-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
540 lines (476 loc) · 21.1 KB
/
train.py
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
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
# MIT License
#
# Copyright (c) 2019 seq2struct contributors
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import argparse
import functools
import collections
import datetime
import itertools
import json
import os
import traceback
from typing import Type, List
import _jsonnet
import torch
import numpy as np
# noinspection PyUnresolvedReferences
from torch.cuda.amp import autocast, GradScaler
from torch.utils.data import DataLoader
# noinspection PyUnresolvedReferences
from duorat import datasets
# noinspection PyUnresolvedReferences
from duorat import preproc
# noinspection PyUnresolvedReferences
from duorat import models
# noinspection PyUnresolvedReferences
from duorat.asdl.lang.spider.spider_transition_system import SpiderTransitionSystem
from duorat.types import RATPreprocItem
# noinspection PyUnresolvedReferences
from duorat.utils import optimizers
from duorat.utils import registry, parallelizer
from duorat.utils import random_state
from duorat.utils import saver as saver_mod
from duorat.utils.evaluation import evaluate_default, load_from_lines
from third_party.spider.evaluation import LEVELS
from preprocess import Preprocessor
class Logger:
def __init__(self, log_path=None, reopen_to_flush=False):
self.log_file = None
self.reopen_to_flush = reopen_to_flush
if log_path is not None:
os.makedirs(os.path.dirname(log_path), exist_ok=True)
self.log_file = open(log_path, "a+")
def log(self, msg):
formatted = "[{}] {}".format(
datetime.datetime.now().replace(microsecond=0).isoformat(), msg
)
print(formatted)
if self.log_file:
self.log_file.write(formatted + "\n")
if self.reopen_to_flush:
log_path = self.log_file.name
self.log_file.close()
self.log_file = open(log_path, "a+")
else:
self.log_file.flush()
class Trainer:
def __init__(self, logger, config):
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
self.logger = logger
self.config = config
self.data_random = random_state.RandomContext(
self.config["train"].get("data_seed", None)
)
self.model_random = random_state.RandomContext(
self.config["train"].get("model_seed", None)
)
self.init_random = random_state.RandomContext(
self.config["train"].get("init_seed", None)
)
with self.init_random:
# 0. Construct preprocessors
self.model_preproc = registry.construct(
"preproc", self.config["model"]["preproc"],
)
self.model_preproc.load()
# 1. Construct model
self.model = registry.construct(
"model", self.config["model"], preproc=self.model_preproc,
)
self.model.to(device=device)
def _log_loss(self, last_step, loss):
self.logger.log("Step {}: loss={:.4f}".format(last_step, loss))
def _log_lr(self, last_step, lrs: List[dict]):
for lr in lrs:
self.logger.log(
"Step {}: lr[{}]={:.4f}".format(last_step, lr["name"], lr["value"])
)
def _log_stats(self, last_step, eval_section, stats):
self.logger.log(
"Step {} stats, {}: {}".format(
last_step,
eval_section,
", ".join("{} = {}".format(k, v) for k, v in stats.items()),
)
)
def train(self, modeldir, load_path=None, step=None):
# Save the config info
with open(
os.path.join(
modeldir,
"config-{}.json".format(
datetime.datetime.now().strftime("%Y%m%dT%H%M%S%Z")
),
),
"w",
) as f:
json.dump(self.config, f, sort_keys=True, indent=4)
# slight difference here vs. unrefactored train: The init_random starts over here. Could be fixed if it was important by saving random state at end of init
with self.init_random:
# We may be able to move optimizer and lr_scheduler to __init__ instead. Empirically it works fine. I think that's because saver.restore
# resets the state by calling optimizer.load_state_dict.
# But, if there is no saved file yet, I think this is not true, so might need to reset the optimizer manually?
# For now, just creating it from scratch each time is safer and appears to be the same speed, but also means you have to pass in the config to train which is kind of ugly.
optimizer = registry.construct(
"optimizer",
self.config["optimizer"],
params=[
{
"name": "no-bert",
"params": (
parameters
for name, parameters in self.model.named_parameters()
if not name.startswith("initial_encoder.bert")
),
},
{
"name": "bert",
"params": (
parameters
for name, parameters in self.model.named_parameters()
if name.startswith("initial_encoder.bert")
),
},
],
)
lr_scheduler = registry.construct(
"lr_scheduler",
self.config.get("lr_scheduler", {"name": "noop"}),
optimizer=optimizer,
)
# 1.5. Initialize Automatic Mixed Precision (AMP) training
scaler = GradScaler(enabled=self.config["train"]["amp_enabled"])
# 2. Restore model parameters
saver = saver_mod.Saver(
self.model, optimizer
)
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
if load_path:
last_step, best_val_all_exact = saver.restore(load_path, step=step, map_location=device)
else:
last_step, best_val_all_exact = saver.restore(modeldir, step=step, map_location=device)
if last_step == 0 and self.config["train"].get("initialize_from", False):
saver.restore(self.config["train"]["initialize_from"])
self.logger.log(
"Model initialized from {}".format(
self.config["train"]["initialize_from"]
)
)
else:
self.logger.log(f"Model restored, the last step is {last_step}, best val_all_exact is {best_val_all_exact}")
# 3. Get training data somewhere
with self.data_random:
data_splits = self.config["train"].get("data_splits", None)
if data_splits is not None:
self.logger.log(f"Using custom training data splits: {data_splits}.")
