import collections as col
import logging
import numpy as np
import random
from tqdm import tqdm, trange
import torch
from torch.utils.data import Dataset, TensorDataset, DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from .core import InputFeatures, Batch, InputExample, TokenizedExample, random_word
from pytorch_pretrained_bert.utils import truncate_seq_pair
from shared.runners import warmup_linear
logger = logging.getLogger(__name__)
def tokenize_example(example, tokenizer):
tokens_a = tokenizer.tokenize(example.text_a)
if example.text_b:
tokens_b = tokenizer.tokenize(example.text_b)
else:
tokens_b = example.text_b
return TokenizedExample(
guid=example.guid,
tokens_a=tokens_a,
tokens_b=tokens_b,
is_next=example.is_next,
)
def convert_example_to_features(example, tokenizer, max_seq_length, select_prob=0.15):
if isinstance(example, InputExample):
example = tokenize_example(example, tokenizer)
tokens_a = example.tokens_a
tokens_b = example.tokens_b
truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
tokens_a, t1_label = random_word(tokens_a, tokenizer, select_prob=select_prob)
tokens_b, t2_label = random_word(tokens_b, tokenizer, select_prob=select_prob)
lm_label_ids = ([-1] + t1_label + [-1] + t2_label + [-1])
tokens = []
segment_ids = []
tokens.append("[CLS]")
segment_ids.append(0)
for token in tokens_a:
tokens.append(token)
segment_ids.append(0)
tokens.append("[SEP]")
segment_ids.append(0)
assert len(tokens_b) > 0
for token in tokens_b:
tokens.append(token)
segment_ids.append(1)
tokens.append("[SEP]")
segment_ids.append(1)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
while len(input_ids) < max_seq_length:
input_ids.append(0)
input_mask.append(0)
segment_ids.append(0)
lm_label_ids.append(-1)
assert len(input_ids) == max_seq_length
assert len(input_mask) == max_seq_length
assert len(segment_ids) == max_seq_length
assert len(lm_label_ids) == max_seq_length
features = InputFeatures(
guid=example.guid,
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
lm_label_ids=lm_label_ids,
is_next=example.is_next,
tokens=tokens,
)
return features
def convert_examples_to_features(examples, max_seq_length, tokenizer,
select_prob=0.15, verbose=True):
features = []
for (ex_index, example) in enumerate(tqdm(examples)):
feature_instance = convert_example_to_features(
example=example,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
select_prob=select_prob,
)
if verbose and ex_index < 5:
logger.info("*** Example ***")
logger.info("guid: %s" % example.guid)
logger.info("tokens: %s" % " ".join([str(x) for x in feature_instance.tokens]))
logger.info("input_ids: %s" % " ".join([str(x) for x in feature_instance.input_ids]))
logger.info("input_mask: %s" % " ".join([str(x) for x in feature_instance.input_mask]))
logger.info(
"segment_ids: %s" % " ".join([str(x) for x in feature_instance.segment_ids]))
logger.info("is_next: %s" % example.is_next)
logger.info("lm_label_ids: %s " % " ".join([
str(x) for x in feature_instance.lm_label_ids]))
features.append(feature_instance)
return features
def convert_to_dataset(features):
full_batch = features_to_data(features)
dataset_ls = [full_batch.input_ids, full_batch.input_mask,
full_batch.segment_ids, full_batch.lm_label_ids]
if full_batch.is_next is not None:
dataset_ls.append(full_batch.is_next)
dataset = TensorDataset(*dataset_ls)
return dataset, full_batch.tokens
def features_to_data(features):
if features[0].is_next is not None:
is_next = torch.tensor([f.is_next for f in features], dtype=torch.long)
else:
is_next = None
return Batch(
input_ids=torch.tensor([f.input_ids for f in features], dtype=torch.long),
input_mask=torch.tensor([f.input_mask for f in features], dtype=torch.long),
segment_ids=torch.tensor([f.segment_ids for f in features], dtype=torch.long),
lm_label_ids=torch.tensor([f.lm_label_ids for f in features], dtype=torch.long),
is_next=is_next,
tokens=[f.tokens for f in features],
)
class HybridLoader:
def __init__(self, dataloader, tokens):
self.dataloader = dataloader
self.tokens = tokens
def __iter__(self):
batch_size = self.dataloader.batch_size
for i, batch in enumerate(self.dataloader):
if len(batch) == 4:
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
elif len(batch) == 3:
input_ids, input_mask, segment_ids, lm_label_ids = batch
is_next = None
else:
raise RuntimeError()
