def pred(model, best_model_path, result_conll_path, fname="test.csv"):
learner = NerLearner(model,
data,
best_model_path=model_dir +
"/norne/bilstm_attn_cased_en.cpt",
lr=0.01,
clip=1.0,
sup_labels=[
l for l in data.id2label
if l not in ['<pad>', '[CLS]', 'X', 'B_O', 'I_']
])
dl = get_bert_data_loader_for_predict(data_path + fname, learner)
learner.load_model(best_model_path)
preds = learner.predict(dl)
tokens, y_true, y_pred, set_labels = bert_preds_to_ys(dl, preds)
clf_report = flat_classification_report(y_true,
y_pred,
set_labels,
digits=3)
# clf_report = get_bert_span_report(dl, preds)
print(clf_report)
write_true_and_pred_to_conll(tokens=tokens,
y_true=y_true,
y_pred=y_pred,
conll_fpath=result_conll_path)
def train(model, num_epochs=20):
learner = NerLearner(model,
data,
best_model_path=model_dir +
"/norne/bilstm_attn_lr0_1_cased_en.cpt",
lr=0.1,
clip=1.0,
sup_labels=[
l for l in data.id2label
if l not in ['<pad>', '[CLS]', 'X', 'B_O', 'I_']
],
t_total=num_epochs * len(data.train_dl))
learner.fit(num_epochs, target_metric='f1')
dl = get_bert_data_loader_for_predict(data_path + "valid.csv", learner)
learner.load_model()
preds = learner.predict(dl)
print(
validate_step(learner.data.valid_dl, learner.model,
learner.data.id2label, learner.sup_labels))
clf_report = get_bert_span_report(dl, preds, [])
print(clf_report)
torch.cuda.set_device(5)
data = NerData.create(train_path,
valid_path,
vocab_file,
is_cls=True,
batch_size=16)
print(data.label2idx)
# model = BertBiLSTMAttnNCRFJoint.create(
# len(data.label2idx), len(data.cls2idx), bert_config_file, init_checkpoint_pt,
# enc_hidden_dim=1024, rnn_layers=1, num_heads=5, input_dropout=0.5, nbest=12)
model = BertBiLSTMAttnNMTJoint.create(len(data.label2idx),
len(data.cls2idx),
bert_config_file,
init_checkpoint_pt,
enc_hidden_dim=512,
rnn_layers=1,
dec_embedding_dim=128,
dec_hidden_dim=512,
input_dropout=0.5,
nbest=8)
# model = torch.nn.DataParallel(model, [2, 3])
num_epochs = 150
learner = NerLearner(
model,
data,
best_model_path="/home/aaemeljanov/models/AGRR-2019/slavic-nmt.cpt",
lr=0.0001,
clip=1.0,
sup_labels=data.id2label[1:],
t_total=num_epochs * len(data.train_dl))
learner.fit(num_epochs, target_metric='f1')
data_path = "/media/liah/DATA/ner_data_other/norne/"
train_path = data_path + "train.txt"
dev_path = data_path + "valid.txt"
test_path = data_path + "test.txt"
dl = get_bert_data_loader_for_predict(data_path + "valid.csv", learner)
model = BertBiLSTMAttnNMT.create(len(data.label2idx),
bert_config_file,
init_checkpoint_pt,
enc_hidden_dim=128,
dec_hidden_dim=128,
dec_embedding_dim=16)
learner = NerLearner(model,
data,
best_model_path=model_dir +
"conll-2003/bilstm_attn_cased.cpt",
lr=0.01,
clip=1.0,
sup_labels=[
l for l in data.id2label
if l not in ['<pad>', '[CLS]', 'X', 'B_O', 'I_']
],
t_total=num_epochs * len(data.train_dl))
learner.load_model(best_model_path)
preds = learner.predict(dl)
data = NerData.create(train_path,
valid_path,
vocab_file,
is_cls=False,
batch_size=16)
print(data.label2idx)
print("data.tokenizer.vocab:", len(data.tokenizer.vocab))
model = BertBiLSTMAttnNCRF.create(len(data.label2idx),
bert_config_file,
init_checkpoint_pt,
enc_hidden_dim=1024,
rnn_layers=1,
num_heads=6,
input_dropout=0.5,
nbest=11)
print(model)
# model = BertBiLSTMAttnNMTJoint.create(
# len(data.label2idx), len(data.cls2idx), bert_config_file, init_checkpoint_pt,
# enc_hidden_dim=128, rnn_layers=1, dec_embedding_dim=32, dec_hidden_dim=128, input_dropout=0.5, nbest=11)
# model = torch.nn.DataParallel(model, [2, 3])
num_epochs = 150
learner = NerLearner(
model,
data,
best_model_path=
"/home/eartemov/ae/work/models/AGRR-2019/slavic_without_clf.cpt",
lr=0.0001,
clip=1.0,
sup_labels=data.id2label[1:],
t_total=num_epochs * len(data.train_dl))
learner.fit(num_epochs, target_metric='f1')