def __init__(self, model_type: str, max_seq_len: int = 128):
self.model_type = model_type
self.max_seq_len = max_seq_len
self.token_indexer = PretrainedBertIndexer(
pretrained_model=self.model_type,
max_pieces=self.max_seq_len,
do_lowercase=True,
)
self.vocab = Vocabulary()
self.token_indexer._add_encoding_to_vocabulary(self.vocab)
self.full_vocab = {v: k for k, v in self.token_indexer.vocab.items()}
def __init__(self,
lazy: bool = False,
question_token_indexers: Dict[str, TokenIndexer] = None,
keep_if_unparsable: bool = True,
tables_file: str = None,
dataset_path: str = 'dataset/database',
load_cache: bool = True,
save_cache: bool = True,
loading_limit=-1):
super().__init__(lazy=lazy)
# default spacy tokenizer splits the common token 'id' to ['i', 'd'], we here write a manual fix for that
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
do_lowercase=True,
use_starting_offsets=True)
self._tokenizer = Bert_Tokenizer(token_indexer.wordpiece_tokenizer)
self._utterance_token_indexers = question_token_indexers or {
'tokens': token_indexer
}
self._keep_if_unparsable = keep_if_unparsable
self._tables_file = tables_file
self._dataset_path = dataset_path
self._load_cache = load_cache
self._save_cache = save_cache
self._loading_limit = loading_limit
def __init__(self,
word_indexer: Optional[TokenIndexer] = None,
is_bert: bool = False,
conceptnet_path: Optional[Path] = None):
super().__init__(lazy=False)
self.pos_indexers = {"pos_tokens": PosTagIndexer()}
self.ner_indexers = {"ner_tokens": NerTagIndexer()}
self.rel_indexers = {
"rel_tokens": SingleIdTokenIndexer(namespace='rel_tokens')
}
if is_bert:
splitter = BertBasicWordSplitter()
else:
splitter = SpacyWordSplitter()
self.tokeniser = WordTokenizer(word_splitter=splitter)
self.word_indexers = {'tokens': word_indexer}
word_splitter = SpacyWordSplitter(pos_tags=True, ner=True, parse=True)
self.word_tokeniser = WordTokenizer(word_splitter=word_splitter)
bert_splitter = BertBasicWordSplitter()
self.bert_tokeniser = WordTokenizer(word_splitter=bert_splitter)
if word_indexer is None:
if is_bert:
word_indexer = PretrainedBertIndexer(
pretrained_model='bert-base-uncased',
truncate_long_sequences=False)
else:
word_indexer = SingleIdTokenIndexer(lowercase_tokens=True)
self.word_indexers = {'tokens': word_indexer}
self.conceptnet = ConceptNet(conceptnet_path=conceptnet_path)
def get_bert_token_indexers(
path_to_bert_weights: str,
maximum_number_of_tokens: int,
is_model_lowercase: bool = True) -> PretrainedBertIndexer:
"""
Retrieving bert based token indexers (which will do sub-word tokenizing)
Parameters
----------
path_to_bert_weights: `str`, required
The path to the pytorch bert weights
maximum_number_of_tokens: `int`, required
The maximum number of tokens (truncation or sliding window based on allennlp design)
is_model_lowercase: `bool` optional (default=True)
Force lower casing the input to the model
Returns
----------
Returns the required instance of :class:`PretrainedBertIndexer`.