batch_tokens = self.tokens[i * batch_size: (i+1) * batch_size]
yield Batch(
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
lm_label_ids=lm_label_ids,
is_next=is_next,
tokens=batch_tokens,
)
def __len__(self):
return len(self.dataloader)
"""
class LMDataset(Dataset):
def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8",
corpus_lines=None, on_memory=True):
self.vocab = tokenizer.vocab
self.tokenizer = tokenizer
self.seq_len = seq_len
self.on_memory = on_memory
self.corpus_lines = corpus_lines # number of non-empty lines in input corpus
self.corpus_path = corpus_path
self.encoding = encoding
self.current_doc = 0 # to avoid random sentence from same doc
# for loading samples directly from file
self.sample_counter = 0 # used to keep track of full epochs on file
self.line_buffer = None # keep second sentence of a pair in memory and use as first sentence in next pair
# for loading samples in memory
self.current_random_doc = 0
self.num_docs = 0
self.sample_to_doc = [] # map sample index to doc and line
"""
class LMDataset(Dataset):
def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=True):
self.vocab = tokenizer.vocab
self.tokenizer = tokenizer
self.seq_len = seq_len
self.on_memory = on_memory
self.corpus_lines = corpus_lines # number of non-empty lines in input corpus
self.corpus_path = corpus_path
self.encoding = encoding
self.current_doc = 0 # to avoid random sentence from same doc
# for loading samples directly from file
self.sample_counter = 0 # used to keep track of full epochs on file
self.line_buffer = None # keep second sentence of a pair in memory and use as first sentence in next pair
# for loading samples in memory
self.current_random_doc = 0
self.num_docs = 0
self.sample_to_doc = [] # map sample index to doc and line
# load samples into memory
if on_memory:
self.all_docs = []
doc = []
self.corpus_lines = 0
with open(corpus_path, "r", encoding=encoding) as f:
for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
line = line.strip()
if line == "":
self.all_docs.append(doc)
doc = []
#remove last added sample because there won't be a subsequent line anymore in the doc
self.sample_to_doc.pop()
else:
#store as one sample
sample = {"doc_id": len(self.all_docs),
"line": len(doc)}
self.sample_to_doc.append(sample)
doc.append(line)
self.corpus_lines = self.corpus_lines + 1
# if last row in file is not empty
if self.all_docs[-1] != doc:
self.all_docs.append(doc)
self.sample_to_doc.pop()
self.num_docs = len(self.all_docs)
# load samples later lazily from disk
else:
if self.corpus_lines is None:
with open(corpus_path, "r", encoding=encoding) as f:
self.corpus_lines = 0
for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
if line.strip() == "":
self.num_docs += 1
else:
self.corpus_lines += 1
# if doc does not end with empty line
if line.strip() != "":
self.num_docs += 1
self.file = open(corpus_path, "r", encoding=encoding)
self.random_file = open(corpus_path, "r", encoding=encoding)
def __len__(self):
# last line of doc won't be used, because there's no "nextSentence". Additionally, we start counting at 0.
return self.corpus_lines - self.num_docs - 1
def __getitem__(self, item):
cur_id = self.sample_counter
self.sample_counter += 1
if not self.on_memory:
# after one epoch we start again from beginning of file
if cur_id != 0 and (cur_id % len(self) == 0):
self.file.close()
self.file = open(self.corpus_path, "r", encoding=self.encoding)
t1, t2, is_next_label = self.random_sent(item)
"""
# tokenize
tokens_a = self.tokenizer.tokenize(t1)
tokens_b = self.tokenizer.tokenize(t2)
# combine to one sample
cur_example = TokenizedExample(
guid=cur_id, tokens_a=tokens_a, tokens_b=tokens_b, is_next=is_next_label,
)
# transform sample to features
cur_features = convert_example_to_features(
example=cur_example,
max_seq_length=self.seq_len,
tokenizer=self.tokenizer,
)
cur_tensors = (torch.tensor(cur_features.input_ids),
torch.tensor(cur_features.input_mask),
torch.tensor(cur_features.segment_ids),
torch.tensor(cur_features.lm_label_ids),
torch.tensor(cur_features.is_next))
return cur_tensors
"""
return InputExample(
guid=cur_id,
text_a=t1,
text_b=t2,
is_next=is_next_label,
)
def random_sent(self, index):
"""
Get one sample from corpus consisting of two sentences. With prob. 50% these are two subsequent sentences
from one doc. With 50% the second sentence will be a random one from another doc.
