def main():
parser = argparse.ArgumentParser()
parser.add_argument('input_file', type=argparse.FileType())
parser.add_argument('output_file', type=argparse.FileType('w'))
parser.add_argument('vector_size', type=int)
parser.add_argument('context_size', type=int)
parser.add_argument('vocabulary_size', type=int)
args = parser.parse_args()
sentences = list(lower(tokenize(args.input_file)))
dictionary = build_dictionary(sentences, args.vocabulary_size)
indices = to_indices(sentences, dictionary)
inputs, outputs = create_context(indices, args.context_size)
cost_gradient = bind_cost_gradient(skip_gram_cost_gradient, inputs, outputs, sampler=get_stochastic_sampler(100))
initial_parameters = np.random.normal(size=(2, len(dictionary) + 1, args.vector_size))
parameters, cost_history = gradient_descent(cost_gradient, initial_parameters, 10000)
input_vectors, output_vectors = parameters
word_vectors = input_vectors + output_vectors
sorted_pairs = sorted(dictionary.items(), key=operator.itemgetter(1))
words = [word for word, index in sorted_pairs]
for word in words:
vector = word_vectors[dictionary[word]]
vector_string = ' '.join(str(element) for element in vector)
print(word, vector_string, file=args.output_file)
def train(self, sentences, iterations=1000):
# Preprocess sentences to create indices of context and next words
self.dictionary = build_dictionary(sentences, self.vocabulary_size)
indices = to_indices(sentences, self.dictionary)
self.reverse_dictionary = {
index: word
for word, index in self.dictionary.items()
}
inputs, outputs = self.create_context(indices)
# Create cost and gradient function for gradient descent
shapes = [self.W_shape, self.U_shape, self.H_shape, self.C_shape]
flatten_nplm_cost_gradient = flatten_cost_gradient(
nplm_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(flatten_nplm_cost_gradient,
inputs,
outputs,
sampler=get_stochastic_sampler(10))
# Train neural network
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
self.parameters, cost_history = gradient_descent(
cost_gradient, initial_parameters, iterations)
return cost_history
def test_supervised_gradient_descent(self):
def linear_regression_cost_gradient(parameters, input, output):
prediction = np.dot(parameters, input)
cost = (prediction - output)**2
gradient = 2.0 * (prediction - output) * input
return cost, gradient
inputs = np.random.normal(0.0, size=(10, 2))
outputs = np.random.normal(0.0, size=10)
initial_parameters = np.random.uniform(-1.0, 1.0, size=2)
# Create cost and gradient function for supervised SGD and check its gradient
cost_gradient = bind_cost_gradient(linear_regression_cost_gradient,
inputs,
outputs,
sampler=batch_sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Run gradient descent on the function and see if it minimizes cost function
actual, cost_history = gradient_descent(cost_gradient,
initial_parameters, 10)
# Compute exact solution of linear regression by closed form
expected = np.linalg.solve(np.dot(inputs.T, inputs),
np.dot(inputs.T, outputs))
for e, a in zip(expected, actual):
self.assertAlmostEqual(e, a, places=0)
def assertMultinomialLogisticRegression(self, sampler):
data_size = 3
input_size = 5
output_size = 4
inputs = np.random.uniform(-10.0, 10.0, size=(data_size, input_size))
outputs = np.random.randint(0, output_size, size=data_size)
initial_parameters = np.random.normal(size=(input_size, output_size))
# Create cost and gradient function for gradient descent and check its gradient
cost_gradient = bind_cost_gradient(
multinomial_logistic_regression_cost_gradient,
inputs,
outputs,
sampler=sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train multinomial logistic regression and see if it predicts correct labels
final_parameters, cost_history = gradient_descent(
cost_gradient, initial_parameters, 100)
predictions = np.argmax(softmax(np.dot(final_parameters.T, inputs.T)),
axis=0)
for output, prediction in zip(outputs, predictions):
self.assertEqual(output, prediction)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('input_file', type=argparse.FileType())
parser.add_argument('output_file', type=argparse.FileType('w'))
parser.add_argument('vector_size', type=int)
parser.add_argument('context_size', type=int)
parser.add_argument('vocabulary_size', type=int)
args = parser.parse_args()
sentences = list(lower(tokenize(args.input_file)))
dictionary = build_dictionary(sentences, args.vocabulary_size)
indices = to_indices(sentences, dictionary)
inputs, outputs = create_context(indices, args.context_size)
cost_gradient = bind_cost_gradient(skip_gram_cost_gradient,
inputs,
outputs,
sampler=get_stochastic_sampler(100))
initial_parameters = np.random.normal(size=(2, len(dictionary) + 1,