"""
token_indexers = PretrainedBertIndexer(pretrained_model=os.path.abspath(
os.path.join(path_to_bert_weights, 'vocab.txt')),
max_pieces=maximum_number_of_tokens,
do_lowercase=is_model_lowercase,
use_starting_offsets=True)
return token_indexers
def get_embedder_info(
embedder_type: str
) -> Tuple[TokenEmbedder, TokenIndexer, str, Dict[str, Any]]:
embedder_type = embedder_type.lower()
text_field_embedder_kwargs: Dict[str, Any] = {}
if embedder_type == 'ner_elmo':
return NERElmoTokenEmbedder(), ELMoTokenCharactersIndexer(
), text_field_embedder_kwargs
elif embedder_type == 'elmo':
return ElmoTokenEmbedder(ELMO_OPTIONS_FILE,
ELMO_WEIGHT_FILE), ELMoTokenCharactersIndexer(
), text_field_embedder_kwargs
elif embedder_type == 'bert':
bert_embedder = PretrainedBertEmbedder(
pretrained_model="bert-base-uncased",
top_layer_only=True, # conserve memory
)
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=512, # max pieces allowed for positional embeddings
do_lowercase=True,
use_starting_offsets=True,
)
text_field_embedder_kwargs['allow_unmatched_keys'] = True
text_field_embedder_kwargs['embedder_to_indexer_map'] = {
"tokens": ["tokens", "tokens-offsets"]
}
return bert_embedder, token_indexer, text_field_embedder_kwargs
else:
raise Exception(f'Unknown embedder type: {embedder_type}')
def get_token_utils(name: str = config.embedder):
if name == 'elmo':
from allennlp.data.tokenizers.word_splitter import SpacyWordSplitter
from allennlp.data.token_indexers.elmo_indexer import ELMoCharacterMapper, ELMoTokenCharactersIndexer
# the token indexer is responsible for mapping tokens to integers
token_indexer = ELMoTokenCharactersIndexer()
def tokenizer(x: str):
return [
w.text for w in SpacyWordSplitter(language='en_core_web_sm',
pos_tags=False).split_words(
x)[:config.max_seq_len]
]
return token_indexer, tokenizer
elif name == 'bert':
from allennlp.data.token_indexers import PretrainedBertIndexer
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=config.max_seq_len,
do_lowercase=True,
)
def tokenizer(s: str):
return token_indexer.wordpiece_tokenizer(s)[:config.max_seq_len -
2]
return token_indexer, tokenizer
def load_data(train_dir, test_dir):
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=config.max_seq_len,
do_lowercase=True,
)
reader = LoadData(tokenizer=tokenizer,
token_indexers={"tokens": token_indexer},
col_name=config.col_name)
train_data = pd.read_csv(train_dir)
test_data = pd.read_csv(test_dir)
test_y = test_data[config.col_name[1]].tolist()
train_data, val_data = train_test_split(train_data,
test_size=0.1,
random_state=42)
train_data = reader.read(train_data)
val_data = reader.read(val_data)
test_data = reader.read(test_data)
return train_data, val_data, test_data, test_y
def get_dataset_reader():
"""
Get dataset reader instance
"""
# Fixed configuration
config = Config(
use_percentage=1.0, # how much data to use
max_seq_len=512, # necessary to limit memory usage
use_extracted_emb=False, # set False since we are using BERT
batch_size=32 # for testing, this does not matter
)
# BERT wordpiece tokenizer
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-multilingual-cased", # recommended by Google
max_pieces=config.max_seq_len,
do_lowercase=False,
)
def tokenizer(s: str, max_seq_len: int=config.max_seq_len) -> List[str]:
return [Token(x) for x in token_indexer.wordpiece_tokenizer(s)[:max_seq_len]]
# dataset reader
reader = PairDatasetReaderSpecial(
tokenizer=tokenizer,
token_indexers={"tokens": token_indexer},
use_percentage=config.use_percentage,
use_extracted_emb=config.use_extracted_emb # set False if finetuning
)
return config, reader
def predict(vocab2):
bert_token_indexer = PretrainedBertIndexer(
pretrained_model="bert-large-uncased",
max_pieces=config.max_seq_len,
do_lowercase=True,
)
reader = BertAnalogyDatasetReader(
tokenizer=bert_tokenizer,
token_indexers={'tokens':bert_token_indexer}
)
train_dataset, test_dataset, dev_dataset = (reader.read(DATA_ROOT + "/" + fname) for fname in ["train_all.txt", "test_all.txt", "val_all.txt"])
bert_embedder = PretrainedBertEmbedder(
pretrained_model='bert-large-uncased',
top_layer_only=True, # conserve memory
)
word_embeddings: TextFieldEmbedder = BasicTextFieldEmbedder({"tokens": bert_embedder},
# we'll be ignoring masks so we'll need to set this to True
allow_unmatched_keys = True)
BERT_DIM = word_embeddings.get_output_dim()
