grt/train.py at main · WiseLabCMU/grt · 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
"""Train radar model."""

import json
import os
import time
from argparse import ArgumentParser

import lightning as L
import torch
from lightning.pytorch.callbacks import EarlyStopping, ModelCheckpoint
from lightning.pytorch.loggers import TensorBoardLogger
from lightning.pytorch.strategies import DDPStrategy

from deepradar import DeepRadar, config


def _parse():
    p = ArgumentParser(description="Train radar model.")

    g = p.add_argument_group("Path")
    g.add_argument(
        "-p", "--path", default="data", help="Root dataset directory.")
    g.add_argument(
        "-o", "--out", default="results", help="Root results directory.")

    g = p.add_argument_group("Training")
    g.add_argument(
        "-c", "--cfg", nargs='+', default=None,
        help="Training configuration; see `deepradar.config` for parsing "
        "rules. Must be specified unless resuming with "
        "`--checkpoint <checkpoint>`.")
    g.add_argument(
        "--cfg_dir", default="config", help="Configuration base directory.")
    g.add_argument(
        "-k", "--checkpoint", default=None,
        help="Checkpoint to load, if specified. Should have the structure "
        "`<folder>/checkpoints/<checkpoint>.ckpt`, where `folder` contains a "
        "`hparams.yaml` file.")
    g.add_argument(
        "--epochs", default=-1, type=int, help="Maximum number of epochs.")
    g.add_argument(
        "--metric", default="loss/val",
        help="Metric to watch for convergence (e.g. `loss/val`).")
    g.add_argument(
        "--patience", default=3, type=int,
        help="Stop after this many validation checks with no improvement.")
    g.add_argument(
        "--find_unused", default=False, action='store_true',
        help="Find unused parameters during training; only necessary for some "
        "models with some underlying library bugs.")

    g = p.add_argument_group("Fine tuning")
    g.add_argument(
        "-b", "--base_model", default=None,
        help="Base model to load, if specified. Should be a "
        "experiment directory containing a `hparams.yaml` file and "
        "`checkpoints` directory.")
    g.add_argument(
        "-d", "--load_decoder", default=False, action='store_true',
        help="Load decoder weights (skipping `unpatch.*`) as well.")
    g.add_argument(
        "--load_full_decoder", default=False, action='store_true',
        help="Load decoder weights (including unpatch.*).")
    g.add_argument(
        "-f", "--freeze", action='store_true', default=False,
        help="Freeze encoder (i.e. don't allow tuning the encoder).")

    g = p.add_argument_group("Logging")
    g.add_argument(
        "-n", "--name", default=None,
        help="Method name (for experiment tracking only).")
    g.add_argument(
        "-v", "--version", default=None,
        help="Experiment version (for experiment tracking only).")
    g.add_argument(
        "--val_interval", default=0.5, type=float,
        help="Validation interval, as a fraction of each epoch.")
    g.add_argument(
        "--log_example_interval", default=500, type=int,
        help="Interval to log example train images.")
    g.add_argument(
        "--log_interval", default=100, type=int,
        help="Logging interval for training statistics.")
    g.add_argument(
        "--num_checkpoints", default=-1, type=int,
        help="Number of checkpoints to save.")

    g = p.add_argument_group("Environment")
    g.add_argument(
        "-e", "--environment", default="local",
        help="Current system environment. Known options: local (generic "
        "system with possibly multiple GPUs), psc (bridges-2 @ PSC).")
    g.add_argument(
        "--workers", default=None, type=int,
        help="Number of dataloader workers. By default, the dataloader will "
        "use the number of (virtual) cores in the system.")

    return p


def _main(args):

    if args.environment == "psc":
        torch.multiprocessing.set_sharing_strategy('file_system')

    if args.cfg is None:
        if args.checkpoint is not None:
            experiment_dir = os.path.dirname(os.path.dirname(args.checkpoint))
            args.cfg = [os.path.join(experiment_dir, "hparams.yaml")]
        else:
            print(
                "Must specify a `config.yaml` file if not resuming training.")
            exit(1)
    else:
        args.cfg = [os.path.join(args.cfg_dir, c) for c in args.cfg]

    # Resume training
    model_cfg = config.load_config(*args.cfg)
    if args.checkpoint is None:
        model = DeepRadar(**model_cfg)
    else:
        model = DeepRadar.load_from_checkpoint(
            args.checkpoint, hparams_file=args.cfg[0])

    # Use base model
    if args.base_model is not None:
        base = DeepRadar.load_from_experiment(args.base_model, checkpoint=None)
        model.encoder.load_state_dict(base.encoder.state_dict())
        if args.freeze:
            model.encoder.freeze()

        if args.load_decoder:
            decoder_params = {
                k: v for k, v in base.decoder.state_dict().items()
                if not k.startswith("unpatch")}
            model.decoder.load_state_dict(decoder_params, strict=False)

        if args.load_full_decoder:
            model.decoder.load_state_dict(
                base.decoder.state_dict(), strict=False)

    # Metadata/logging-related config bypasses save_hyperparameters
    model.configure(log_interval=args.log_example_interval, num_examples=16)

    data = model.get_dataset(args.path, n_workers=args.workers)
    checkpoint = ModelCheckpoint(
        save_top_k=args.num_checkpoints, monitor=args.metric,
        save_last=True, dirpath=None)
    stopping = EarlyStopping(
        monitor=args.metric, min_delta=0.0, patience=args.patience, mode="min")
    logger = TensorBoardLogger(
        args.out, name=args.name, version=args.version,
        default_hp_metric=False)
    strategy = DDPStrategy(find_unused_parameters=args.find_unused)
    trainer = L.Trainer(
        logger=logger, log_every_n_steps=args.log_interval,
        callbacks=[checkpoint, stopping], max_steps=-1, max_epochs=args.epochs,
        val_check_interval=args.val_interval, strategy=strategy,
        precision="16-mixed")

    start = time.perf_counter()
    trainer.fit(model=model, datamodule=data)
    duration = time.perf_counter() - start

    with open(os.path.join(logger.log_dir, "meta.json"), 'w') as f:
        json.dump({
            "best": os.path.basename(checkpoint.best_model_path),
            "duration": duration
        }, f)


if __name__ == '__main__':
    torch.set_float32_matmul_precision('high')
    _main(_parse().parse_args())