else:
data_splits = [
section for section in self.config["data"] if "train" in section
]
train_data = list(
itertools.chain.from_iterable(
self.model_preproc.dataset(split) for split in data_splits
)
)
self.logger.log(f"{len(train_data)} training examples")
train_data_loader = self._yield_batches_from_epochs(
DataLoader(
train_data,
batch_size=self.config["train"]["batch_size"],
shuffle=True,
drop_last=True,
collate_fn=lambda x: x,
pin_memory=True if torch.cuda.is_available() else False,
num_workers=8 if torch.cuda.is_available() else 0,
)
)
train_eval_data_loader = DataLoader(
train_data,
batch_size=self.config["train"]["eval_batch_size"],
collate_fn=lambda x: x,
pin_memory=True if torch.cuda.is_available() else False,
num_workers=8 if torch.cuda.is_available() else 0,
)
val_data = self.model_preproc.dataset("val")
val_data_loader = DataLoader(
val_data,
batch_size=self.config["train"]["eval_batch_size"],
collate_fn=lambda x: x,
pin_memory=True if torch.cuda.is_available() else False,
num_workers=8 if torch.cuda.is_available() else 0,
)
def _evaluate_model():
# A model that is not evaluated (if eval_on_val=False, or if step < infer_min_n)
# is given a performance of 0
val_all_exact = 0
train_all_exact = 0
if self.config["train"]["eval_on_train"]:
self.logger.log("Evaluate on the training set")
train_all_exact = self._eval_model(modeldir, last_step, train_eval_data_loader, "train")
if self.config["train"]["eval_on_val"]:
self.logger.log("Evaluate on the validation set")
val_all_exact = self._eval_model(modeldir, last_step, val_data_loader, "val")
return val_all_exact, train_all_exact
# Counter for grad aggregation
grad_accumulation_counter = 0
losses = []
# 4. Start training loop
with self.data_random:
self.logger.log("Enter training loop")
for batch in train_data_loader:
# Quit if too long
if last_step >= self.config["train"]["max_steps"]:
break
# Compute and apply gradient
with self.model_random:
with autocast(enabled=self.config["train"]["amp_enabled"]):
loss = self.model.compute_loss(batch)
if loss:
with self.model_random:
with autocast(enabled=self.config["train"]["amp_enabled"]):
loss /= self.config["train"]["n_grad_accumulation_steps"]
scaler.scale(loss).backward()
losses.append(
loss.item() * self.config["train"]["n_grad_accumulation_steps"]
)
grad_accumulation_counter += 1
# Update params every `n_grad_accumulation_steps` times
if (
grad_accumulation_counter
% self.config["train"]["n_grad_accumulation_steps"]
== 0
):
# may call unscale_ here to allow clipping unscaled gradients,
# see https://pytorch.org/docs/stable/notes/amp_examples.html#gradient-accumulation
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
lr_scheduler.update_lr(last_step)
try:
scaler.step(optimizer)
scaler.update()
except RuntimeError as e:
self.logger.log(
f"Caught error '{e}' while updating model. Continuing training ..."
)
optimizer.zero_grad()
last_step += 1
# Report metrics
cur_loss = np.mean(losses)
cur_lrs = [
{"name": param_group["name"], "value": param_group["lr"]}
for param_group in optimizer.param_groups
]
if last_step % self.config["train"]["report_every_n"] == 0:
self._log_stats(
last_step,
"train",
{
"step": last_step,
},
)
self._log_loss(last_step, cur_loss)
self._log_lr(last_step, cur_lrs)
# Evaluate model
if last_step % self.config["train"]["eval_every_n"] == 0 \
and last_step != 0:
# try:
# val_all_exact, train_all_exact = _evaluate_model()
# print('val_all_exact: ', val_all_exact)
# except:
val_all_exact, train_all_exact = 0, 0
# print('infer error')
# Run saver
saver.save(
modeldir,
last_step,
is_best=val_all_exact >= best_val_all_exact,
best_validation_metric=max(val_all_exact, best_val_all_exact)
)
best_val_all_exact = max(val_all_exact, best_val_all_exact)
# Reset the list of losses
losses = []
@staticmethod
def _yield_batches_from_epochs(loader):
while True:
for batch in loader:
yield batch
def _eval_model(self, modeldir, last_step, eval_data_loader, eval_section):
num_eval_items = self.config["train"]["num_eval_items"]
stats = collections.defaultdict(float)
self.model.eval()
with torch.no_grad():
for eval_batch in eval_data_loader:
batch_res = self.model.eval_on_batch(eval_batch)
for k, v in batch_res.items():
stats[k] += v
if num_eval_items and stats["total"] > num_eval_items:
break
self.model.train()
# Divide each stat by 'total'
for k in stats:
if k != "total":
stats[k] /= stats["total"]
if "total" in stats:
del stats["total"]
all_exact = 0
if last_step >= self.config["train"]["infer_min_n"]:
metrics = self._infer(modeldir, last_step, eval_section)
stats.update(
{
"{}_exact".format(level): metrics["total_scores"][level]["exact"]
for level in LEVELS
}
)
all_exact = metrics["total_scores"]["all"]["exact"]
json.dump(metrics,
open(os.path.join(modeldir,
f'eval-{eval_section}-{last_step}.json'),
'w'))
self._log_stats(last_step, eval_section, stats)
return all_exact
def _infer(self, modeldir, last_step, eval_section):
self.logger.log("Inferring...")