class BertSentencePooler(Seq2VecEncoder):
def forward(self, embs: torch.tensor,
mask: torch.tensor=None) -> torch.tensor:
# extract first token tensor
return embs[:, 0]
@overrides
def get_output_dim(self) -> int:
return BERT_DIM
# if not vocab2:
# vocab2 = Vocabulary.from_files("./bert_vocabulary")
bert_encoder = BertSentencePooler(vocab2)
model2 = LstmModel(word_embeddings, bert_encoder, vocab2)
if USE_GPU: model2.cuda()
else: model2
with open("./bert_model.th", 'rb') as f:
model2.load_state_dict(torch.load(f))
predictor2 = SentenceClassifierPredictor(model2, dataset_reader=reader)
with open('bert_predictions.txt', 'w+') as f:
top_10_words_list = []
for analogy_test in test_dataset:
logits = predictor2.predict_instance(analogy_test)['logits']
label_id = np.argmax(logits)
label_predict = model2.vocab.get_token_from_index(label_id, 'labels')
top_10_ids = np.argsort(logits)[-10:]
top_10_words = [model2.vocab.get_token_from_index(id, 'labels') for id in top_10_ids]
top_10_words_list.append(top_10_words)
f.write(label_predict + "\n")
top_10_words_list = np.array(top_10_words_list)
print(top_10_words_list.shape)
np.save('bert_top_10_words_list.npy', np.array(top_10_words_list))
def __init__(self):
super().__init__(lazy=False)
self.token_indexers = {
"tokens":
PretrainedBertIndexer(pretrained_model="bert-base-cased",
do_lowercase=False
#max_pieces=config.max_seq_length
)
}
def __init__(self, word_indexer: Optional[TokenIndexer] = None):
super().__init__(lazy=False)
splitter = BertBasicWordSplitter()
self.tokeniser = WordTokenizer(word_splitter=splitter)
if word_indexer is None:
word_indexer = PretrainedBertIndexer(
pretrained_model='bert-base-uncased',
truncate_long_sequences=False)
self.word_indexers = {'tokens': word_indexer}
def get_indexer(embedding_type: str,
xlnet_vocab_file: Path
) -> TokenIndexer:
if embedding_type == 'bert':
return PretrainedBertIndexer(
pretrained_model='bert-base-uncased',
truncate_long_sequences=False)
if embedding_type == 'glove':
return SingleIdTokenIndexer(lowercase_tokens=True)
if embedding_type == 'xlnet':
return XLNetIndexer(vocab_file=str(xlnet_vocab_file))
def __init__(
self,
target_namespace: str,
lazy: bool = False,
) -> None:
super().__init__(lazy)
self._target_namespace = target_namespace
self._source_token_indexers: Dict[str, TokenIndexer] = {
"bert": PretrainedBertIndexer('bert-base-uncased')
}
self._target_token_indexers: Dict[str, TokenIndexer] = {
"tokens": SingleIdTokenIndexer(namespace=self._target_namespace)
}
def __init__(self,
tokenizer: Tokenizer = None,
token_indexers: Dict[str, TokenIndexer] = None,
lazy: bool = False) -> None:
super().__init__(lazy)
self._tokenizer = tokenizer or JiebaTokenizer(
word_splitter=JiebaSplitter())
self._token_indexers = token_indexers or {
'tokens':
PretrainedBertIndexer(
"./bert/multi_cased_L-12_H-768_A-12/bert-base-multilingual-cased-vocab.txt"
)
}
def load_stackex_data(data_root,
assign_id_file,
encoder,
train_type,
max_seq_len,
toy_data,
use_tokenizer=True,
is_rank_task=True):
# the token indexer is responsible for mapping tokens to integers
def bert_tokenizer(s: str):
return token_indexer.wordpiece_tokenizer(s)[:max_seq_len - 2]
def normal_tokenizer(x: str):
return [
w.text for w in SpacyWordSplitter(language='en_core_web_sm',
pos_tags=False).split_words(x)
[:max_seq_len]
]
if encoder == "bert":
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=max_seq_len,
do_lowercase=True,
)
tokenizer = bert_tokenizer
else:
token_indexer = SingleIdTokenIndexer()
tokenizer = normal_tokenizer
# init dataset reader
reader = StackExDataReader(assignee_id_file=assign_id_file,
tokenizer=tokenizer,
token_indexers={"tokens": token_indexer},
is_rank_task=is_rank_task,
toy_data=toy_data,
use_tokenizer=use_tokenizer)
if train_type == "seq":
train_pkl, val_pkl, test_pkl = "train.pkl", "val.pkl", "test.pkl"
else:
train_pkl, val_pkl, test_pkl = "train_" + train_type + ".pkl", "val_" + train_type + ".pkl", "test_" + train_type + ".pkl"
# elif train_type == "single":
# train_pkl, test_pkl = "train_single.pkl", "test_single.pkl"
train_ds, val_ds, test_ds = (reader.read(data_root / fname)
for fname in [train_pkl, val_pkl, test_pkl])
return reader, train_ds, val_ds, test_ds
def get_pretrained_bert_indexer(model_name: str = 'bert-base-uncased',
cache_dir: str = PATH_ALLENNLP_CACHE,
max_length=512,
lowercase=True,
**kwargs):
model_path = path.join(cache_dir, 'bert', f'{model_name}-vocab.txt')