args.vector_size))
parameters, cost_history = gradient_descent(cost_gradient,
initial_parameters, 10000)
input_vectors, output_vectors = parameters
word_vectors = input_vectors + output_vectors
sorted_pairs = sorted(dictionary.items(), key=operator.itemgetter(1))
words = [word for word, index in sorted_pairs]
for word in words:
vector = word_vectors[dictionary[word]]
vector_string = ' '.join(str(element) for element in vector)
print(word, vector_string, file=args.output_file)
def test_supervised_gradient_descent(self):
def linear_regression_cost_gradient(parameters, input, output):
prediction = np.dot(parameters, input)
cost = (prediction - output) ** 2
gradient = 2.0 * (prediction - output) * input
return cost, gradient
inputs = np.random.normal(0.0, size=(10, 2))
outputs = np.random.normal(0.0, size=10)
initial_parameters = np.random.uniform(-1.0, 1.0, size=2)
# Create cost and gradient function for supervised SGD and check its gradient
cost_gradient = bind_cost_gradient(linear_regression_cost_gradient,
inputs, outputs, sampler=batch_sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Run gradient descent on the function and see if it minimizes cost function
actual, cost_history = gradient_descent(cost_gradient, initial_parameters, 10)
# Compute exact solution of linear regression by closed form
expected = np.linalg.solve(np.dot(inputs.T, inputs), np.dot(inputs.T, outputs))
for e, a in zip(expected, actual):
self.assertAlmostEqual(e, a, places=0)
def gradient_check(self, inputs, outputs):
# Create cost and gradient function for gradient check
shapes = [self.W_shape, self.U_shape, self.H_shape, self.C_shape]
flatten_nplm_cost_gradient = flatten_cost_gradient(nplm_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(flatten_nplm_cost_gradient, inputs, outputs)
# Gradient check!
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
result = gradient_check(cost_gradient, initial_parameters)
return result
def test_neural_network(self):
np.random.seed(0)
input_size = 2
hidden_size = 2
output_size = 2
# Classic XOR test data
inputs = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
outputs = np.array([0, 1, 1, 0])
# Create cost and gradient function for gradient descent
shapes = [(hidden_size, (input_size)), (output_size, (hidden_size))]
flatten_neural_network_cost_gradient = flatten_cost_gradient(
neural_network_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(
flatten_neural_network_cost_gradient,
inputs,
outputs,
sampler=batch_sampler)
# Check gradient with initial parameters
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train neural network (this is slow even such a simple task!)
final_parameters, cost_history = gradient_descent(
cost_gradient, initial_parameters, 1000)
# Check if cost monotonically decrease (no guarantee in theory, but works in practice)
previous_cost = None
for cost in cost_history:
if previous_cost is not None:
self.assertLessEqual(cost, previous_cost)
previous_cost = cost
# TODO: extract duplicated code for prediction to reusable component
split_index = hidden_size * (input_size)
W1, W2 = np.split(final_parameters, [split_index])
W1 = W1.reshape((hidden_size, input_size))
W2 = W2.reshape((output_size, hidden_size))
for input, output in zip(inputs, outputs):
input = input.reshape(-1, 1)
hidden_layer = expit(W1.dot(input))
inside_softmax = W2.dot(hidden_layer)
prediction = softmax(inside_softmax.reshape(-1)).reshape(-1, 1)
label = np.argmax(prediction)
# Check if output is correctly predicted
self.assertEqual(output, label)
def gradient_check(self, inputs, outputs):
# Create cost and gradient function for gradient check
shapes = [self.W_shape, self.U_shape, self.H_shape, self.C_shape]
flatten_nplm_cost_gradient = flatten_cost_gradient(
nplm_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(flatten_nplm_cost_gradient, inputs,
outputs)
# Gradient check!
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
result = gradient_check(cost_gradient, initial_parameters)
return result
def test_neural_network(self):
np.random.seed(0)
input_size = 2
hidden_size = 2
output_size = 2
# Classic XOR test data
inputs = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
outputs = np.array([0, 1, 1, 0])
# Create cost and gradient function for gradient descent
shapes = [(hidden_size, (input_size)), (output_size, (hidden_size))]
flatten_neural_network_cost_gradient = flatten_cost_gradient(neural_network_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(flatten_neural_network_cost_gradient, inputs, outputs, sampler=batch_sampler)
# Check gradient with initial parameters
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train neural network (this is slow even such a simple task!)
