def main():
from model import E2EETModel, MentionLevelModel
from bert_serving.client import BertClient
import jsonlines
bc = BertClient()
logger.info("Loading files...")
data_loaders = dutils.load_obj_from_pkl_file('data loaders', cf.ASSET_FOLDER + '/data_loaders.pkl')
word_vocab = dutils.load_obj_from_pkl_file('word vocab', cf.ASSET_FOLDER + '/word_vocab.pkl')
wordpiece_vocab = dutils.load_obj_from_pkl_file('wordpiece vocab', cf.ASSET_FOLDER + '/wordpiece_vocab.pkl')
hierarchy = dutils.load_obj_from_pkl_file('hierarchy', cf.ASSET_FOLDER + '/hierarchy.pkl')
total_wordpieces = dutils.load_obj_from_pkl_file('total wordpieces', cf.ASSET_FOLDER + '/total_wordpieces.pkl')
logger.info("Building model.")
model = create_model(data_loaders, word_vocab, wordpiece_vocab, hierarchy, total_wordpieces)
model.cuda()
model.load_state_dict(torch.load(cf.BEST_MODEL_FILENAME))
modelEvaluator = ModelEvaluator(model, data_loaders['test'], word_vocab, wordpiece_vocab, hierarchy, bc, mode="test")
with jsonlines.open(cf.BEST_MODEL_JSON_FILENAME, "r") as reader:
for line in reader:
f1_score, epoch = line['f1_score'], line['epoch']
modelEvaluator.evaluate_model(epoch)
def main(opts):
if len(opts) == 0:
raise ValueError("Usage: evaluate.py <dataset>")
dataset = opts[0]
if dataset not in ['cateringServices', 'automotiveEngineering', 'bbn']:
raise ValueError(
"Dataset must be either cateringServices, automotiveEngineering or bbn."
)
cf.load_config(dataset)
logger.info("Loading data loaders...")
data_loader_train = dutils.load_obj_from_pkl_file(
'data loader (train)', cf.ASSET_FOLDER + '/data_loader_train.pkl')
data_loader_dev = dutils.load_obj_from_pkl_file(
'data loader (dev)', cf.ASSET_FOLDER + '/data_loader_dev.pkl')
dataset_dev = pd.read_csv(cf.DEV_FILENAME)
ground_truth_triples_df = pd.read_csv(cf.GROUND_TRUTH_TRIPLES_FILE)
ground_truth_triples = parse_ground_truth_triples(ground_truth_triples_df)
logger.info("Building model.")
model = CandidateFilteringModel(
embedding_dim=cf.EMBEDDING_DIM,
hidden_dim=cf.HIDDEN_DIM,
)
model.cuda()
train(model, data_loader_train, data_loader_dev, dataset_dev,
ground_truth_triples)
def __init__(self, dataset, cf):
from model import E2EETModel
from bert_serving.client import BertClient
import jsonlines
logger.info("Loading files...")
data_loaders = dutils.load_obj_from_pkl_file(
'data loaders', cf.ASSET_FOLDER + '/data_loaders.pkl')
# Note: the word and wordpiece vocab are stored as attributes so that they may be expanded
# if necessary during evaluation (if a new word appears)
self.word_vocab = dutils.load_obj_from_pkl_file(
'word vocab', cf.ASSET_FOLDER + '/word_vocab.pkl')
self.wordpiece_vocab = dutils.load_obj_from_pkl_file(
'wordpiece vocab', cf.ASSET_FOLDER + '/wordpiece_vocab.pkl')
hierarchy = dutils.load_obj_from_pkl_file(
'hierarchy', cf.ASSET_FOLDER + '/hierarchy.pkl')
total_wordpieces = dutils.load_obj_from_pkl_file(
'total wordpieces', cf.ASSET_FOLDER + '/total_wordpieces.pkl')
# Initialise the coref pipeline for use in end-user evaluation
self.nlp = spacy.load('en')
self.coref = neuralcoref.NeuralCoref(self.nlp.vocab)
self.nlp.add_pipe(self.coref, name='neuralcoref')
logger.info("Building model.")