msgex.assert_path_exist(
path_str=model_path,
arg_name='model_path',
extra_msg=f"the specified BERT model '{model_name}' is not found")
return PretrainedBertIndexer(pretrained_model=model_path,
max_pieces=max_length,
do_lowercase=lowercase,
**kwargs)
def get_token_indexer(self, token_indexers):
self.token_indexers = token_indexers or {
"tokens": SingleIdTokenIndexer()
}
# the token indexer is responsible for mapping tokens to integers
if self.embeddings == 'elmo':
self.token_indexers = {"tokens": ELMoTokenCharactersIndexer()}
elif self.embeddings == 'bert':
self.ber_embedder = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=128,
do_lowercase=True,
)
self.token_indexers = {"bert": self.ber_embedder}
def test_reader_from_file(self):
reader = BaselineReader(
idiom_vector_path=self.IDIOM_VECTOR_PATH,
pretrained_model=self.PRETRAINED_MODEL,
content_token_indexer={
"bert": PretrainedBertIndexer(
self.PRETRAINED_MODEL
)
},
max_seq_length=256
)
instances = reader.read(str(self.FIXTURES_ROOT / "data" / "realcount_3_sample.txt"))
instances = ensure_list(instances)
assert len(instances) == 3
def __init__(self, lazy: bool = False,
load_cache: bool = True,
save_cache: bool = True,
loading_ratio: float = 1.0,
enable_unparse: bool = True,
use_bert: bool = False,
# insert before/after/both
super_mode: str = 'before',
word_level: bool = False,
# if joint encoding, means that
joint_encoding: bool = False,
language: str = 'zh',
extra_stop_words: List = None):
super().__init__(lazy=lazy)
# different languages share the same word splitter
self._tokenizer = WordTokenizer(JustSpacesWordSplitter())
if use_bert:
pretrain_model_name = 'bert-base-chinese' if language == 'zh' else 'bert-base-uncased'
self._indexer = {'bert': PretrainedBertIndexer(pretrained_model=pretrain_model_name,
use_starting_offsets=False,
do_lowercase=True,
never_lowercase=['[UNK]', '[PAD]', '[CLS]', '[SEP]'],
truncate_long_sequences=False)}
else:
self._indexer = {'tokens': SingleIdTokenIndexer(namespace='tokens')}
# loading ratio is designed for hyper-parameter fine-tuning
self._loading_ratio = loading_ratio
self._load_cache = load_cache
self._save_cache = save_cache
self._use_bert = use_bert
self._language = language
self._word_level = word_level
self._super_mode = super_mode
self._enable_unparse = enable_unparse
self._joint_encoding = joint_encoding
self._extra_stop_words = extra_stop_words
def set_values(max_sequence_length: Optional[int] = -1,
concat_title_abstract: Optional[bool] = None,
data_source: Optional[str] = None,
included_text_fields: Optional[str] = None) -> None:
# set global values
# note: a class with __init__ would have been a better design
# we have this structure for efficiency reasons: to support multiprocessing
# as multiprocessing with class methods is slower
global _tokenizer
global _token_indexers
global _token_indexer_author_id
global _token_indexer_author_position
global _token_indexer_venue
global _token_indexer_id
global _max_sequence_length
global _concat_title_abstract
global _data_source
global _included_text_fields
if _tokenizer is None: # if not initialized, initialize the tokenizers and token indexers
_tokenizer = WordTokenizer(word_splitter=BertBasicWordSplitter(
do_lower_case=bert_params["do_lowercase"]))
_token_indexers = {
"bert": PretrainedBertIndexer.from_params(Params(bert_params))
}
_token_indexer_author_id = {
"tokens": SingleIdTokenIndexer(namespace='author')
}
_token_indexer_author_position = {
"tokens": SingleIdTokenIndexer(namespace='author_positions')
}
_token_indexer_venue = {
"tokens": SingleIdTokenIndexer(namespace='venue')
}
_token_indexer_id = {"tokens": SingleIdTokenIndexer(namespace='id')}
_max_sequence_length = max_sequence_length
_concat_title_abstract = concat_title_abstract
_data_source = data_source
_included_text_fields = included_text_fields
def __init__(self,
is_bert: bool,
conceptnet_path: Path,
word_indexer: Optional[TokenIndexer] = None):
super().__init__(lazy=False)
if is_bert:
splitter = BertBasicWordSplitter()
else:
splitter = SpacyWordSplitter()
self.tokeniser = WordTokenizer(word_splitter=splitter)
if word_indexer is None:
if is_bert:
word_indexer = PretrainedBertIndexer(
pretrained_model='bert-base-uncased',
truncate_long_sequences=True)
else:
word_indexer = SingleIdTokenIndexer(lowercase_tokens=True)
self.word_indexers = {'tokens': word_indexer}
