コード例 #1
0
ファイル: main.py プロジェクト: wonlee2019/EGAD
def start_exp(args):
    trainer = Trainer(dataset_dict[args.dataset], args.cuda)
    results = trainer.train_model(args)
    if not path.exists('results'):
        os.mkdir('results')
    f = open(
        "results/" + args.dataset + '_teacher_emb_' +
        str(args.teacher_embed_size) + '_teacher_heads_' +
        str(args.teacher_n_heads) + '_window_' + str(args.window) +
        '_student_emb_' + str(args.student_emb) + '_student_heads_' +
        str(args.student_heads) + '_distillation_' + str(args.distillation) +
        '.pkl', "wb")
    pkl.dump(results, f)
    f.close()
コード例 #2
0
        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)
コード例 #3
0
                                          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)