def train(path):
# Trainer
trainer = Trainer(**train_args.data)
def _define_optim(obj):
lr = optim_args.data['lr']
embed_params = set(obj._model.word_embedding.parameters())
decay_params = set(obj._model.arc_predictor.parameters()) | set(
obj._model.label_predictor.parameters())
params = [
p for p in obj._model.parameters()
if p not in decay_params and p not in embed_params
]
obj._optimizer = torch.optim.Adam([{
'params': list(embed_params),
'lr': lr * 0.1
}, {
'params': list(decay_params),
**optim_args.data
}, {
'params': params
}],
lr=lr,
betas=(0.9, 0.9))
obj._scheduler = torch.optim.lr_scheduler.LambdaLR(
obj._optimizer, lambda ep: max(.75**(ep / 5e4), 0.05))
def _update(obj):
# torch.nn.utils.clip_grad_norm_(obj._model.parameters(), 5.0)
obj._scheduler.step()
obj._optimizer.step()
trainer.define_optimizer = lambda: _define_optim(trainer)
trainer.update = lambda: _update(trainer)
trainer.set_validator(
Tester(**test_args.data, evaluator=ParserEvaluator(ignore_label)))
model.word_embedding = torch.nn.Embedding.from_pretrained(embed,
freeze=False)
model.word_embedding.padding_idx = word_v.padding_idx
model.word_embedding.weight.data[word_v.padding_idx].fill_(0)
model.pos_embedding.padding_idx = pos_v.padding_idx
model.pos_embedding.weight.data[pos_v.padding_idx].fill_(0)
# try:
# ModelLoader.load_pytorch(model, "./save/saved_model.pkl")
# print('model parameter loaded!')
# except Exception as _:
# print("No saved model. Continue.")
# pass
# Start training
trainer.train(model, train_data, dev_data)
print("Training finished!")
# Saver
saver = ModelSaver("./save/saved_model.pkl")
saver.save_pytorch(model)
print("Model saved!")
def train():
# Trainer
trainer = Trainer(**train_args.data)
def _define_optim(obj):
obj._optimizer = torch.optim.Adam(obj._model.parameters(),
**optim_args.data)
obj._scheduler = torch.optim.lr_scheduler.LambdaLR(
obj._optimizer, lambda ep: .75**(ep / 5e4))
def _update(obj):
obj._scheduler.step()
obj._optimizer.step()
trainer.define_optimizer = lambda: _define_optim(trainer)
trainer.update = lambda: _update(trainer)
trainer.get_loss = lambda predict, truth: trainer._loss_func(
**predict, **truth)
trainer._create_validator = lambda x: MyTester(**test_args.data)
# Model
model = BiaffineParser(**model_args.data)
# use pretrain embedding
embed, _ = EmbedLoader.load_embedding(
model_args['word_emb_dim'], emb_file_name, 'glove', word_v,
os.path.join(processed_datadir, 'word_emb.pkl'))
model.word_embedding = torch.nn.Embedding.from_pretrained(embed,
freeze=False)
model.word_embedding.padding_idx = word_v.padding_idx
model.word_embedding.weight.data[word_v.padding_idx].fill_(0)
model.pos_embedding.padding_idx = pos_v.padding_idx
model.pos_embedding.weight.data[pos_v.padding_idx].fill_(0)
try:
ModelLoader.load_pytorch(model, "./save/saved_model.pkl")
print('model parameter loaded!')
except Exception as _:
print("No saved model. Continue.")
pass
# Start training
trainer.train(model, train_data, dev_data)
print("Training finished!")
# Saver
saver = ModelSaver("./save/saved_model.pkl")
saver.save_pytorch(model)
print("Model saved!")