def get_data(config):
train, dev, test = smt_dataset(
directory="../data/",
train=True,
dev=True,
test=True,
fine_grained=True,
)
def filter_neutrals(data, labels):
logger.info("Filtering neutral labels for binary task")
new_data, new_labels = [], []
for d, l in zip(data, labels):
# l positive or very positive
if "positive" in l:
new_data.append(d)
new_labels.append("positive")
# l negative or very negative
elif "negative" in l:
new_data.append(d)
new_labels.append("negative")
else:
continue
return new_data, new_labels
raw_train = [d["text"] for d in train]
labels_train = [d["label"] for d in train]
raw_dev = [d["text"] for d in dev]
labels_dev = [d["label"] for d in dev]
raw_test = [d["text"] for d in dev]
labels_test = [d["label"] for d in dev]
num_labels = 5
if config.binary:
raw_train, labels_train = filter_neutrals(raw_train, labels_train)
raw_dev, labels_dev = filter_neutrals(raw_dev, labels_dev)
raw_test, labels_test = filter_neutrals(raw_test, labels_test)
num_labels = 2
return (
raw_train,
labels_train,
raw_dev,
labels_dev,
raw_test,
labels_test,
num_labels,
)
def test_smt_dataset_row(mock_urlretrieve):
mock_urlretrieve.side_effect = urlretrieve_side_effect
# Check a row are parsed correctly
train, dev, test = smt_dataset(directory=directory,
test=True,
dev=True,
train=True)
assert len(train) > 0
assert len(dev) > 0
assert len(test) > 0
assert train[5] == {
'text':
"Whether or not you 're enlightened by any of Derrida 's lectures on `` the other '' "
+
"and `` the self , '' Derrida is an undeniably fascinating and playful fellow .",
'label':
'positive'
}
train = smt_dataset(directory=directory, train=True, subtrees=True)
assert train[3] == {'text': 'Rock', 'label': 'neutral'}
train = smt_dataset(directory=directory,
train=True,
subtrees=True,
fine_grained=True)
assert train[4] == {
'text':
"is destined to be the 21st Century 's new `` Conan '' and that he 's going to make a"
+
" splash even greater than Arnold Schwarzenegger , Jean-Claud Van Damme or Steven"
+ " Segal .",
'label':
'very positive'
}
# Clean up
shutil.rmtree(os.path.join(directory, 'trees'))
def __init__(self,
mode='train',
subtrees=False,
embedder=None,
tokenizer=None,
granularity=2,
threshold=3):
if granularity == 5:
fine_grained = True
else:
fine_grained = False
if tokenizer:
self.tokenizer = Tokenizer(tokenizer)
self.subtrees = subtrees
if mode == 'train':
self.data = list(
smt_dataset('sst/',
train=True,
fine_grained=fine_grained,
subtrees=self.subtrees))
if mode == 'val':
self.data = list(
smt_dataset('sst/',
train=False,
dev=True,
fine_grained=fine_grained,
subtrees=self.subtrees))
if mode == 'test':
self.data = list(
smt_dataset('sst/',
train=False,
test=True,
fine_grained=fine_grained,
subtrees=self.subtrees))
if fine_grained is True:
label_to_id = {}
label_to_id['very negative'] = 0
label_to_id['negative'] = 1
label_to_id['neutral'] = 2
label_to_id['positive'] = 3
label_to_id['very positive'] = 4
else:
label_to_id = {}
label_to_id['very negative'] = 0
label_to_id['negative'] = 0
label_to_id['neutral'] = 2
label_to_id['positive'] = 1
label_to_id['very positive'] = 1
if self.subtrees == False:
for i in self.data:
i['label'] = label_to_id[i['label']]
if tokenizer:
i['text'] = self.tokenizer.tokenize(i['text'])
else:
data_list = []
count = 0
for i in self.data:
if len(i['text'].split()) >= threshold:
label = label_to_id[i['label']]
if tokenizer:
text = self.tokenizer.tokenize(i['text'])
else:
text = i['text']
data_list.append({'text': text, 'label': label})
count += 1
self.data = data_list
del data_list
del_idxs = []
if fine_grained is False:
for i in range(len(self.data)):
if self.data[i]['label'] == 2:
del_idxs.append(i)
self.data = [
self.data[x] for x in range(len(self.data)) if x not in del_idxs
]
from slp.plbind.module import RnnPLModule
from slp.util.log import configure_logging
from slp.data.collators import SequenceClassificationCollator
from slp.modules.classifier import Classifier
from slp.modules.rnn import WordRNN
from slp.plbind.trainer import make_trainer, watch_model
from slp.plbind.helpers import FromLogits
collate_fn = SequenceClassificationCollator(device="cpu")
if __name__ == "__main__":
EXPERIMENT_NAME = "smt-words-sentiment-classification"
configure_logging(f"logs/{EXPERIMENT_NAME}")
train, dev = smt_dataset(directory="../data/", train=True, dev=True)
raw_train = [d["text"] for d in train]
labels_train = [d["label"] for d in train]
raw_dev = [d["text"] for d in dev]
labels_dev = [d["label"] for d in dev]
ldm = PLDataModuleFromCorpus(
raw_train,
labels_train,
val=raw_dev,
val_labels=labels_dev,
batch_size=8,
batch_size_eval=32,
collate_fn=collate_fn,
to_token_ids = ToTokenIds(word2idx)
to_tensor = ToTensor(device='cpu')
def create_dataloader(d):
d = (DatasetWrapper(d).map(tokenizer).map(to_token_ids).map(to_tensor))
return DataLoader(
d,
batch_size=8,
num_workers=1,
pin_memory=True,
shuffle=True,
collate_fn=collate_fn)
train_loader, dev_loader = map(
create_dataloader,
smt_dataset(directory='../data/', train=True, dev=True))
model = Classifier(
WordRNN(256, embeddings, bidirectional=True, merge_bi='cat',
packed_sequence=True, attention=False, device=DEVICE),
512, 3)
optimizer = Adam([p for p in model.parameters() if p.requires_grad],
lr=1e-3)
criterion = nn.CrossEntropyLoss()
metrics = {
'accuracy': Accuracy(),
'loss': Loss(criterion)
}
trainer = SequentialTrainer(model, optimizer,
checkpoint_dir='../checkpoints',
neut_indices = [i for i, x in enumerate(labels) if x == "neutral"]
neg_indices = [i for i, x in enumerate(labels) if x == "negative"]
very_neg_indices = [
i for i, x in enumerate(labels) if x == "very negative"
]
labels_tensor[very_pos_indices] = 0
labels_tensor[pos_indices] = 1
labels_tensor[neut_indices] = 2
labels_tensor[neg_indices] = 3
labels_tensor[very_neg_indices] = 4
return labels_tensor
train = smt_dataset(train=True, fine_grained=True)
valid = smt_dataset(dev=True, fine_grained=True)
test = smt_dataset(test=True, fine_grained=True)
train_labels = create_SMT_labels(train, len(train))
train_text = np.array(train.__getitem__('text'))
valid_labels = create_SMT_labels(valid, len(valid))
valid_text = np.array(valid.__getitem__('text'))
test_labels = create_SMT_labels(test, len(test))
test_text = np.array(test.__getitem__('text'))
np.save('sst_train_text', train_text)
np.save('sst_train_labels', train_labels)
np.save('sst_valid_text', valid_text)
np.save('sst_valid_labels', valid_labels)
np.save('sst_test_text', test_text)