# self.rel_indexers = {
# "rel_tokens": SingleIdTokenIndexer(namespace='rel_tokens')}
self.conceptnet = ConceptNet(conceptnet_path=conceptnet_path)
assert(len(sent_embeddings) == len(rel_labels))
return sentences, sent_embeddings, rel_labels
"""
Example
"""
if __name__ == "__main__":
working_dir = "../data/ICNALE/original/pairwise_link_labelling/test/"
# BERT wordpiece tokenizer
max_seq_len = 512 # maximum number of tokens when tokenizing a text
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-multilingual-cased", # recommended by Google
max_pieces=max_seq_len,
do_lowercase=False,
)
def tokenizer(s: str, max_seq_len: int=512) -> List[str]:
return [Token(x) for x in token_indexer.wordpiece_tokenizer(s)[:max_seq_len]]
# reading dataset
reader = PairDatasetReader(
tokenizer=None, # None if you do not want to tokenize
token_indexers={"tokens": token_indexer},
use_percentage=0.1,
use_extracted_emb=False # set False if finetuning
)
train_ds = reader.read(working_dir)
TextField(tk, self._token_indexers)
for tk in aug_unsup_tokenized_sents
])
fields['ori_unsup_sentences'] = ori_unsup_sentence_sequence
fields['aug_unsup_sentences'] = aug_unsup_sentence_sequence
# Fake data
# fields['ori_unsup_sentences'] = sentence_sequence
# fields['aug_unsup_sentences'] = sentence_sequence
return Instance(fields)
# %%
token_indexer = PretrainedBertIndexer(
pretrained_model="./biobert_v1.1_pubmed/vocab.txt",
do_lowercase=True,
)
# %%
reader = ClaimAnnotationReaderJSON(token_indexers={"tokens": token_indexer},
lazy=True)
merge_reader = MergeDatasetReader(token_indexers={"tokens": token_indexer},
lazy=True)
# train_dataset = reader.read(TRAIN_PATH)
train_dataset = merge_reader.read(COMBINED_TRAIN_PATH)
validation_dataset = reader.read(VALIDATION_PATH)
test_dataset = reader.read(TEST_PATH)
# %%
vocab = Vocabulary()
fileds["label"] = label_field
return Instance(fileds)
# def tokenizer(x: str):
# return [w.text for w in SpacyWordSplitter(language='en_core_web_sm', pos_tags=False).split_words(x)[: config.max_seq_len]]
# special tokenizer for bert
def tokenizer(s: str):
return token_indexer.wordpiece_tokenizer(s)[:config.max_seq_len - 2]
# read data
# token_indexer = SingleIdTokenIndexer()
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased",
max_pieces=config.max_seq_len,
do_lowercase=True,
)
reader = JigsawDatasetReader(tokenizer=tokenizer, token_indexers={"tokens": token_indexer})
DATA_ROOT = Path("data") / "jigsaw"
train_ds, test_ds = (reader.read(DATA_ROOT / fname) for fname in ["train.csv", "test_proced.csv"])
val_ds = None
# prepare vocab
# vocab = Vocabulary.from_instances(train_ds, max_vocab_size=config.max_vocab_size)
vocab = Vocabulary()
# prepare iterator
iterator = BucketIterator(batch_size=config.batch_size, sorting_keys=[("tokens", "num_tokens")])
iterator.index_with(vocab)
model: torch.nn.Module = SentencePairModel(
extractor=CoupledSentencePairFeatureExtractor(
joiner=BERTConcatenator(),
encoder=BertSeq2VecEncoderForPairs.from_pretrained(
"bert-base-uncased")),
mlp=torch.nn.Sequential(torch.nn.Linear(768, 1), torch.nn.Sigmoid()),
loss_func=torch.nn.BCELoss(),
mode="regression")
model.cuda()
if ARGS.pretrained != "":
model.load_state_dict(torch.load(ARGS.pretrained))
reader = QRelsPointwiseReader(
lazy=True,
token_indexers={"wordpiece": PretrainedBertIndexer("bert-base-uncased")},
left_tokenizer=BertTokenizer(),
right_tokenizer=BertTokenizer())
iterator = BasicIterator()
iterator.index_with(vocab)
trainer = Trainer(model=model,
optimizer=Adam(params=model.parameters(), lr=0.00001),
grad_norm=1.0,
train_dataset=reader.read(f"{ARGS.data}/train"),
validation_dataset=reader.read(f"{ARGS.data}/dev"),
iterator=iterator,
validation_metric="+pearson",
num_epochs=3,
patience=3,
serialization_dir=ARGS.out,
# In[6]:
from pytorch_pretrained_bert import BertConfig, BertForMaskedLM
masked_lm = BertForMaskedLM.from_pretrained(config.model_type)
masked_lm.eval()
# In[ ]:
# In[7]:
from allennlp.data import Token
from allennlp.data.token_indexers import PretrainedBertIndexer
token_indexer = PretrainedBertIndexer(
pretrained_model=config.model_type,
max_pieces=config.max_seq_len,
do_lowercase=True,
)
# if len(toks) < config.max_seq_len:
# return toks + (["[PAD]"] * (maxlen - len(toks)))
# else:
def tokenizer(s: str):
maxlen = config.max_seq_len - 2
toks = token_indexer.wordpiece_tokenizer(s)[:maxlen]
