def load_train_and_test_bow(train_ixes,
test_ixes,
top_n_to_inc=None,
include_auto_labeled=None):
# total of 39
X, Y = load_subtask1_data(train_ixes, tokenized_folder=tokenized_dir)
#X_auto = open(os.path.join(ext_data_dir,
# 'top_3_auto_labeled_from_brown_external.txt')).readlines()
if top_n_to_inc is not None:
X_auto, Y_auto = load_subtask1_brown_auto_labeled(top_n=top_n_to_inc)
#p = os.path.join(ext_data_dir, include_auto_labeled)
#X_auto = open(p, encoding='utf-8').readlines()
#Y_auto = np.ones(len(X_auto))
X = np.concatenate([X, X_auto])
Y = np.concatenate([Y, Y_auto])
X_test, Y_test = load_subtask1_data(test_ixes,
tokenized_folder=tokenized_dir)
cvec = CountVectorizer(ngram_range=(1, 2), stop_words='english', min_df=0)
cvec_X = cvec.fit_transform(X).toarray()
cvec_X_test = cvec.transform(X_test).toarray()
return cvec_X, Y, cvec_X_test, Y_test
def fit_committee(models=None, feature_type='bow'):
if models is None:
models = dict(nb=MultinomialNB(alpha=3.25),
gb=GradientBoostingClassifier(n_estimators=170,
max_depth=5,
learning_rate=0.5,
min_samples_leaf=3,
min_samples_split=4),
dt=DecisionTreeClassifier(criterion='gini',
max_depth=25,
max_leaf_nodes=None,
min_samples_leaf=3,
min_samples_split=4))
raw_X, Y = load_subtask1_data(list(range(53)))
if feature_type == 'bow':
cvec = CountVectorizer(ngram_range=(1, 2),
stop_words='english',
min_df=3)
X = cvec.fit_transform(raw_X).toarray()
vote_clf = VotingClassifier(estimators=[(k, v) for k, v in models.items()],
n_jobs=3,
flatten_transform=False,
voting='soft').fit(X, Y)
return vote_clf
def build_train_count_vectorizer(raw_data=None):
if raw_data is None:
raw_data, _ = load_subtask1_data(list(range(53)))
cvec = CountVectorizer(ngram_range=(1, 2), stop_words='english', min_df=3)
return cvec.fit(raw_data)
def run():
np.random.seed(42)
file_ixs = list(range(39))
np.random.shuffle(file_ixs)
X, Y = load_subtask1_data(file_ixs[:25])
X_test, Y_test = load_subtask1_data(file_ixs[25:35])
tokenizer, sequences = preprocessing.tokenize_texts(X, nb_words=1000)
test_sequences = tokenizer.texts_to_sequences(X_test)
# TODO: replace with pad_packed_sequence in torch?
test_sequences = pad_sequences(test_sequences, maxlen=100)
np_embeddings = loaders.load_glove_wiki_embedding(tokenizer.word_index)
######
print("Torch Time")
torch_Y = torch.from_numpy(np.array(Y).astype('float32'))
torch_Y = Variable(torch_Y)
torch_sequences = torch.LongTensor(sequences.astype('long'))
var_torch_sequences = Variable(torch_sequences)
###
torch_Y_test = torch.from_numpy(np.array(Y_test).astype('float32'))
torch_Y_test = Variable(torch_Y_test)
torch_test_sequences = torch.LongTensor(test_sequences.astype('long'))
var_torch_test_sequences = Variable(torch_test_sequences)
# Try various learning rates
res = {
lr: train(Task1TorchRNN(np_embeddings, n_lstm_layers=2, hidden_dim=30),
var_torch_sequences,
torch_Y,
var_torch_test_sequences,
torch_Y_test,
lr=lr)
for lr in [0.0025, 0.003, 0.005, 0.01, 0.001]
}
print(res)
def load_train_and_test_bow(train_ixes,
test_ixes,
top_n_to_inc=None,
include_auto_labeled=None,
resample=None):
# total of 39
X, Y = load_subtask1_data(train_ixes, tokenized_folder=tokenized_dir)
#X_auto = open(os.path.join(ext_data_dir,
# 'top_3_auto_labeled_from_brown_external.txt')).readlines()
if top_n_to_inc is not None:
#X_auto, Y_auto = load_subtask1_brown_auto_labeled(top_n=top_n_to_inc)
X_auto, Y_auto = load_auto_labeled(top_n=top_n_to_inc)
#p = os.path.join(ext_data_dir, include_auto_labeled)
#X_auto = open(p, encoding='utf-8').readlines()
#Y_auto = np.ones(len(X_auto))
X = np.concatenate([X, X_auto])
Y = np.concatenate([Y, Y_auto])
if resample is not None:
print("Resample N: %s" % str(resample))
# Easiest to just load into a DF for group-by-then-sample
_df = pd.DataFrame(X)
