist_ok=True)

writer = SummaryWriter(tensorboard_dir)

min_loss = 1e20
max_acc = 0

steps_cnt = 0
epoch_cnt = 0


for epoch in tqdm(range(1, cfg.OPTIMIZER.MAX_EPOCH + 1)):
    opt.zero_grad()
    net.zero_grad()
    net.train()
    loss = 0
    for data_dic in tqdm(train_dataloader):

        data_dic = net(data_dic)
        loss, loss_dict = net.get_loss(data_dic, smoothing = True, is_segmentation = cfg.DATASET.IS_SEGMENTATION)
        loss = loss
        loss.backward()
        steps_cnt += 1

        # if (steps_cnt)%(cfg.OPTIMIZER.GRAD_ACCUMULATION) == 0:
        torch.nn.utils.clip_grad_norm_(net.parameters(), cfg.OPTIMIZER.GRAD_CLIP)
        opt.step()
        opt.zero_grad()
        lr = scheduler.get_last_lr()[0]
        scheduler.step()
        writer.add_scalar('steps/loss', loss, steps_cnt)
        writer.add_scalar('steps/lr', lr, steps_cnt)

        for k,v in loss_dict.items():
            writer.add_scalar('steps/loss_' + k, v, steps_cnt)

    if (epoch % args.val_steps) == 0:
        val_dict = validate(net, val_dataloader, net.get_loss, 'cuda', is_segmentation = cfg.DATASET.IS_SEGMENTATION)

        print('='*20, 'Epoch ' + str(epoch+1), '='*20)

        if cfg.DATASET.IS_SEGMENTATION:
            writer.add_scalar('epochs/val_miou', val_dict['miou'], epoch_cnt)
            print('Val mIoU: ', val_dict['miou'])

        else:
            writer.add_scalar('epochs/val_loss', val_dict['loss'], epoch_cnt)
            writer.add_scalar('epochs/val_acc', val_dict['acc'], epoch_cnt)
            writer.add_scalar('epochs/val_acc_avg', val_dict['acc_avg'], epoch_cnt)
            print('Val Loss: ', val_dict['loss'], 'Val Accuracy: ', val_dict['acc'], 'Val Avg Accuracy: ', val_dict['acc_avg'])

            for k,v in val_dict['loss_dic'].items():
                writer.add_scalar('epochs/val_loss_'+ k, v, epoch_cnt)

        epoch_cnt += 1


        if cfg.DATASET.IS_SEGMENTATION:
            if val_dict['miou'] > max_acc:
                torch.save({
                    'epoch': epoch,
                    'model_state_dict': net.state_dict(),
                    'optimizer_state_dict': opt.state_dict(),
                    }, ckpt_dir / 'ckpt-best.pth')

                max_acc = val_dict['miou']
        else:

            if val_dict['acc'] > max_acc:
                torch.save({
                    'epoch': epoch,
                    'model_state_dict': net.state_dict(),
                    'optimizer_state_dict': opt.state_dict(),
                    'loss': val_dict['loss'],
                    }, ckpt_dir / 'ckpt-best.pth')

                max_acc = val_dict['acc']

    torch.save({
                'epoch': epoch,
                'model_state_dict': net.state_dict(),
                'optimizer_state_dict': opt.state_dict(),
                }, ckpt_dir / 'ckpt-last.pth')