return toks
# In[8]:
def main():
parser = argparse.ArgumentParser(description='Evidence Inference experiments')
parser.add_argument('--cuda_device', type=int, default=0,
help='GPU number (default: 0)')
parser.add_argument('--epochs', type=int, default=2,
help='upper epoch limit (default: 2)')
parser.add_argument('--patience', type=int, default=1,
help='trainer patience (default: 1)')
parser.add_argument('--batch_size', type=int, default=32,
help='batch size (default: 32)')
parser.add_argument('--dropout', type=float, default=0.2,
help='dropout for the model (default: 0.2)')
parser.add_argument('--model_name', type=str, default='baseline',
help='model name (default: baseline)')
parser.add_argument('--tunable', action='store_true',
help='tune the underlying embedding model (default: False)')
args = parser.parse_args()
annotations = pd.read_csv('data/data/annotations_merged.csv')
prompts = pd.read_csv('data/data/prompts_merged.csv')
feature_dictionary = {}
prompts_dictionary = {}
for index, row in prompts.iterrows():
prompts_dictionary[row['PromptID']] = [row['Outcome'], row['Intervention'], row['Comparator']]
for index, row in annotations.iterrows():
if row['PMCID'] not in feature_dictionary:
feature_dictionary[row['PMCID']] = []
feature_dictionary[row['PMCID']].append([row['Annotations'], row['Label']]
+ prompts_dictionary[row['PromptID']])
train = []
valid = []
test = []
with open('data/splits/train_article_ids.txt') as train_file:
for line in train_file:
train.append(int(line.strip()))
with open('data/splits/validation_article_ids.txt') as valid_file:
for line in valid_file:
valid.append(int(line.strip()))
with open('data/splits/test_article_ids.txt') as test_file:
for line in test_file:
test.append(int(line.strip()))
bert_token_indexer = {'bert': PretrainedBertIndexer('scibert/vocab.txt', max_pieces=512)}
reader = EIDatasetReader(bert_token_indexer, feature_dictionary)
train_data = reader.read(train)
valid_data = reader.read(valid)
test_data = reader.read(test)
vocab = Vocabulary.from_instances(train_data + valid_data + test_data)
bert_token_embedding = PretrainedBertEmbedder(
'scibert/weights.tar.gz', requires_grad=args.tunable
)
word_embeddings = BasicTextFieldEmbedder(
{"bert": bert_token_embedding},
{"bert": ['bert']},
allow_unmatched_keys=True
)
model = Baseline(word_embeddings, vocab)
cuda_device = args.cuda_device
if torch.cuda.is_available():
model = model.cuda(cuda_device)
else:
cuda_device = -1
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
iterator = BucketIterator(batch_size=args.batch_size,
sorting_keys=[('intervention', 'num_tokens')],
padding_noise=0.1)
iterator.index_with(vocab)
serialization_dir = 'model_checkpoints/' + args.model_name
trainer = Trainer(model=model,
optimizer=optimizer,
iterator=iterator,
train_dataset=train_data,
validation_dataset=test_data,
patience=args.patience,
validation_metric='+accuracy',
num_epochs=args.epochs,
cuda_device=cuda_device,
serialization_dir=serialization_dir)
result = trainer.train()
for key in result:
print(str(key) + ': ' + str(result[key]))
test_metrics = evaluate(trainer.model, test_data, iterator,
cuda_device=cuda_device,
batch_weight_key="")
print('Test Data statistics:')
for key, value in test_metrics.items():
print(str(key) + ': ' + str(value))
HIDDEN_DIM = 256
PRINT_EVERY = 1000
EVALUATE_EVERY_EPOCH = 1
ENCODER_LAYER = 2
DROPOUT_RATE = 0.1
BATCH_SIZE = 32
INIT_LEARNING_RATE = 0.0
EPOCH = 1000
WARMUP_STEPS = 5000
PATIENCE = 10
BERT = False
torch.manual_seed(1)
if BERT:
token_indexer = PretrainedBertIndexer(
pretrained_model="bert-base-uncased", do_lowercase=True)
reader = DisfluencyDatasetReader(
token_indexers={"tokens": token_indexer})
else:
reader = DisfluencyDatasetReader()
train_dataset = reader.read('../train.txt')
validation_dataset = reader.read('../val.txt')
test_dataset = reader.read('../test.txt')
vocab = Vocabulary.from_instances(train_dataset + validation_dataset +
test_dataset)
if BERT:
bert_embedder = PretrainedBertEmbedder(
pretrained_model="bert-base-uncased",
def run_model(name, context, conf, double_input, use_elmo=False, save_predictions=False, save_model=False):
"""
Runs the given model 'name' for the given 'context' and agreement level 'conf'. If double_input is True, runs the combined model using context comment text. Optionally saves the trained model & its vocabulary, and predictions.