_df['target'] = Y
rs_df = _df.groupby('target').apply(
lambda df: df.sample(resample, replace=True))
X = rs_df.drop('target', axis=1).values.reshape(-1)
Y = rs_df['target'].values.reshape(-1)
X_test, Y_test = load_subtask1_data(test_ixes,
tokenized_folder=tokenized_dir)
cvec = CountVectorizer(ngram_range=(1, 2), stop_words='english', min_df=0)
cvec_X = cvec.fit_transform(X).toarray()
cvec_X_test = cvec.transform(X_test).toarray()
return cvec_X, Y, cvec_X_test, Y_test
def load_data(embedding_dim=100, return_holdout=False):
file_ixs = list(range(65))
X, Y = load_subtask1_data(file_ixs[:40])
# Add in auto labeled
X_auto, Y_auto = load_subtask1_brown_auto_labeled()
X = np.concatenate([X, X_auto])
Y = np.concatenate([Y, Y_auto])
ix = list(range(len(X)))
np.random.shuffle(ix)
X = X[ix]
Y = Y[ix]
X_test, Y_test = load_subtask1_data(file_ixs[40:53])
test_ix = list(range(len(X_test)))
np.random.shuffle(test_ix)
X_test = X_test[test_ix]
Y_test = Y_test[test_ix]
tokenizer, sequences = preprocessing.tokenize_texts(X, nb_words=5000)
# TODO: Masking implementation rather than padding?
sequences = pad_sequences(sequences, maxlen=100)
test_sequences = tokenizer.texts_to_sequences(X_test)
test_sequences = pad_sequences(test_sequences, maxlen=100)
embeddings = loaders.load_glove_wiki_embedding(tokenizer.word_index,
embedding_dim=embedding_dim)
if return_holdout:
X_holdout, Y_holdout = load_subtask1_data(file_ixs[53:])
holdout_sequences = tokenizer.texts_to_sequences(X_holdout)
holdout_sequences = pad_sequences(holdout_sequences, maxlen=100)
return embeddings, sequences, Y, test_sequences, Y_test, holdout_sequences, Y_holdout
else:
return embeddings, sequences, Y, test_sequences, Y_test
def run():
np.random.seed(42)
file_ixs = list(range(39))
np.random.shuffle(file_ixs)
X, Y = load_subtask1_data(file_ixs[:20])
X_test, Y_test = load_subtask1_data(file_ixs[20:30])
# Turn sentences into sequences of integers, with each integer representing
# a word. Only the top occuring nb_words are kept
tokenizer, sequences = preprocessing.tokenize_texts(X, nb_words=1000)
sequences = pad_sequences(sequences, maxlen=50)
# Tokenizer does conversion on new texts - treat test as unseen
test_sequences = tokenizer.texts_to_sequences(X_test)
# TODO: replace with pad_packed_sequence in torch?
test_sequences = pad_sequences(test_sequences, maxlen=50)
np_embeddings = loaders.load_glove_wiki_embedding(tokenizer.word_index)
######
print("Torch Time")
torch_Y = torch.from_numpy(np.array(Y).astype('float32'))
torch_Y = Variable(torch_Y)
#torch_sequences = torch.LongTensor(np.array([np.array(s).astype('int')
# for s in sequences]))
#torch_sequences = [torch.LongTensor(np.array(s).astype('long'))
# for s in sequences]
# Need to pad if doing this way
torch_sequences = torch.LongTensor(np.array(sequences).astype('long'))
var_torch_sequences = Variable(torch_sequences)
###
torch_Y_test = torch.from_numpy(np.array(Y_test).astype('float32'))
torch_Y_test = Variable(torch_Y_test)
torch_test_sequences = torch.LongTensor(test_sequences.astype('long'))
var_torch_test_sequences = Variable(torch_test_sequences)
# Try various learning rates
res = dict()
for lr in [0.001, 0.0009, 0.0011]: #[0.0001, 0.00017, 0.00025, 0.0035]:
print("=" * 20)
print(lr)
res[lr] = train(Task1TorchRNN(np_embeddings,
n_lstm_layers=2,
hidden_dim=10),
var_torch_sequences,
torch_Y,
var_torch_test_sequences,
torch_Y_test,
lr=lr,
weight_decay=0.0031)
print(max([perf['val_acc'] for perf in res[lr]]))
for lr, r in res.items():
print("lr: %f" % lr)
print(max([perf['val_acc'] for perf in r]))
fname = './task1_lstm_torch_results_%d.pkl' % int(time.time())
print("Saving results to %s" % fname)
with open(fname, 'wb') as f:
pickle.dump(res, f)
