File size: 10,597 Bytes
0b32ad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
import argparse
import logging
import math
import os
from copy import deepcopy
from pathlib import Path

import torch
from tqdm import tqdm

from s3prl import Container, Logs, Object, Output
from s3prl.dataset.base import AugmentedDynamicItemDataset, DataLoader
from s3prl.nn import S3PRLUpstream, UpstreamDownstreamModel
from s3prl.sampler import DistributedBatchSamplerWrapper
from s3prl.util.configuration import parse_override, qualname_to_cls
from s3prl.util.seed import fix_random_seeds

device = "cuda" if torch.cuda.is_available() else "cpu"
logger = logging.getLogger(__name__)

DRYRUN_CONFIG = dict(
    Trainer=dict(
        total_steps=10,
        log_step=2,
        valid_step=5,
        save_step=5,
        eval_batch=5,
    ),
)


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("upstream", help="The upstream name. E.g. wav2vec2")
    parser.add_argument(
        "problem",
        help="The problem module. E.g. s3prl.problem.SuperbSID",
    )
    parser.add_argument(
        "dataset_root",
        help="The dataset root of your problem.",
    )
    parser.add_argument("save_to", help="The directory to save checkpoint")
    parser.add_argument("--feature_selection", default="hidden_states")
    parser.add_argument("--n_jobs", type=int, default=6)
    parser.add_argument(
        "--override",
        default=None,
        help=(
            "Override the default_config of the problem module. "
            "E.g. --override ValidSampler.batch_size=4,,TestSampler.batch_size=4"
        ),
    )
    parser.add_argument("--resume", action="store_true")
    parser.add_argument("--dryrun", action="store_true")
    parser.add_argument("--seed", type=int, default=1337)
    args = parser.parse_args()

    fix_random_seeds(args.seed)
    problem = qualname_to_cls(args.problem)
    config = Container(deepcopy(problem.default_config))

    for key, value in vars(args).items():
        if key not in ["override"]:
            config[key] = value

    if args.dryrun:
        config.override(DRYRUN_CONFIG)

    if isinstance(args.override, str) and len(args.override) > 0:
        override_dict = parse_override(args.override)
        config.override(override_dict)

    return problem, config


def main():
    logging.basicConfig(level=logging.INFO)

    problem, config = parse_args()
    save_to = Path(config.save_to)
    save_to.mkdir(exist_ok=True, parents=True)

    # configure any upstream
    upstream = S3PRLUpstream(config.upstream, config.feature_selection)
    stats = Container(upstream_rate=upstream.downsample_rate)

    logger.info("Preparing corpus")
    corpus = problem.Corpus(config.dataset_root, **config.Corpus)
    train_data, valid_data, test_data, corpus_stats = corpus().split(3)
    stats.add(corpus_stats)

    logger.info("Preparing train data")
    train_dataset = AugmentedDynamicItemDataset(train_data, tools=stats)
    train_dataset = problem.TrainData(**config.TrainData)(train_dataset)
    train_sampler = DistributedBatchSamplerWrapper(
        problem.TrainSampler(train_dataset, **config.TrainSampler),
        num_replicas=1,
        rank=0,
    )
    train_dataloader = DataLoader(
        train_dataset,
        train_sampler,
        num_workers=config.n_jobs,
    )
    stats.add(train_dataset.all_tools())

    logger.info("Preparing valid data")
    valid_dataset = AugmentedDynamicItemDataset(valid_data, tools=stats)
    valid_dataset = problem.ValidData(**config.ValidData)(valid_dataset)
    valid_sampler = DistributedBatchSamplerWrapper(
        problem.ValidSampler(valid_dataset, **config.ValidSampler),
        num_replicas=1,
        rank=0,
    )
    valid_dataloader = DataLoader(
        valid_dataset,
        valid_sampler,
        num_workers=12,
    )

    logger.info("Preparing test data")
    test_dataset = AugmentedDynamicItemDataset(test_data, tools=stats)
    test_dataset = problem.TestData(**config.TestData)(test_dataset)
    test_sampler = DistributedBatchSamplerWrapper(
        problem.ValidSampler(test_dataset, **config.TestSampler),
        num_replicas=1,
        rank=0,
    )
    test_dataloader = DataLoader(
        test_dataset,
        test_sampler,
        num_workers=12,
    )

    sorted_ckpt_dirs = sorted(
        [
            file
            for file in save_to.iterdir()
            if file.is_dir() and str(file).endswith(".ckpts")
        ],
        key=os.path.getmtime,
    )

    if config.resume and len(sorted_ckpt_dirs) > 0:
        logger.info("Last checkpoint found. Load model and optimizer from checkpoint")
        task = Object.load_checkpoint(sorted_ckpt_dirs[1] / "task.ckpt").to(device)
    else:
        logger.info("Create a new model")
        downstream = problem.Downstream(
            upstream.output_size,
            **stats,
        )
        model = UpstreamDownstreamModel(upstream, downstream)
        # task = problem.Task(model, **{**stats, **config.Task})
        task = problem.Task(model, **stats, **config.Task)
        task = task.to(device)