Allowed names: lstm | bilstm | stacked_bilstm | cnn | dense_lstm | dense_bilstm | dense_stacked_bilstm | dense_cnn | nli_cnn | bert | dense_bert
If use_elmo=True, uses ELMo's pre-trained language model for embeddings.
"""
if use_elmo:
token_indexer = ELMoTokenCharactersIndexer() # token indexer is responsible for mapping tokens to integers: this makes sure that the mapping is consistent with what was used in the original ELMo training.
elif name == 'bert':
global bert_token_indexer
bert_token_indexer = PretrainedBertIndexer(pretrained_model=BERT_MODEL, do_lowercase=True)
else:
token_indexer = SingleIdTokenIndexer()
if name == 'bert': # BERT uses a special wordpiece tokenizer
reader = data_reader.UnpalatableDatasetReader(main_input=context, additional_context=double_input,
tokenizer=tokenizer_bert, token_indexers={"tokens": bert_token_indexer},
label_cols=LABEL_COLS)
else:
reader = data_reader.UnpalatableDatasetReader(main_input=context, additional_context=double_input, tokenizer=tokenizer,
token_indexers={"tokens": token_indexer}, label_cols=LABEL_COLS)
map_reply_id_pred_probability = {}; n_epochs = []
f1s, AUROCs, weighted_f1s, precision_s, recall_s, accuracies, AUPRCs = [], [], [], [], [], [], []
for fold_number in range(1,6): # 5-fold cross validation
train_fname = 'train_data_fold_'+str(fold_number)+'_OneHot.csv'
val_fname = 'val_data_fold_'+str(fold_number)+'_OneHot.csv'
test_fname = 'test_data_fold_'+str(fold_number)+'_OneHot.csv'
train_dataset = reader.read(file_path=DATA_ROOT / conf / train_fname)
validation_dataset = reader.read(file_path=DATA_ROOT / conf / val_fname)
test_dataset = reader.read(file_path=DATA_ROOT / conf / test_fname)
print("\n#####################################################\n", double_input, context, conf, name, len(train_dataset), len(validation_dataset), len(test_dataset))
# Train model:
if name == 'lstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=False,
num_layers=1, bidirectional=False, use_elmo=use_elmo, double_input=double_input)
elif name == 'dense_lstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=True,
col_name=context, num_layers=1, bidirectional=False, use_elmo=use_elmo,
double_input=double_input)
elif name == 'bilstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=False,
num_layers=1, bidirectional=True, use_elmo=use_elmo, double_input=double_input)
elif name == 'dense_bilstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=True,
col_name=context, num_layers=1, bidirectional=True, use_elmo=use_elmo,
double_input=double_input)
elif name == 'stacked_bilstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=False,
num_layers=2, bidirectional=True, use_elmo=use_elmo, double_input=double_input)
elif name == 'dense_stacked_bilstm':
model, vocab, ep = train.train_lstm(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=True,
col_name=context, num_layers=2, bidirectional=True, use_elmo=use_elmo,
double_input=double_input)
elif name == 'cnn':
if context == 'reply_text': filter_sizes = (2,3) # kernels can not be bigger than the shortest sentence
else: filter_sizes = (2,)
model, vocab, ep = train.train_cnn(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=False,
num_filters=100, filter_sizes=filter_sizes, use_elmo=use_elmo,
double_input=double_input)
elif name == 'dense_cnn':
if context == 'reply_text': filter_sizes = (2,3) # kernels can not be bigger than the shortest sentence
else: filter_sizes = (2,)
model, vocab, ep = train.train_cnn(train_dataset, validation_dataset, BATCH_SIZE, dense_vector=True,
col_name=context, num_filters=100, filter_sizes=filter_sizes, use_elmo=use_elmo,
double_input=double_input)
elif name == 'nli_cnn':
if double_input == False:
print("Error: NLI-inspired architecture only accepts double-input.")
return [None]*9
filter_sizes = (2,3)
model, vocab, ep = train.train_nli(train_dataset, validation_dataset, BATCH_SIZE, use_elmo=use_elmo,
num_filters=100, filter_sizes=filter_sizes)
elif name == 'bert':
model, vocab, ep = train.train_bert(train_dataset, validation_dataset, BATCH_SIZE, pretrained_model=BERT_MODEL,
dense_vector=False, double_input=double_input)
elif name == 'dense_bert':
model, vocab, ep = train.train_bert(train_dataset, validation_dataset, BATCH_SIZE, pretrained_model=BERT_MODEL,
dense_vector=True, col_name=context, double_input=double_input)
else:
sys.exit("'name' not valid")
n_epochs.append(ep) # keep track of number of actual training epochs for each fold
# Predict and evaluate model on test set:
preds = evaluate.make_predictions(model, vocab, test_dataset, BATCH_SIZE, use_gpu=False) # NOTE: preds is of shape (number of samples, 2) - the columns represent the probabilities for the two classes in order ['yes_unp', 'not_unp']
f1, auroc, w_f1, precision, recall, acc, auprc = evaluate.compute_metrics(preds, test_dataset)
if save_predictions: # save predictions for error analysis
replyid_pred = evaluate.map_id_prediction(preds, test_dataset)
if set(replyid_pred.keys()).intersection(set(map_reply_id_pred_probability.keys())) != set(): # sanity check
sys.exit("Error: There is overlap in Test IDs across folds.")