model = E2EETModel(
embedding_dim=cf.EMBEDDING_DIM + cf.POSITIONAL_EMB_DIM,
hidden_dim=cf.HIDDEN_DIM,
vocab_size=len(self.wordpiece_vocab),
label_size=len(hierarchy),
total_wordpieces=total_wordpieces,
category_counts=hierarchy.get_train_category_counts(),
hierarchy_matrix=hierarchy.hierarchy_matrix,
max_seq_len=cf.MAX_SENT_LEN,
batch_size=1)
model.cuda()
model.load_state_dict(torch.load(cf.BEST_MODEL_FILENAME))
self.modelEvaluator = ModelEvaluator(model, None, self.word_vocab,
self.wordpiece_vocab, hierarchy,
None, cf)
self.cf = cf
def main():
logger.info("Loading files...")
data_loaders = dutils.load_obj_from_pkl_file('data loaders', cf.ASSET_FOLDER + '/data_loaders.pkl')
word_vocab = dutils.load_obj_from_pkl_file('word vocab', cf.ASSET_FOLDER + '/word_vocab.pkl')
wordpiece_vocab = dutils.load_obj_from_pkl_file('wordpiece vocab', cf.ASSET_FOLDER + '/wordpiece_vocab.pkl')
hierarchy = dutils.load_obj_from_pkl_file('hierarchy', cf.ASSET_FOLDER + '/hierarchy.pkl')
total_wordpieces = dutils.load_obj_from_pkl_file('total wordpieces', cf.ASSET_FOLDER + '/total_wordpieces.pkl')
print(len(data_loaders['train']))
logger.info("Building model.")
model = create_model(data_loaders, word_vocab, wordpiece_vocab, hierarchy, total_wordpieces)
model.cuda()
train(model, data_loaders, word_vocab, wordpiece_vocab, hierarchy)
def main(opts):
if len(opts) == 0:
raise ValueError("Usage: train.py <dataset>")
dataset = opts[0]
if dataset not in ['cateringServices', 'automotiveEngineering', 'bbn']:
raise ValueError(
"Dataset must be either cateringServices, automotiveEngineering, or bbn."
)
cf.load_config(dataset)
logger.info("Loading files...")
data_loaders = dutils.load_obj_from_pkl_file(
'data loaders', cf.ASSET_FOLDER + '/data_loaders.pkl')
word_vocab = dutils.load_obj_from_pkl_file(
'word vocab', cf.ASSET_FOLDER + '/word_vocab.pkl')
wordpiece_vocab = dutils.load_obj_from_pkl_file(
'wordpiece vocab', cf.ASSET_FOLDER + '/wordpiece_vocab.pkl')
hierarchy = dutils.load_obj_from_pkl_file(
'hierarchy', cf.ASSET_FOLDER + '/hierarchy.pkl')
total_wordpieces = dutils.load_obj_from_pkl_file(
'total wordpieces', cf.ASSET_FOLDER + '/total_wordpieces.pkl')
ground_truth_triples_df = pd.read_csv(cf.GROUND_TRUTH_TRIPLES_FILE)
ground_truth_triples = parse_ground_truth_triples(ground_truth_triples_df)
logger.info("Building model.")
model = E2EETModel(embedding_dim=cf.EMBEDDING_DIM + cf.POSITIONAL_EMB_DIM,
hidden_dim=cf.HIDDEN_DIM,
vocab_size=len(wordpiece_vocab),
label_size=len(hierarchy),
total_wordpieces=total_wordpieces,
category_counts=hierarchy.get_train_category_counts(),
hierarchy_matrix=hierarchy.hierarchy_matrix,
max_seq_len=cf.MAX_SENT_LEN,
batch_size=cf.BATCH_SIZE)
model.cuda()
train(model, data_loaders, word_vocab, wordpiece_vocab, hierarchy,
ground_truth_triples)
