CasCast/train.py at master · OpenEarthLab/CasCast · 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
170
171
172
173
174
175
176
177
178
179
import argparse
import os
import torch
from utils.builder import ConfigBuilder
import utils.misc as utils
import yaml
from utils.logger import get_logger
from megatron_utils.tensor_parallel.data import get_data_loader_length


#----------------------------------------------------------------------------

def subprocess_fn(args):
    utils.setup_seed(args.seed * args.world_size + args.rank)

    logger = get_logger("train", args.run_dir, utils.get_rank(), filename='iter.log', resume=args.resume)

    # args.logger = logger
    args.cfg_params["logger"] = logger

    # build config
    logger.info('Building config ...')
    builder = ConfigBuilder(**args.cfg_params)

    logger.info('Building dataloaders ...')

    train_dataloader = builder.get_dataloader(split = 'train')
    logger.info('Train dataloaders build complete')
    valid_dataloader = builder.get_dataloader(split = 'valid')
    logger.info('Valid dataloaders build complete')

    ## set lr_scheduler (by epoch/step) ##
    model_params = args.cfg_params['model']['params']
    ## by step ##
    if "sampler" in args.cfg_params.keys() and "TrainingSampler" in args.cfg_params["sampler"]["type"]:
         lr_scheduler_params = model_params['lr_scheduler']
         for key in lr_scheduler_params:
             for key1 in lr_scheduler_params[key]:
                        if "epochs" in key1:
                            lr_scheduler_params[key][key1] = int(builder.get_max_step() * lr_scheduler_params[key][key1])
    else:
        ## by epoch ##
        steps_per_epoch = get_data_loader_length(train_dataloader)
        if 'lr_scheduler' in model_params:
            lr_scheduler_params = model_params['lr_scheduler']
            for key in lr_scheduler_params:
                if 'by_step' in lr_scheduler_params[key]:
                    if lr_scheduler_params[key]['by_step']:
                        for key1 in lr_scheduler_params[key]:
                            if "epochs" in key1:
                                lr_scheduler_params[key][key1] *= steps_per_epoch


    # build model
    logger.info('Building models ...')
    model = builder.get_model()

    if model.use_ceph:
        model_checkpoint = os.path.join(args.relative_checkpoint_dir, 'checkpoint_latest.pth')
    else:
        model_checkpoint = os.path.join(args.run_dir, 'checkpoint_latest.pth')


    model_without_ddp = utils.DistributedParallel_Model(model, args.local_rank)

    if args.world_size > 1:
        for key in model_without_ddp.model:
            utils.check_ddp_consistency(model_without_ddp.model[key])

    for key in model_without_ddp.model:
        params = [p for p in model_without_ddp.model[key].parameters() if p.requires_grad]
        cnt_params = sum([p.numel() for p in params])
        logger.info("params {key}: {cnt_params}".format(key=key, cnt_params=cnt_params))

    logger.info('begin training ...')

    model_without_ddp.trainer(train_dataloader, valid_dataloader, builder.get_max_epoch(), builder.get_max_step(), checkpoint_savedir=args.relative_checkpoint_dir if model_without_ddp.use_ceph else args.run_dir, resume=args.resume)


def main(args):
    if args.world_size > 1:
        utils.init_distributed_mode(args)
    else:
        args.rank = 0
        args.distributed = False
        # args.local_rank = 0
        torch.cuda.set_device(args.local_rank)
    desc = f'world_size{args.world_size:d}'

    if args.desc is not None:
        desc += f'-{args.desc}'

