INVITATION/train.py at master · CVEO/INVITATION · 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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
import os
import sys
import time
import argparse
from tqdm import tqdm

import torch
import torch.nn as nn

import torch.distributed as dist
import torch.backends.cudnn as cudnn
from torch.nn.parallel import DistributedDataParallel

from config import config
from dataloader.dataloader import get_train_loader
from models.builder import DualSegModel as DualSegModel
from dataloader.UAVDataset import UAVDataset
from utils.init_func import group_weight
from utils.lr_policy import WarmUpPolyLR
from engine.engine import Engine
# from utils.loss import Lovas

from utils.pyt_utils import all_reduce_tensor

from tensorboardX import SummaryWriter

parser = argparse.ArgumentParser()


with Engine(custom_parser=parser) as engine:
    args = parser.parse_args()

    cudnn.benchmark = True
    seed = config.seed
    if engine.distributed:
        seed = engine.local_rank
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed(seed)

    criterion = nn.CrossEntropyLoss(reduction='mean', ignore_index=255)
    if engine.distributed:
        BatchNorm2d = nn.SyncBatchNorm
    else:
        BatchNorm2d = nn.BatchNorm2d

    model = DualSegModel(cfg=config, criterion=criterion, norm_layer=BatchNorm2d)

    # group weight and config optimizer
    base_lr = config.lr
    if engine.distributed:
        base_lr = config.lr
        niters_per_epoch = config.num_train_imgs // (
            config.batch_size * engine.world_size) + 1
    else:
        niters_per_epoch = config.niters_per_epoch

    params_list = []
    params_list = group_weight(params_list, model, BatchNorm2d, base_lr)

    if config.optimizer == 'AdamW':
        optimizer = torch.optim.AdamW(params_list, lr=base_lr, betas=(
            0.9, 0.999), weight_decay=config.weight_decay)
    elif config.optimizer == 'SGDM':
        optimizer = torch.optim.SGD(
            params_list, lr=base_lr, momentum=config.momentum, weight_decay=config.weight_decay)
    else:
        raise NotImplementedError

    # config lr policy
    total_iteration = config.nepochs * niters_per_epoch
    warm_up_iteration = config.warm_up_epoch * niters_per_epoch
    lr_policy = WarmUpPolyLR(
        base_lr, config.lr_power, total_iteration, warm_up_iteration, config.warmup_ratio)

    if engine.distributed:
        if torch.cuda.is_available():
            model.cuda()
            model = DistributedDataParallel(model, device_ids=[engine.local_rank],
                                            output_device=engine.local_rank, find_unused_parameters=False)
    else:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model.to(device)

    train_loader, train_sampler = get_train_loader(engine, UAVDataset)

    if (engine.distributed and (engine.local_rank == 0)) or (not engine.distributed):
        tb_dir = config.tb_dir + \
            '/{}'.format(time.strftime("%b%d_%d-%H-%M", time.localtime()))
        generate_tb_dir = config.tb_dir + '/tb'
        tb = SummaryWriter(log_dir=tb_dir)
        engine.link_tb(tb_dir, generate_tb_dir)

    engine.register_state(dataloader=train_loader, model=model,
                          optimizer=optimizer)
    if engine.continue_state_object:
        engine.restore_checkpoint()

    optimizer.zero_grad()
    model.train()

    for epoch in range(engine.state.epoch, config.nepochs+1):
        if engine.distributed:
            train_sampler.set_epoch(epoch-1)
        bar_format = '{desc}[{elapsed}<{remaining},{rate_fmt}]'
        pbar = tqdm(range(niters_per_epoch), file=sys.stdout,
                    bar_format=bar_format)
        dataloader = iter(train_loader)

        sum_loss = 0

        for idx in pbar:
            engine.update_iteration(epoch, idx)

            minibatch = dataloader.next()

            imgs = minibatch['patch']
            gts = minibatch['label']
            depths = minibatch['depth']

            imgs = imgs.cuda(non_blocking=True)
            gts = gts.cuda(non_blocking=True)
            depths = depths.cuda(non_blocking=True)

            out, loss = model(imgs, depths, gts)

            # reduce the whole loss over multi-gpu
            if engine.distributed:
                reduce_loss = all_reduce_tensor(
                    loss, world_size=engine.world_size)

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            current_idx = (epoch - 1) * niters_per_epoch + idx
            lr = lr_policy.get_lr(current_idx)

            for i in range(len(optimizer.param_groups)):
                optimizer.param_groups[i]['lr'] = lr

            if engine.distributed:
                sum_loss += reduce_loss.item()
                print_str = 'Epoch {}/{}'.format(epoch, config.nepochs) \
                    + ' Iter {}/{}:'.format(idx + 1, niters_per_epoch) \
                    + ' lr=%.4e' % lr \
                    + ' loss=%.4f total_loss=%.4f' % (reduce_loss.item(), (sum_loss / (idx + 1)))
            else:
                sum_loss += loss
                print_str = 'Epoch {}/{}'.format(epoch, config.nepochs) \
                    + ' Iter {}/{}:'.format(idx + 1, niters_per_epoch) \
                    + ' lr=%.4e' % lr \
                    + ' loss=%.4f total_loss=%.4f' % (loss, (sum_loss / (idx + 1)))

            del loss
            pbar.set_description(print_str, refresh=False)

        if (engine.distributed and (engine.local_rank == 0)) or (not engine.distributed):
            tb.add_scalar('epoch_train_loss', sum_loss / len(pbar), epoch)

        if (epoch >= config.checkpoint_start_epoch) and (epoch % config.checkpoint_step == 0) or (epoch == config.nepochs):
            if engine.distributed and (engine.local_rank == 0):
                engine.save_and_link_checkpoint(config.checkpoint_dir,
                                                config.log_dir,
                                                config.log_dir_link)
            elif not engine.distributed:
                engine.save_and_link_checkpoint(config.checkpoint_dir,
                                                config.log_dir,
                                                config.log_dir_link)