def create(cls,
train_path,
valid_path,
vocab_file,
batch_size=16,
cuda=True,
is_cls=False,
data_type="bert_cased",
max_seq_len=424,
is_meta=False):
if ipython_info():
global tqdm_notebook
tqdm_notebook = tqdm
if data_type == "bert_cased":
do_lower_case = False
fn = get_bert_data_loaders
elif data_type == "bert_uncased":
do_lower_case = True
fn = get_bert_data_loaders
else:
raise NotImplementedError("No requested mode :(.")
return cls(*fn(train_path, valid_path, vocab_file, batch_size, cuda,
is_cls, do_lower_case, max_seq_len, is_meta),
batch_size=batch_size,
cuda=cuda,
is_meta=is_meta)
def create(cls,
train_path,
valid_path,
model_dir,
config_name,
batch_size=16,
cuda=True,
is_cls=False,
oov='<oov>',
pad='<pad>'):
if ipython_info():
global tqdm_notebook
tqdm_notebook = tqdm
fn = get_elmo_data_loaders
return cls(*fn(train_path, valid_path, model_dir, config_name,
batch_size, cuda, is_cls, oov, pad),
batch_size=batch_size,
cuda=cuda)
def __init__(self,
model,
data,
best_model_path,
lr=0.001,
betas=list([0.8, 0.9]),
clip=5,
verbose=True,
sup_labels=None,
t_total=-1,
warmup=0.1,
weight_decay=0.01):
if ipython_info():
global tqdm_notebook
tqdm_notebook = tqdm
self.model = model
self.optimizer = BertAdam(model,
lr,
t_total=t_total,
b1=betas[0],
b2=betas[1],
max_grad_norm=clip)
self.optimizer_defaults = dict(model=model,
lr=lr,
warmup=warmup,
t_total=t_total,
schedule='warmup_linear',
b1=betas[0],
b2=betas[1],
e=1e-6,
weight_decay=0.01,
max_grad_norm=clip)
self.data = data
if sup_labels is None:
sup_labels = data.id2label[1:]
self.sup_labels = sup_labels
self.best_model_path = best_model_path
self.verbose = verbose
self.history = []
self.cls_history = []
self.epoch = 0
self.clip = clip
self.best_target_metric = 0.
self.lr_scheduler = None