final_parameters, cost_history = gradient_descent(cost_gradient, initial_parameters, 1000)
# Check if cost monotonically decrease (no guarantee in theory, but works in practice)
previous_cost = None
for cost in cost_history:
if previous_cost is not None:
self.assertLessEqual(cost, previous_cost)
previous_cost = cost
# TODO: extract duplicated code for prediction to reusable component
split_index = hidden_size * (input_size)
W1, W2 = np.split(final_parameters, [split_index])
W1 = W1.reshape((hidden_size, input_size))
W2 = W2.reshape((output_size, hidden_size ))
for input, output in zip(inputs, outputs):
input = input.reshape(-1, 1)
hidden_layer = expit(W1.dot(input))
inside_softmax = W2.dot(hidden_layer)
prediction = softmax(inside_softmax.reshape(-1)).reshape(-1, 1)
label = np.argmax(prediction)
# Check if output is correctly predicted
self.assertEqual(output, label)
def train(self, sentences, iterations=1000):
# Preprocess sentences to create indices of context and next words
self.dictionary = build_dictionary(sentences, self.vocabulary_size)
indices = to_indices(sentences, self.dictionary)
self.reverse_dictionary = {index: word for word, index in self.dictionary.items()}
inputs, outputs = self.create_context(indices)
# Create cost and gradient function for gradient descent
shapes = [self.W_shape, self.U_shape, self.H_shape, self.C_shape]
flatten_nplm_cost_gradient = flatten_cost_gradient(nplm_cost_gradient, shapes)
cost_gradient = bind_cost_gradient(flatten_nplm_cost_gradient, inputs, outputs,
sampler=get_stochastic_sampler(10))
# Train neural network
parameters_size = np.sum(np.product(shape) for shape in shapes)
initial_parameters = np.random.normal(size=parameters_size)
self.parameters, cost_history = gradient_descent(cost_gradient, initial_parameters, iterations)
return cost_history
def assertMultinomialLogisticRegression(self, sampler):
data_size = 3
input_size = 5
output_size = 4
inputs = np.random.uniform(-10.0, 10.0, size=(data_size, input_size))
outputs = np.random.randint(0, output_size, size=data_size)
initial_parameters = np.random.normal(size=(input_size, output_size))
# Create cost and gradient function for gradient descent and check its gradient
cost_gradient = bind_cost_gradient(multinomial_logistic_regression_cost_gradient,
inputs, outputs, sampler=sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train multinomial logistic regression and see if it predicts correct labels
final_parameters, cost_history = gradient_descent(cost_gradient, initial_parameters, 100)
predictions = np.argmax(softmax(np.dot(final_parameters.T, inputs.T)), axis=0)
for output, prediction in zip(outputs, predictions):
self.assertEqual(output, prediction)
def assertLogisticRegression(self, sampler):
data_size = 3
input_size = 5
inputs = np.random.uniform(-10.0, 10.0, size=(data_size, input_size))
outputs = np.random.randint(0, 2, size=data_size)
initial_parameters = np.random.normal(scale=1e-5, size=input_size)
# Create cost and gradient function for gradient descent and check its gradient
cost_gradient = bind_cost_gradient(logistic_regression_cost_gradient,
inputs, outputs, sampler=sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train logistic regression and see if it predicts correct labels
final_parameters, cost_history = gradient_descent(cost_gradient, initial_parameters, 100)
predictions = expit(np.dot(inputs, final_parameters)) > 0.5
# Binary classification of 3 data points with 5 dimension is always linearly separable
for output, prediction in zip(outputs, predictions):
self.assertEqual(output, prediction)
def assertLogisticRegression(self, sampler):
data_size = 3
input_size = 5
inputs = np.random.uniform(-10.0, 10.0, size=(data_size, input_size))
outputs = np.random.randint(0, 2, size=data_size)
initial_parameters = np.random.normal(scale=1e-5, size=input_size)
# Create cost and gradient function for gradient descent and check its gradient
cost_gradient = bind_cost_gradient(logistic_regression_cost_gradient,
inputs,
outputs,
sampler=sampler)
result = gradient_check(cost_gradient, initial_parameters)
self.assertEqual([], result)
# Train logistic regression and see if it predicts correct labels
final_parameters, cost_history = gradient_descent(
cost_gradient, initial_parameters, 100)
predictions = expit(np.dot(inputs, final_parameters)) > 0.5
# Binary classification of 3 data points with 5 dimension is always linearly separable
for output, prediction in zip(outputs, predictions):
self.assertEqual(output, prediction)