def get_data():
vocab_dict = load_vocab_dict()
data = pickling.load(glovar.PKL_DIR, 'annotated_dep_trees.pkl')
train = SSTDataset(data['train'], vocab_dict)
dev = SSTDataset(data['dev'], vocab_dict)
test = SSTDataset(data['test'], vocab_dict)
return train, dev, test
def train(self, config):
if config['target'] == 'negs':
return pickling.load(glovar.DATA_DIR, 'train_negs', ['arct'])
else:
data = self.data('train-full')
if config['train_subsample'] > 0:
return random.sample(data, config['train_subsample'])
else:
return data
def get(pkl_dir, name, override, arg_config):
print('Getting history with name %s; override=%s...' % (name, override))
pkl_name = 'history_%s.pkl' % name
exists = os.path.exists(os.path.join(pkl_dir, pkl_name))
print('Exists: %s' % exists)
if exists:
if override:
print('Overriding...')
return History(name, models.Config(**arg_config))
else:
print('Loading...')
return pickling.load(pkl_dir, pkl_name)
else:
print('Creating...')
return History(name, models.Config(**arg_config))
def load_vocab_dict():
return pickling.load(glovar.PKL_DIR, 'vocab_dict.pkl')
params, arg_config = parameters.parse_arguments()
# Get or create History
history = histories.get(
glovar.PKL_DIR, params.name, params.override, arg_config)
# Report config to be used
config = history.config
print(config)
# Get vocab dict and embeddings
print('Load vocab dict and embedding matrix...')
vocab_dict = pickling.load(glovar.PKL_DIR, 'vocab_dict.pkl')
embedding_matrix = pickling.load(glovar.PKL_DIR, 'glove_embeddings.pkl')[0]
print('Loading data...')
mnli_train = nli.load_json('mnli', 'train')
snli_train = nli.load_json('snli', 'train')
mnli_dev_matched = nli.load_json('mnli', 'dev_matched')
train_data = nli.NYUDataSet(mnli_train, snli_train, vocab_dict)
tune_data = nli.NLIDataSet(mnli_dev_matched, vocab_dict)
train_loader = nli.get_data_loader(train_data, config.batch_size)
dev_loader = nli.get_data_loader(tune_data, config.batch_size)
print('Loading model...')
model = inference.InferenceModel(params.name, config, embedding_matrix)
def load(name):
pkl_name = 'history_%s.pkl' % name
return pickling.load(glovar.PKL_DIR, pkl_name)
def encoder_state_dict(transfer_name):
np_dict = pickling.load(glovar.DATA_DIR, '%s_dict_np' % transfer_name)
return dict(
zip(np_dict.keys(), [torch.from_numpy(a) for a in np_dict.values()]))
def vocab(self, config):
return pickling.load(glovar.DATA_DIR, 'vocab', ['arct'])
def embeddings(self, config):
return pickling.load(glovar.DATA_DIR, config['embed_type'], ['arct'])
def vocab(self, config):
return pickling.load(glovar.DATA_DIR, 'vocab', [config['target']])
def embeddings(self, config):
print(config)
return pickling.load(glovar.DATA_DIR, config['embed_type'],
[config['target']])
from ext import parameters, pickling, training, histories
from models import sentiment
# Parse configuration settings from command line
params, arg_config = parameters.parse_arguments()
# Get or create History
history = histories.get(glovar.PKL_DIR, params.name, params.override,
arg_config)
# Report config to be used
config = history.config
print(config)
print('Load embedding matrix...')
embedding_matrix = pickling.load(glovar.PKL_DIR, 'glove_embeddings.pkl')[0]
print('Loading data...')
train_data, dev_data, _ = sst.get_data()
train_loader = sst.get_data_loader(train_data, config.batch_size)
dev_loader = sst.get_data_loader(dev_data, config.batch_size)
print('Loading model...')
model = sentiment.SentimentModel(params.name, config, embedding_matrix)
print('Loading trainer...')
trainer = training.PyTorchTrainer(model, history, train_loader, dev_loader,
glovar.CKPT_DIR)
print('Training...')
trainer.train()