/
main_ddp.py
400 lines (340 loc) · 15.6 KB
/
main_ddp.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
import os
os.environ["CUDA_VISIBLE_DEVICES"]="3,4"
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
import resource
import numpy as np
import datetime
import torch
import torch.multiprocessing as mp
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.nn.parallel import DistributedDataParallel
from torch.utils.tensorboard import SummaryWriter
from graph4nlp.pytorch.models.graph2seq import Graph2Seq
from graph4nlp.pytorch.models.graph2seq_loss import Graph2SeqLoss
from graph4nlp.pytorch.modules.evaluation import BLEU
from args import get_args
from build_model import get_model
from dataset import IWSLT14Dataset
from utils import WarmupCosineSchedule, get_log, wordid2str
import random
import numpy as np
import torch
import time
class set_worker_seed_builder():
def __init__(self, rank):
self.rank = rank
def __call__(self, worker_id):
base_seed = time.time_ns() % 666666
worker_seed = base_seed + worker_id + self.rank * 10
random.seed(worker_seed)
np.random.seed(worker_seed + 1)
torch.manual_seed(worker_seed + 2)
torch.cuda.manual_seed(worker_seed + 3)
class NMT:
def __init__(self, opt):
super(NMT, self).__init__()
self.opt = opt
self.use_copy = self.opt["decoder_args"]["rnn_decoder_share"]["use_copy"]
assert self.use_copy is False, print("Copy is not fit to NMT")
self.use_coverage = self.opt["decoder_args"]["rnn_decoder_share"]["use_coverage"]
self.distributed = True
torch.distributed.init_process_group(
backend="nccl",
init_method=self.opt["init_method"],
rank=self.opt["rank"],
world_size=self.opt["world_size"],
timeout=datetime.timedelta(0, 120) # 120 seconds
)
self.device = torch.device("cuda:{}".format(self.opt["rank"]))
torch.cuda.set_device(self.opt["rank"])
print('process {}/{} initialized.'.format(self.opt["rank"] + 1, self.opt["world_size"]))
assert self.opt["rank"] == torch.distributed.get_rank()
assert self.opt["world_size"] == torch.distributed.get_world_size()
if self.opt["rank"] == 0:
sid = datetime.datetime.now().strftime('%b%d_%H-%M-%S')
server_store = torch.distributed.TCPStore(
host_name=self.opt["master_addr"],
port=5678,
world_size=self.opt["world_size"],
is_master=True,
timeout=datetime.timedelta(seconds=30)
)
server_store.set("sid", sid)
self._build_logger(self.opt["log_dir"])
self._build_dataloader()
# os.makedirs(os.path.join(self.run_path, 'checkpoints'), exist_ok=True)
# os.makedirs(os.path.join(self.run_path, 'samples'), exist_ok=True)
# save_yaml(os.path.join(self.run_path, 'config.yml'), self.args.config)
# self.writer = SummaryWriter(log_dir=os.path.join(self.run_path, 'tb'))
# self.writer.add_text('config', str(self.args.config))
# self.writer.flush()
# os.makedirs(os.path.join(self.run_path, 'log'), exist_ok=True)
# self.logger = get_log(os.path.join(self.run_path, 'log/log.txt'))
# sync the sid and run_path for other processes
else:
client_store = torch.distributed.TCPStore(
host_name=self.opt["master_addr"],
port=5678,
world_size=self.opt["world_size"],
is_master=False,
timeout=datetime.timedelta(seconds=30)
)
sid = client_store.get("sid").decode("utf-8")
self._build_logger(self.opt["log_dir"])
torch.distributed.barrier()
self._build_dataloader()
self._build_model()
self._build_optimizer()
self._build_evaluation()
self._build_loss_function()
def _build_logger(self, log_dir):
log_path = os.path.join(log_dir, self.opt["name"])
logger_path = os.path.join(log_path, "txt")
tensorboard_path = os.path.join(log_path, "tensorboard")
if not os.path.exists(logger_path):
os.makedirs(logger_path)
if not os.path.exists(tensorboard_path):
os.makedirs(tensorboard_path)
self.logger = get_log(logger_path + "log.txt")
self.writer = SummaryWriter(log_dir=tensorboard_path)
def _build_dataloader(self):
dataset = IWSLT14Dataset(
root_dir=self.opt["graph_construction_args"]["graph_construction_share"]["root_dir"],
val_split_ratio=self.opt["val_split_ratio"],
merge_strategy=self.opt["graph_construction_args"]["graph_construction_private"][
"merge_strategy"
],
edge_strategy=self.opt["graph_construction_args"]["graph_construction_private"][
"edge_strategy"
],
seed=self.opt["seed"],