orig_data = registry.construct("dataset", self.config["data"][eval_section])
preproc_data: List[RATPreprocItem] = self.model_preproc.dataset(eval_section)
self.model.eval()
with torch.no_grad():
if isinstance(self.model.preproc.transition_system, SpiderTransitionSystem):
# Orig data only used with SpiderTransitionSystem
# print(len(orig_data))
# print(len(preproc_data))
assert len(orig_data) == len(preproc_data)
inferred_lines = list(
self._inner_infer(
model=self.model,
orig_data=orig_data,
preproc_data=preproc_data,
nproc=self.config["train"]["eval_nproc"],
beam_size=self.config["train"]["eval_beam_size"],
decode_max_time_step=self.config["train"][
"eval_decode_max_time_step"
],
)
)
self.model.train()
with open(f"{modeldir}/output-{last_step}", "w") as output_dst:
for line in inferred_lines:
output_dst.write(line)
if isinstance(self.model.preproc.transition_system, SpiderTransitionSystem):
return evaluate_default(orig_data, load_from_lines(inferred_lines))
else:
raise NotImplementedError
@classmethod
def _inner_infer(
cls, model, orig_data, preproc_data, nproc, beam_size, decode_max_time_step
):
if torch.cuda.is_available():
cp = parallelizer.CUDAParallelizer(nproc)
else:
cp = parallelizer.CPUParallelizer(nproc)
return cp.parallel_map(
[
(
functools.partial(
cls._parse_single,
model,
beam_size=beam_size,
decode_max_time_step=decode_max_time_step,
),
list(enumerate(zip(orig_data, preproc_data))),
)
]
)
@staticmethod
def _parse_single(model, item, beam_size, decode_max_time_step):
index, (orig_item, preproc_item) = item
batch = [preproc_item]
try:
beams = model.parse(batch, decode_max_time_step, beam_size)
decoded = []
for beam in beams:
asdl_ast = beam.ast
if isinstance(model.preproc.transition_system, SpiderTransitionSystem):
tree = model.preproc.transition_system.ast_to_surface_code(
asdl_ast=asdl_ast
)
inferred_code = model.preproc.transition_system.spider_grammar.unparse(
tree=tree, spider_schema=orig_item.spider_schema
)
inferred_code_readable = ""
else:
raise NotImplementedError
decoded.append(
{
"model_output": asdl_ast.pretty(),
"inferred_code": inferred_code,
"inferred_code_readable": inferred_code_readable,
"score": beam.score,
}
)
result = {
"index": index,
"beams": decoded,
}
except Exception as e:
result = {"index": index, "error": str(e), "trace": traceback.format_exc()}
return json.dumps(result) + "\n"
def main(
args=None, logdir_suffix: List[str] = None, trainer_class: Type[Trainer] = Trainer
):
parser = argparse.ArgumentParser()
parser.add_argument("--logdir", required='output/xlmr-multilingual')
parser.add_argument("--config", default='configs/duorat/duorat-xlmr-multilingual.jsonnet')
parser.add_argument("--preproc_data_path", default='dataset/pkl/xlmr-multilingual')
parser.add_argument("--load_path", default='')
parser.add_argument("--step", default='')
args = parser.parse_args(args)
config = json.loads(_jsonnet.evaluate_file(args.config))
if logdir_suffix:
args.logdir = os.path.join(args.logdir, *logdir_suffix)
if "model_name" in config:
args.logdir = os.path.join(args.logdir, config["model_name"])
# Initialize the logger
reopen_to_flush = config.get("log", {}).get("reopen_to_flush")
logger = Logger(os.path.join(args.logdir, "log.txt"), reopen_to_flush)
logger.log("Logging to {}".format(args.logdir))
logger.log(f"Overwriting preproc save_path with: {args.preproc_data_path}")
if os.path.exists(args.preproc_data_path):
logger.log("Skip preprocessing..")
else:
logger.log("Running preprocessing...")
sections = config["data"].keys()
keep_vocab = False
preprocessor = Preprocessor(config)
preprocessor.preprocess(sections, keep_vocab)
# Construct trainer and do training
step = None if not args.step else int(args.step)
trainer = trainer_class(logger, config)
if args.load_path != '':
trainer.train(modeldir=args.logdir, load_path=args.load_path, step=step)
else:
trainer.train(modeldir=args.logdir, step=step)
if __name__ == "__main__":
main()