def __init__(self, scope, num_dirs, opt="adagrad", lr=0.025):
print("num_dirs:", num_dirs)
sub_dir = "ecb_" + scope + "/"
# init stuff
print("TORCH VERSION:", torch.__version__)
global args
self.args = config.parse_known_args()
self.args.cuda = self.args.cuda and torch.cuda.is_available()
device = torch.device("cuda:0" if self.args.cuda else "cpu")
torch.manual_seed(self.args.seed)
random.seed(self.args.seed)
print("TREELSTM:", opt, "lr:", lr)
# paths
train_dir = os.path.join(self.args.data, str(num_dirs), 'train/',
sub_dir)
dev_dir = os.path.join(self.args.data, str(num_dirs), 'dev/', sub_dir)
test_dir = os.path.join(self.args.data, str(num_dirs), 'test/',
sub_dir)
print("train_dir:", train_dir)
print("dev_dir:", dev_dir)
# builds vocabulary
sick_vocab_file = Helper.build_entire_vocab(
os.path.join(self.args.data, str(num_dirs), 'sick.vocab'),
train_dir, dev_dir, test_dir)
vocab = Vocab(filename=sick_vocab_file,
data=[
Constants.PAD_WORD, Constants.UNK_WORD,
Constants.BOS_WORD, Constants.EOS_WORD
])
print('==> SICK vocabulary size : %d ' % vocab.size())
# loads SICKDataset: Trees, sentences, and labels
self.train_dataset = Helper.load_data(
train_dir,
os.path.join(self.args.data, str(num_dirs), 'sick_train.pth'),
vocab, self.args.num_classes)
self.dev_dataset = Helper.load_data(
dev_dir, os.path.join(self.args.data, str(num_dirs),
'sick_dev.pth'), vocab,
self.args.num_classes)
self.test_dataset = Helper.load_data(
test_dir,
os.path.join(self.args.data, str(num_dirs), 'sick_test.pth'),
vocab, self.args.num_classes)
# creates the TreeLSTM
model = SimilarityTreeLSTM(vocab.size(), self.args.input_dim, self.args.mem_dim, self.args.hidden_dim, \
self.args.num_classes, self.args.sparse, self.args.freeze_embed, vocab)
criterion = nn.KLDivLoss() #nn.CrossEntropyLoss()
# loads glove embeddings
emb = Helper.load_embeddings(
self.args,
os.path.join(self.args.data, str(num_dirs), 'sick_embed.pth'),
vocab, device)
# sets up the model
model.emb.weight.data.copy_(
emb) # plug these into embedding matrix inside model
model.to(device)
criterion.to(device)
opt = optim.Adagrad(filter(lambda p: p.requires_grad, \
model.parameters()), lr=lr, weight_decay=self.args.wd)
if opt == "adam":
opt = optim.Adam(filter(lambda p: p.requires_grad, \
model.parameters()), lr=lr)
self.metrics = Metrics(self.args.num_classes)
# create trainer object for training and testing
self.trainer = Trainer(self.args, model, criterion, opt, device, vocab)
