# unfreeze top layers and train
# for param in list(model.parameters())[:-4]:
# param.requires_grad = False
trainer = Trainer(model=model,
optimizer=optimizer,
iterator=iterator,
validation_iterator=iterator,
train_dataset=train_dataset,
validation_dataset=validation_dataset,
patience=3,
num_epochs=1,
cuda_device=-1)
trainer.train()
# precision, recall, f-score on validation set
validation_list = read_json(cached_path(VALIDATION_PATH))
claim_predictor = ClaimCrfPredictor(model, dataset_reader=reader)
y_pred, y_true = [], []
for val in validation_list:
pred = claim_predictor.predict_json(val)
logits = torch.FloatTensor(pred['logits'])
best_paths = model.crf.viterbi_tags(
torch.FloatTensor(pred['logits']).unsqueeze(0),
torch.LongTensor(pred['mask']).unsqueeze(0))
predicted_labels = best_paths[0][0]
y_pred.extend(predicted_labels)
y_true.extend(val['labels'])
y_true = np.array(y_true).astype(int)
y_pred = np.array(y_pred).astype(int)
print(precision_recall_fscore_support(y_true, y_pred, average='binary'))
num_classes, constraints, include_start_end_transitions = 2, None, False
model.classifier_feedforward._linear_layers = ModuleList([
torch.nn.Linear(2 * EMBEDDING_DIM, EMBEDDING_DIM),
torch.nn.Linear(EMBEDDING_DIM, num_classes)
])
model.crf = ConditionalRandomField(
num_classes,
constraints,
include_start_end_transitions=include_start_end_transitions)
model.label_projection_layer = TimeDistributed(
Linear(2 * EMBEDDING_DIM, num_classes))
model.load_state_dict(torch.load(cached_path(WEIGHT_PATH)))
reader = CrfPubmedRCTReader()
claim_predictor = ClaimCrfPredictor(model, dataset_reader=reader)
fixture_path = os.path.join('..', 'pubmed-rct', 'PubMed_200k_RCT',
'fixtures_crf.json')
examples = read_json(fixture_path)
pred_list = []
for example in examples:
sentences = sent_tokenize(example['abstract'])
instance = reader.text_to_instance(sents=sentences)
pred = claim_predictor.predict_instance(instance)
logits = torch.FloatTensor(pred['logits'])
best_paths = model.crf.viterbi_tags(
torch.FloatTensor(pred['logits']).unsqueeze(0),
torch.LongTensor(pred['mask']).unsqueeze(0))
pred_list.append(best_paths[0][0])
print(pred_list)
"""
Example of discourse prediction using ``discourse`` library
created using AllenNLP
"""
import os
import sys
import json
sys.path.insert(0, '..')
from discourse.predictors import DiscourseClassifierPredictor
from discourse.dataset_readers import PubmedRCTReader
from discourse.models import DiscourseClassifier
from discourse import read_json
from allennlp.models.archival import load_archive
from allennlp.common.file_utils import cached_path
from allennlp.service.predictors import Predictor
MODEL_PATH = 'https://s3-us-west-2.amazonaws.com/pubmed-rct/model.tar.gz'
archive = load_archive(cached_path(MODEL_PATH))
predictor = Predictor.from_archive(archive, 'discourse_predictor')
if __name__ == '__main__':
fixture_path = os.path.join('..', 'pubmed-rct', 'PubMed_200k_RCT',
'fixtures.json')
json_sentences = read_json(fixture_path)
for json_sentence in json_sentences:
output = predictor.predict_json(json_sentence)
output['sentence'] = json_sentence['sentence']
print(output)
def flatten_dataset(df):
"""
Flatten Gold Standard JSON data for Claim Extraction
"""
sentence_data = []
for _, r in df.iterrows():
sentence_data.extend(list(zip(r.sentences, r.labels)))
flatten_df = pd.DataFrame(sentence_data, columns=['sentence', 'label'])
return flatten_df
if __name__ == '__main__':
# read dataset
train_df = pd.DataFrame(read_json(cached_path(TRAIN_PATH)))
validation_df = pd.DataFrame(read_json(cached_path(VALIDATION_PATH)))
test_df = pd.DataFrame(read_json(cached_path(TEST_PATH)))
train_df = flatten_dataset(train_df)
validation_df = flatten_dataset(validation_df)
test_df = flatten_dataset(test_df)
# prepare dataset
train_df['class_probability'] = train_df.sentence.map(
lambda x: predictor.predict_json({'sentence': x})['class_probabilities'
])
train_df['sentence_vector'] = train_df.sentence.map(
lambda x: get_sentence_vector(x, ft_model))
X_train = np.hstack((np.vstack(train_df['sentence_vector'].values),
np.vstack(train_df['class_probability'].values)))