all_gram_idx = torch.arange(
self.n_gram).cuda() if use_cuda else torch.arange(self.n_gram)
all_vocab_idx = torch.arange(
self.vocab_size).cuda() if use_cuda else torch.arange(
self.vocab_size)
position_matrix = self.C(all_gram_idx).reshape(-1, self.embedding_dim,
self.embedding_dim)
context_features = torch.tensordot(context_word_features,
position_matrix)
all_word = self.embeddings_word_output(all_vocab_idx)
decoded = torch.mm(
context_features,
all_word.T) + self.embedding_bias(all_vocab_idx).view(-1)
logits = F.log_softmax(decoded, dim=1)
return logits
TEXT, train_iter, val_iter, test_iter = get_iter(args.batch_size)
model = LBLModel(TEXT=TEXT,
embedding_dim=args.embedding_dim,
batch_size=args.batch_size,
n_gram=args.n_gram)
if use_cuda:
model.cuda()
trainer = Trainer(train_iter=train_iter,
val_iter=val_iter,
TEXT=TEXT,
lr=args.lr,
n_gram=args.n_gram)
trainer.train_model(model=model, num_epochs=args.num_epochs)
position_matrix,
dim=1)
if args.sep_output:
all_word = self.embeddings_word_output(all_vocab_idx)
else:
all_word = self.embeddings_word(all_vocab_idx)
decoded = torch.mm(
context_features,
all_word.T) + self.embedding_bias(all_vocab_idx).view(-1)
logits = F.log_softmax(decoded, dim=1)
return logits
TEXT, train_iter, val_iter, test_iter = get_iter(args.batch_size, args.dataset)
model = LBLModel(TEXT=TEXT,
embedding_dim=args.embedding_dim,
batch_size=args.batch_size,
n_gram=args.n_gram)
if use_cuda:
model.cuda()
trainer = Trainer(train_iter=train_iter,
val_iter=val_iter,
TEXT=TEXT,
lr=args.lr,
n_gram=args.n_gram)
trainer.train_model(model=model,
num_epochs=args.num_epochs,
path=wandb.run.dir)