word_emb_size=self.opt["word_emb_size"],
share_vocab=self.opt["share_vocab"],
graph_name=self.opt["graph_construction_args"]["graph_construction_share"][
"graph_name"
],
topology_subdir=self.opt["graph_construction_args"]["graph_construction_share"][
"topology_subdir"
],
)
self.train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset=dataset.train,
num_replicas=self.opt["world_size"],
rank=self.opt["rank"],
shuffle=True,
seed=0, # must be same for all process
drop_last=True,
)
self.train_dataloader = DataLoader(dataset.train, sampler=self.train_sampler, pin_memory=True,
persistent_workers=False,
worker_init_fn=set_worker_seed_builder(self.opt["rank"]),
batch_size=self.opt["batch_size"],
num_workers=0,
collate_fn=dataset.collate_fn)
self.val_dataloader = DataLoader(
dataset.val,
batch_size=self.opt["batch_size"],
shuffle=False,
num_workers=8,
collate_fn=dataset.collate_fn,
)
self.test_dataloader = DataLoader(
dataset.test,
batch_size=self.opt["batch_size"],
shuffle=False,
num_workers=8,
collate_fn=dataset.collate_fn,
)
self.vocab = dataset.vocab_model
def _build_model(self):
self.model = get_model(self.opt, vocab_model=self.vocab, device=self.device).to(self.device)
self.model = DistributedDataParallel(self.model, device_ids=[self.device], output_device=self.device)
def _build_optimizer(self):
parameters = [p for p in self.model.parameters() if p.requires_grad]
self.optimizer = optim.Adam(parameters, lr=self.opt["learning_rate"])
self.scheduler = WarmupCosineSchedule(
self.optimizer, warmup_steps=self.opt["warmup_steps"], t_total=self.opt["max_steps"]
)
def _build_evaluation(self):
self.metrics = {"BLEU": BLEU(n_grams=[1, 2, 3, 4])}
def _build_loss_function(self):
self.loss = Graph2SeqLoss(
ignore_index=self.vocab.out_word_vocab.PAD,
use_coverage=self.use_coverage,
coverage_weight=0.3,
)
def train(self):
max_score = -1
self._best_epoch = -1
self.global_steps = 0
for epoch in range(200):
print("epoch: {}, rank: {}".format(epoch, self.opt["rank"]), flush=True)
self.model.train()
self.train_epoch(epoch, split="train")
# self._adjust_lr(epoch)
score = self.evaluate(epoch=epoch, split="val")
if score >= max_score and self.opt["rank"] == 0:
self.logger.info("Best model saved, epoch {}".format(epoch))
self.model.module.save_checkpoint(
os.path.join("examples/pytorch/nmt/save", self.opt["name"]), "best.pth"
)
# self._best_epoch = epoch
max_score = max(max_score, score)
torch.distributed.barrier()
# if epoch >= 30 and self._stop_condition(epoch):
# break
print("epoch: {}, rank: {}".format(epoch, self.opt["rank"]), "===========", flush=True)
return max_score
def _stop_condition(self, epoch, patience=20):
return epoch > patience + self._best_epoch
def _adjust_lr(self, epoch):
def set_lr(optimizer, decay_factor):
for group in optimizer.param_groups:
group["lr"] = group["lr"] * decay_factor
epoch_diff = epoch - self.opt["lr_start_decay_epoch"]
if epoch_diff >= 0 and epoch_diff % self.opt["lr_decay_per_epoch"] == 0:
if self.opt["learning_rate"] > self.opt["min_lr"]:
set_lr(self.optimizer, self.opt["lr_decay_rate"])
self.opt["learning_rate"] = self.opt["learning_rate"] * self.opt["lr_decay_rate"]
self.logger.info("Learning rate adjusted: {:.5f}".format(self.opt["learning_rate"]))
def train_epoch(self, epoch, split="train"):
assert split in ["train"]
self.logger.info("Start training in split {}, Epoch: {}".format(split, epoch))
loss_collect = []
dataloader = self.train_dataloader
step_all_train = len(dataloader)
self.train_sampler.set_epoch(epoch)
for step, data in enumerate(dataloader):
graph, tgt = data["graph_data"], data["tgt_seq"]
tgt = tgt.to(self.device)
graph = graph.to(self.device)
oov_dict = None
prob, enc_attn_weights, coverage_vectors = self.model(graph, tgt, oov_dict=oov_dict)
loss = self.loss(
logits=prob,
label=tgt,
enc_attn_weights=enc_attn_weights,
coverage_vectors=coverage_vectors,
)
# add graph regularization loss if available
if graph.graph_attributes.get("graph_reg", None) is not None:
loss = loss + graph.graph_attributes["graph_reg"]
loss_collect.append(loss.item())
self.global_steps += 1
if step % self.opt["loss_display_step"] == 0 and step != 0:
if self.opt["rank"] == 0:
self.logger.info(
"Epoch {}: [{} / {}] loss: {:.3f}".format(
epoch, step, step_all_train, np.mean(loss_collect)
)
)
loss_collect = []
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
self.scheduler.step()
if self.opt["rank"] == 0:
self.writer.add_scalar(
"train/loss", scalar_value=loss.item(), global_step=self.global_steps
)
self.writer.add_scalar(
"train/lr", scalar_value=self.scheduler.get_lr()[0], global_step=self.global_steps
)
def remove_bpe(self, str_with_subword):
if isinstance(str_with_subword, list):
return [self.remove_bpe(ss) for ss in str_with_subword]
symbol = "@@ "
return str_with_subword.replace(symbol, "").strip()
@torch.no_grad()
def evaluate(self, epoch, split="val"):
self.model.eval()
pred_collect = []
gt_collect = []
assert split in ["val", "test"]
dataloader = self.val_dataloader if split == "val" else self.test_dataloader
for data in dataloader:
graph, gt_str = data["graph_data"], data["output_str"]
graph = graph.to(self.device)
oov_dict = None
ref_dict = self.vocab.out_word_vocab
prob, _, _ = self.model(graph, oov_dict=oov_dict)
pred = prob.argmax(dim=-1)
pred_str = wordid2str(pred.detach().cpu(), ref_dict)
pred_collect.extend(self.remove_bpe(pred_str))
gt_collect.extend(self.remove_bpe(gt_str))
score = self.metrics["BLEU"].calculate_scores(ground_truth=gt_collect, predict=pred_collect)
self.logger.info(
"Evaluation results in `{}` split: BLEU-1:{:.4f}\tBLEU-2:{:.4f}\tBLEU-3:{:.4f}\t"
"BLEU-4:{:.4f}".format(split, score[0][0], score[0][1], score[0][2], score[0][3])
)
self.writer.add_scalar(split + "/BLEU@1", score[0][0] * 100, global_step=epoch)
self.writer.add_scalar(split + "/BLEU@2", score[0][1] * 100, global_step=epoch)
self.writer.add_scalar(split + "/BLEU@3", score[0][2] * 100, global_step=epoch)
self.writer.add_scalar(split + "/BLEU@4", score[0][3] * 100, global_step=epoch)
return score[0][-1]
@torch.no_grad()
def translate(self):
self.model.eval()
pred_collect = []
gt_collect = []
dataloader = self.test_dataloader
for data in dataloader:
batch_graph, gt_str = data["graph_data"], data["output_str"]
oov_dict = None
ref_dict = self.vocab.out_word_vocab
batch_graph = batch_graph.to(self.device)
pred = self.model.translate(
batch_graph=batch_graph, oov_dict=oov_dict, beam_size=3, topk=1
)
pred_ids = pred[:, 0, :] # we just use the top-1
pred_str = wordid2str(pred_ids.detach().cpu(), ref_dict)
pred_collect.extend(pred_str)
gt_collect.extend(gt_str)
score = self.metrics["BLEU"].calculate_scores(ground_truth=gt_collect, predict=pred_collect)
self.logger.info(
"Evaluation results in `{}` split: BLEU-1:{:.4f}\tBLEU-2:{:.4f}\tBLEU-3:{:.4f}\t"
"BLEU-4:{:.4f}".format("test", score[0][0], score[0][1], score[0][2], score[0][3])
)
self.writer.add_scalar("test" + "/BLEU@1", score[0][0] * 100, global_step=0)
self.writer.add_scalar("test" + "/BLEU@2", score[0][1] * 100, global_step=0)
self.writer.add_scalar("test" + "/BLEU@3", score[0][2] * 100, global_step=0)
self.writer.add_scalar("test" + "/BLEU@4", score[0][3] * 100, global_step=0)
return score
def run(rank, opt, distributed_meta_info):
opt["rank"] = rank
opt["init_method"] = distributed_meta_info["init_method"]
opt["world_size"] = distributed_meta_info["world_size"]
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
runner = NMT(opt)
runner.train()
if __name__ == "__main__":
opt = get_args()
world_size = int(opt["world_size"])
# runner.logger.info("------ Running Training ----------")
# runner.logger.info("\tRunner name: {}".format(opt["name"]))
# save_config(opt, os.path.join(opt["checkpoint_save_path"], opt["name"]))
init_method = "tcp://{}:{}".format(opt["master_addr"], opt["master_port"])
distributed_meta_info = {
"world_size": world_size,
"master_addr": opt["master_addr"],
"init_method": init_method,
# rank will be added in spawned processes
}
mp.freeze_support()
mp.spawn(
fn=run,
args=(opt, distributed_meta_info),
nprocs=world_size,
join=True,
daemon=False
)
# max_score = runner.train()
# runner.logger.info("Train finish, best val score: {:.3f}".format(max_score))
# runner.model = Graph2Seq.load_checkpoint(
# os.path.join("examples/pytorch/nmt/save", opt["name"]), "best.pth"
# ).to(runner.device)
# runner.translate()