:param index: int, index of sample.
:return: (str, str, int), sentence 1, sentence 2, isNextSentence Label
"""
t1, t2 = self.get_corpus_line(index)
if random.random() > 0.5:
label = 0
else:
t2 = self.get_random_line()
label = 1
assert len(t1) > 0
assert len(t2) > 0
return t1, t2, label
def get_corpus_line(self, item):
"""
Get one sample from corpus consisting of a pair of two subsequent lines from the same doc.
:param item: int, index of sample.
:return: (str, str), two subsequent sentences from corpus
"""
t1 = ""
t2 = ""
assert item < self.corpus_lines
if self.on_memory:
sample = self.sample_to_doc[item]
t1 = self.all_docs[sample["doc_id"]][sample["line"]]
t2 = self.all_docs[sample["doc_id"]][sample["line"]+1]
# used later to avoid random nextSentence from same doc
self.current_doc = sample["doc_id"]
return t1, t2
else:
if self.line_buffer is None:
# read first non-empty line of file
while t1 == "" :
t1 = next(self.file).strip()
t2 = next(self.file).strip()
else:
# use t2 from previous iteration as new t1
t1 = self.line_buffer
t2 = next(self.file).strip()
# skip empty rows that are used for separating documents and keep track of current doc id
while t2 == "" or t1 == "":
t1 = next(self.file).strip()
t2 = next(self.file).strip()
self.current_doc = self.current_doc+1
self.line_buffer = t2
assert t1 != ""
assert t2 != ""
return t1, t2
def get_random_line(self):
"""
Get random line from another document for nextSentence task.
:return: str, content of one line
"""
# Similar to original tf repo: This outer loop should rarely go for more than one iteration for large
# corpora. However, just to be careful, we try to make sure that
# the random document is not the same as the document we're processing.
for _ in range(10):
if self.on_memory:
rand_doc_idx = random.randint(0, len(self.all_docs)-1)
rand_doc = self.all_docs[rand_doc_idx]
line = rand_doc[random.randrange(len(rand_doc))]
else:
rand_index = random.randint(1, self.corpus_lines if self.corpus_lines < 1000 else 1000)
#pick random line
for _ in range(rand_index):
line = self.get_next_line()
#check if our picked random line is really from another doc like we want it to be
if self.current_random_doc != self.current_doc:
break
return line
def get_next_line(self):
""" Gets next line of random_file and starts over when reaching end of file"""
try:
line = next(self.random_file).strip()
#keep track of which document we are currently looking at to later avoid having the same doc as t1
if line == "":
self.current_random_doc = self.current_random_doc + 1
line = next(self.random_file).strip()
except StopIteration:
self.random_file.close()
self.random_file = open(self.corpus_path, "r", encoding=self.encoding)
line = next(self.random_file).strip()
return line
class TrainEpochState:
def __init__(self):
self.tr_loss = 0
self.global_step = 0
self.nb_tr_examples = 0
self.nb_tr_steps = 0
class RunnerParameters:
def __init__(self, select_prob, max_seq_length, local_rank, n_gpu, fp16,
learning_rate, gradient_accumulation_steps, t_total, warmup_proportion,
num_train_epochs, train_batch_size):
self.select_prob = select_prob
self.max_seq_length = max_seq_length
self.local_rank = local_rank
self.n_gpu = n_gpu
self.fp16 = fp16
self.learning_rate = learning_rate
self.gradient_accumulation_steps = gradient_accumulation_steps
self.t_total = t_total
self.warmup_proportion = warmup_proportion
self.num_train_epochs = num_train_epochs
self.train_batch_size = train_batch_size
class LMRunner:
def __init__(self, model, optimizer, tokenizer, device, rparams):
self.model = model
self.optimizer = optimizer
self.tokenizer = tokenizer
self.device = device
self.rparams = rparams
def run_train(self, train_examples, verbose=True):
if verbose:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", self.rparams.train_batch_size)
logger.info(" Num steps = %d", self.rparams.t_total)
train_dataloader = self.get_train_dataloader(train_examples, verbose=verbose)
for _ in trange(int(self.rparams.num_train_epochs), desc="Epoch"):
self.run_train_epoch(train_dataloader)
def run_train_val(self, train_examples, val_examples):