map_reply_id_pred_probability.update(replyid_pred)
if save_model: # save the model weights and vocabulary
with open('./tmp/'+name+'_model_conf_'+conf.split('-')[1]+'_fold_'+str(fold_number)+'.th', 'wb') as f:
torch.save(model.state_dict(), f)
vocab.save_to_files("./tmp/"+name+"_vocabulary_"+conf.split('-')[1]+"_fold_"+str(fold_number))
print("\nFold #{} | F1 = {} | AUROC = {} | AUPRC = {}".format(fold_number, f1, auroc, auprc))
f1s.append(f1); AUROCs.append(auroc); weighted_f1s.append(w_f1); precision_s.append(precision);
recall_s.append(recall); accuracies.append(acc); AUPRCs.append(auprc)
mean_f1 = np.array(f1s).mean(); mean_auroc = np.array(AUROCs).mean(); mean_weighted_f1 = np.array(weighted_f1s).mean();
mean_precision = np.array(precision_s).mean(); mean_recall = np.array(recall_s).mean(); mean_accuracy = np.array(accuracies).mean(); mean_auprc = np.array(AUPRCs).mean()
print("Total predictions: {} | Save Predictions: {}".format(len(map_reply_id_pred_probability), save_predictions))
return mean_f1, mean_auroc, mean_weighted_f1, mean_precision, mean_recall, mean_accuracy, mean_auprc, map_reply_id_pred_probability, n_epochs
def multiprocess_single_sequence_loader(process_number: int, _config,
_queue: mp.Queue,
_wait_for_exit: mp.Event, _local_file,
_fasttext_vocab_cached_mapping,
_fasttext_vocab_cached_data):
torch.manual_seed(_config["random_seed"])
numpy.random.seed(_config["random_seed"])
random.seed(_config["random_seed"])
if _config["token_embedder_type"] == "bert_cls":
_tokenizer = BlingFireTokenizer()
_ind = PretrainedBertIndexer(
pretrained_model=_config["bert_pretrained_model"],
do_lowercase=True)
_token_indexers = {"tokens": _ind}
_tuple_loader = IrSingleSequenceDatasetReader(
lazy=True,
tokenizer=_tokenizer,
token_indexers=_token_indexers,
max_seq_length=_config["max_doc_length"],
min_seq_length=_config["min_doc_length"],
)
_iterator = BucketIterator(batch_size=int(_config["batch_size_eval"]),
sorting_keys=[("seq_tokens", "num_tokens")])
_iterator.index_with(Vocabulary.from_files(_config["vocab_directory"]))
else:
_tokenizer = BlingFireTokenizer()
if _config["token_embedder_type"] == "embedding":
_token_indexers = {
"tokens": SingleIdTokenIndexer(lowercase_tokens=True)
}
_vocab = Vocabulary.from_files(_config["vocab_directory"])
elif _config["token_embedder_type"] == "fasttext":
_token_indexers = {
"tokens":
FastTextNGramIndexer(_config["fasttext_max_subwords"])
}
_vocab = FastTextVocab(_fasttext_vocab_cached_mapping,
_fasttext_vocab_cached_data,
_config["fasttext_max_subwords"])
elif _config["token_embedder_type"] == "elmo":
_token_indexers = {"tokens": ELMoTokenCharactersIndexer()}
_vocab = None
_tuple_loader = IrSingleSequenceDatasetReader(
lazy=True,
tokenizer=_tokenizer,
token_indexers=_token_indexers,
max_seq_length=_config["max_doc_length"],
min_seq_length=_config["min_doc_length"],
)
_iterator = BucketIterator(batch_size=int(_config["batch_size_eval"]),
sorting_keys=[("seq_tokens", "num_tokens")])
_iterator.index_with(_vocab)
for training_batch in _iterator(_tuple_loader.read(_local_file),
num_epochs=1):
_queue.put(
training_batch) # this moves the tensors in to shared memory
_queue.put(None) # signal end of queue
_queue.close() # indicate this local thread is done
_wait_for_exit.wait(
) # keep this process alive until all the shared memory is used and not needed anymore