def main():
""" main """
DATASET_LOCATION = '../datasets/'
POS_DATASET_NAME = 'brown_pos_dataset.hdf5'
POS_DATASET_PATH = os.path.join(DATASET_LOCATION, POS_DATASET_NAME)
WORD_BY_WORD_MATRIX = 'brown.word-by-word.normalized.npy'
VECTOR_INDEX_PATH = os.path.join(DATASET_LOCATION, WORD_BY_WORD_MATRIX)
CUTOFF = 100000
MODELS_PATH = "../trained_models/"
MODEL_NAME = "logistic_regression_model.pkl"
save_path = os.path.join(MODELS_PATH, MODEL_NAME)
train_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH, which_sets=['train'], cutoff=CUTOFF)
dev_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH, which_sets=['dev'], cutoff=CUTOFF)
# get the functions and params that we need for our models
initialization_data = initialize_logistic_regression(train_dataset, dev_dataset,
learning_rate=0.1, batch_size=100)
classifier, train_model_func, validate_model_func, n_train_batches, n_valid_batches = initialization_data
train_model(train_model_func, n_train_batches, validate_model=validate_model_func,
n_valid_batches=n_valid_batches, training_epochs=200)
save_model(classifier, save_path)
def main():
""" main """
DATASET_LOCATION = '../../datasets/'
# the pos dataset consists of windows around words
POS_DATASET_NAME = 'brown_pos_dataset.hdf5'
POS_DATASET_PATH = os.path.join(DATASET_LOCATION, POS_DATASET_NAME)
CORPUS_INDICES = 'brown_pos_dataset.indices'
WORD_BY_WORD_MATRIX = 'brown.word-by-word.normalized.npy'
# load the training data
VECTOR_INDEX_PATH = os.path.join(DATASET_LOCATION, WORD_BY_WORD_MATRIX)
CUTOFF = 10000
train_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH,
which_sets=['train'], cutoff=CUTOFF)
initialization_data = initialize_dA(train_dataset, learning_rate=0.1,
corruption_level=0.3, batch_size=50,
n_hidden=2)
classifier, train_model_func, validate_model_func, n_train_batches, n_valid_batches = initialization_data
train_model(train_model_func, n_train_batches,
validate_model=validate_model_func,
n_valid_batches=n_valid_batches, training_epochs=10)
# make a theano function to get predictions from a trained model
training_data = theano.tensor.matrix('training_X')
predictions = classifier.predict(training_data)
get_predictions = theano.function([training_data], predictions)
# get predictions and evaluate
p = get_predictions(train_dataset[0].get_value())
# get train_y without the cast
train_y = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH, which_sets=['train'],
cast_y=False, cutoff=CUTOFF)[1].get_value().astype('int32')
CUTOFF_BEGIN=0
CUTOFF_END=1000
y_vals = train_y[CUTOFF_BEGIN:CUTOFF_END]
norm_y_vals = y_vals / float(np.amax(y_vals))
jitter1 = np.random.normal(loc=0.0, scale=0.05, size=CUTOFF_END-CUTOFF_BEGIN)
jitter2 = np.random.normal(loc=0.0, scale=0.05, size=CUTOFF_END-CUTOFF_BEGIN)
x1 = p[CUTOFF_BEGIN:CUTOFF_END,0] + jitter1
x2 = p[CUTOFF_BEGIN:CUTOFF_END,1] + jitter2
plt.scatter(x1, x2, c=norm_y_vals, s=20)
plt.show()
def main():
DATASET_LOCATION = '../datasets/'
POS_DATASET_NAME = 'brown_pos_dataset.hdf5'
POS_DATASET_PATH = os.path.join(DATASET_LOCATION, POS_DATASET_NAME)
WORD_BY_WORD_MATRIX = 'brown.word-by-word.normalized.npy'
VECTOR_INDEX_PATH = os.path.join(DATASET_LOCATION, WORD_BY_WORD_MATRIX)
CUTOFF = 10000
MODELS_PATH = "../trained_models/"
MODEL_NAME = "mlp_model.pkl"
save_path = os.path.join(MODELS_PATH, MODEL_NAME)
train_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH,
which_sets=['train'], cutoff=CUTOFF)
dev_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH,
which_sets=['dev'], cutoff=CUTOFF)
# initialize the MLP
initialization_data = initialize_mlp(train_dataset, dev_dataset,
learning_rate=0.01, batch_size=50)
classifier, train_model_func, validate_model_func, n_train_batches, n_valid_batches = initialization_data
# train the MLP model
train_model(train_model_func, n_train_batches, validate_model=validate_model_func,
n_valid_batches=n_valid_batches, training_epochs=10)
test_dataset = prep_dataset(POS_DATASET_PATH, VECTOR_INDEX_PATH,
which_sets=['test'], cutoff=CUTOFF, cast_y=False)
test_X, test_y = test_dataset
test_y = test_y.get_value().astype('int32')
predictions = predict(classifier, test_X.get_value())
CORPUS_INDICES = 'brown_pos_dataset.indices'
with open(os.path.join(DATASET_LOCATION, CORPUS_INDICES)) as indices_file:
corpus_indices = cPickle.load(indices_file)
# map tag ids back to strings
y_test_actual = [corpus_indices['idx2tag'][tag_idx] for tag_idx in test_y]
y_test_hat = [corpus_indices['idx2tag'][tag_idx] for tag_idx in predictions]
# Quick Evaluation
acc = sum([y==p for y,p in zip(predictions, test_y)]) / float(len(predictions))
print "ACC: {}".format((acc))
# get class names
class_names = list(set(y_test_actual))
# Compute confusion matrix
cm = confusion_matrix(y_test_actual, y_test_hat, labels=class_names)
# Normalize the confusion matrix by row (i.e by the number of samples in each class)
cm_normalized = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
plt.figure()
plot_confusion_matrix(cm_normalized, class_names, title='Normalized confusion matrix')
plt.show()