epoch_result_dict = col.OrderedDict()
for i in trange(int(self.rparams.num_train_epochs), desc="Epoch"):
train_dataloader = self.get_train_dataloader(train_examples, verbose=False)
self.run_train_epoch(train_dataloader)
epoch_result = self.run_val(val_examples, verbose=False)
del epoch_result["logits"]
epoch_result_dict[i] = epoch_result
return epoch_result_dict
def run_train_epoch(self, train_dataloader):
for _ in self.run_train_epoch_context(train_dataloader):
pass
def run_train_epoch_context(self, train_dataloader):
self.model.train()
train_epoch_state = TrainEpochState()
for step, batch in enumerate(tqdm(train_dataloader, desc="Training")):
self.run_train_step(
step=step,
batch=batch,
train_epoch_state=train_epoch_state,
)
yield step, batch, train_epoch_state
def run_train_step(self, step, batch, train_epoch_state):
batch = batch.to(self.device)
loss = self.model(
batch.input_ids, batch.segment_ids, batch.input_mask,
batch.lm_label_ids, batch.is_next,
)
if self.rparams.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if self.rparams.gradient_accumulation_steps > 1:
loss = loss / self.rparams.gradient_accumulation_steps
if self.rparams.fp16:
self.optimizer.backward(loss)
else:
loss.backward()
train_epoch_state.tr_loss += loss.item()
train_epoch_state.nb_tr_examples += batch.input_ids.size(0)
train_epoch_state.nb_tr_steps += 1
if (step + 1) % self.rparams.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
if self.rparams.fp16:
lr_this_step = self.rparams.learning_rate * warmup_linear(
train_epoch_state.global_step / self.rparams.t_total, self.rparams.warmup_proportion)
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr_this_step
self.optimizer.step()
self.optimizer.zero_grad()
train_epoch_state.global_step += 1
def run_val(self, val_examples, verbose=True):
val_dataloader = self.get_eval_dataloader(val_examples, verbose=verbose)
self.model.eval()
total_eval_loss = 0
nb_eval_steps, nb_eval_examples = 0, 0
all_logits = []
for step, batch in enumerate(tqdm(val_dataloader, desc="Evaluating (Val)")):
batch = batch.to(self.device)
with torch.no_grad():
tmp_eval_loss = self.model(
batch.input_ids, batch.segment_ids, batch.input_mask,
batch.lm_label_ids, batch.is_next,
)
logits = self.model(batch.input_ids, batch.segment_ids, batch.input_mask)
logits = logits.detach().cpu().numpy()
total_eval_loss += tmp_eval_loss.mean().item()
nb_eval_examples += batch.input_ids.size(0)
nb_eval_steps += 1
all_logits.append(logits)
eval_loss = total_eval_loss / nb_eval_steps
all_logits = np.concatenate(all_logits, axis=0)
return {
"logits": all_logits,
"loss": eval_loss,
}
def run_test(self, test_examples, verbose=True):
test_dataloader = self.get_eval_dataloader(test_examples, verbose=verbose)
self.model.eval()
all_logits = []
for step, batch in enumerate(tqdm(test_dataloader, desc="Predictions (Test)")):
batch = batch.to(self.device)
with torch.no_grad():
logits = self.model(batch.input_ids, batch.segment_ids, batch.input_mask)
logits = logits.detach().cpu().numpy()
all_logits.append(logits)
all_logits = np.concatenate(all_logits, axis=0)
return all_logits
def get_train_dataloader(self, train_examples, verbose=True):
train_features = convert_examples_to_features(
examples=train_examples,
max_seq_length=self.rparams.max_seq_length,
tokenizer=self.tokenizer,
select_prob=self.rparams.select_prob,
verbose=verbose,
)
train_data, train_tokens = convert_to_dataset(train_features)
if self.rparams.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(
train_data, sampler=train_sampler, batch_size=self.rparams.train_batch_size,
)
return HybridLoader(train_dataloader, train_tokens)
def get_eval_dataloader(self, eval_examples, verbose=True):
eval_features = convert_examples_to_features(
examples=eval_examples,
max_seq_length=self.rparams.max_seq_length,
tokenizer=self.tokenizer,
select_prob=self.rparams.select_prob,
verbose=verbose,
)
eval_data, eval_tokens = convert_to_dataset(eval_features)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(
eval_data, sampler=eval_sampler, batch_size=self.rparams.eval_batch_size,
)
return HybridLoader(eval_dataloader, eval_tokens)