    ##############################################
    # ## trace net output unused ##
    # os.environ['TORCH_DISTRIBUTED_DEBUG'] = "INFO"
    ## or trace unused params in this way ##
    # for name, param in self.model[list(self.model.keys())[0]].named_parameters():
    #         if param.grad is None:
    #             print(name)
    ###############################################
    alg_dir = args.cfg.split("/")[-1].split(".")[0]
    ## check if outdir exists ##
    if not os.path.exists(args.outdir):
        raise ValueError(f"outdir {args.outdir} not exists, should link a oss path")
    args.outdir = args.outdir + "/" + alg_dir
    run_dir = os.path.join(args.outdir, f'{desc}')
    relative_checkpoint_dir = alg_dir + "/" + f'{desc}'
    args.relative_checkpoint_dir = relative_checkpoint_dir
    print(run_dir)
    os.makedirs(run_dir, exist_ok=True)
    train_config_file = os.path.join(run_dir, 'training_options.yaml')

    if not args.resume:
        print("load yaml from config")
        with open(args.cfg, 'r') as cfg_file:
            cfg_params = yaml.load(cfg_file, Loader = yaml.FullLoader)
    else:
        print("load yaml from resume")
        train_config_file = args.resume_cfg_file
        with open(train_config_file, 'r') as cfg_file:
            cfg_params = yaml.load(cfg_file, Loader = yaml.FullLoader)
        del_keys = []
        for key in cfg_params:
            if key in args:
                del_keys.append(key)
        for key in del_keys:
            del cfg_params[key]

    # cfg_params['dataloader']['num_workers'] = args.per_cpus

    if args.rank == 0:
        with open(os.path.join(run_dir, 'training_options.yaml'), 'wt') as f:
            yaml.dump(vars(args), f, indent=2, sort_keys=False)
            yaml.dump(cfg_params, f, indent=2, sort_keys=False)

    args.cfg_params = cfg_params
    args.run_dir = run_dir

    args.cfg_params['model']['params']['run_dir']=args.run_dir

    if args.debug:
        args.cfg_params['model']['params']['debug'] = True

    if args.visual_vars is not None:
        args.cfg_params['model']['params']['visual_vars'] = args.visual_vars
        args.cfg_params['model']['params']['run_dir'] = args.run_dir

    if 'sampler' in args.cfg_params.keys():
        args.cfg_params['model']['params']['sampler_type'] = args.cfg_params['sampler']['type']

    print('Launching processes...')
    subprocess_fn(args)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--tensor_model_parallel_size', type = int,     default = 1,                                            help = 'tensor_model_parallel_size')
    parser.add_argument('--resume',                     action = "store_true",                                                  help = 'resume')
    parser.add_argument('--resume_from_config',         action = "store_true",                                                  help = 'resume from config')
    parser.add_argument('--seed',                       type = int,     default = 0,                                            help = 'seed')
    parser.add_argument('--cuda',                       type = int,     default = 0,                                            help = 'cuda id')
    parser.add_argument('--world_size',                 type = int,     default = 1,                                            help = 'Number of progress')
    parser.add_argument('--per_cpus',                   type = int,     default = 1,                                            help = 'Number of perCPUs to use')
    parser.add_argument('--local_rank',                 type=int,       default=0,                                              help='local rank')
    # parser.add_argument('--world_size',     type = int,     default = -1,                                           help = 'number of distributed processes')
    parser.add_argument('--init_method',                type = str,     default='tcp://127.0.0.1:23456',                        help = 'multi process init method')
    parser.add_argument('--outdir',                     type = str,     default='/mnt/petrelfs/chenkang/code/wpredict/IF_WKP',  help = 'Where to save the results')
    parser.add_argument('--cfg', '-c',                  type = str,     default = os.path.join('configs', 'default.yaml'),      help = 'path to the configuration file')
    parser.add_argument('--desc',                       type=str,       default='STR',                                          help = 'String to include in result dir name')
    parser.add_argument('--visual_vars',                nargs='+',       default=None,                                          help = 'visual vars')
    # debug mode for quick debug #
    parser.add_argument('--debug', action='store_true', help='debug or not')
    parser.add_argument('--resume_checkpoint', type=str, default=None, help='resume checkpoint')
    parser.add_argument('--resume_cfg_file', type=str, default=None, help='resume from cfg file')

    args = parser.parse_args()

    main(args)