    # ALL THE FOLLOWING CODES ARE FOR TRAINER
    # WHICH CAN BE LARGELY SIMPLIFIED WHEN USING OTHER TRAINER PACKAGES

    opt_cls_qualname, opt_cfgs = config.Optimizer.split(1)
    optimizer = qualname_to_cls(opt_cls_qualname)(task.parameters(), **opt_cfgs)
    if config.resume and len(sorted_ckpt_dirs) > 0:
        optimizer.load_state_dict(torch.load(sorted_ckpt_dirs[-1] / "optimizer.ckpt"))

    if config.Trainer.use_valid:
        if config.resume and len(sorted_ckpt_dirs) > 0:
            valid_best_score = torch.load(
                sorted_ckpt_dirs[-1] / "valid_best_score.ckpt"
            )[config.Trainer.valid_metric]
        else:
            valid_best_score = -100000 if config.Trainer.valid_higher_better else 100000

    def save_checkpoint(name):
        ckpt_dir: Path = save_to / f"{name}.ckpts"
        ckpt_dir.mkdir(parents=True, exist_ok=True)
        logger.info(f"Save checkpoint to: {ckpt_dir}")

        task.save_checkpoint(ckpt_dir / "task.ckpt")
        torch.save(optimizer.state_dict(), ckpt_dir / "optimizer.ckpt")
        torch.save(
            {config.Trainer.valid_metric: valid_best_score},
            ckpt_dir / "valid_best_score.ckpt",
        )

    pbar = tqdm(total=config.Trainer.total_steps, desc="Total")
    train_completed = False
    accum_grad_steps = 0
    while not train_completed:
        batch_results = []
        for batch in tqdm(train_dataloader, desc="Train", total=len(train_dataloader)):
            pbar.update(1)
            global_step = pbar.n

            assert isinstance(batch, Output)
            batch = batch.to(device)

            task.train()
            result = task.train_step(**batch)
            assert isinstance(result, Output)

            result.loss /= config.Trainer.gradient_accumulate_steps
            result.loss.backward()

            grad_norm = torch.nn.utils.clip_grad_norm_(
                task.parameters(), max_norm=config.Trainer.gradient_clipping
            )

            if math.isnan(grad_norm):
                logger.warning(f"Grad norm is NaN at step {global_step}")
                optimizer.zero_grad()
                accum_grad_steps = 0
            else:
                accum_grad_steps += 1
                if accum_grad_steps == config.Trainer.gradient_accumulate_steps:
                    optimizer.step()
                    optimizer.zero_grad()
                    accum_grad_steps = 0
                batch_results.append(result.cacheable())

            if global_step % config.Trainer.log_step == 0:
                logs: Logs = task.train_reduction(batch_results).logs
                logger.info(f"[Train] step {global_step}")
                for name, value in logs.Scalar.items():
                    if name == "loss":
                        value *= config.Trainer.gradient_accumulate_steps
                    logger.info(f"{name}: {value}")
                batch_results = []

            if global_step % config.Trainer.valid_step == 0:
                with torch.no_grad():
                    if config.Trainer.use_valid:
                        valid_results = []
                        for batch_idx, batch in enumerate(
                            tqdm(
                                valid_dataloader,
                                desc="Valid",
                                total=len(valid_dataloader),
                            )
                        ):
                            if batch_idx == config.Trainer.get("eval_batch", -1):
                                break
                            batch = batch.to(device)
                            task.eval()
                            result = task.valid_step(**batch)
                            valid_results.append(result.cacheable())

                        logs: Logs = task.valid_reduction(valid_results).slice(1)
                        logger.info(f"[Valid] step {global_step}")
                        for name, value in logs.Scalar.items():
                            logger.info(f"{name}: {value}")
                            if name == config.Trainer.valid_metric:
                                cond1 = config.Trainer.valid_higher_better and (
                                    value > valid_best_score
                                )
                                cond2 = (not config.Trainer.valid_higher_better) and (
                                    value < valid_best_score
                                )
                                if cond1 or cond2:
                                    valid_best_score = value
                                    save_checkpoint("valid_best")

            if (
                global_step % config.Trainer.save_step == 0
                or global_step == config.Trainer.total_steps
            ):
                save_checkpoint(f"global_step_{global_step}")

                if global_step == config.Trainer.total_steps:
                    train_completed = True
                    break

    test_results = []
    for batch_idx, batch in enumerate(
        tqdm(test_dataloader, desc="Test", total=len(test_dataloader))
    ):
        if batch_idx == config.Trainer.get("eval_batch", -1):
            break
        batch = batch.to(device)
        result = task.test_step(**batch)
        test_results.append(result.cacheable())

    logs: Logs = task.test_reduction(test_results).slice(1)
    logger.info(f"[Test] step {global_step}")
    for name, value in logs.Scalar.items():
        logger.info(f"{name}: {value}")


if __name__ == "__main__":
    main()