def update_step(x_batch, y_batch, model, learning_rate):
"""Performs on single update step, (i.e. forward then backward).
Args:
x_batch(numpy.ndarray): input data of dimension (N, ndims).
y_batch(numpy.ndarray): label data of dimension (N, 1).
model(LinearModel): Initialized linear model.
"""
f = LinearRegression.forward(model, x_batch)
grad = learning_rate * LinearRegression.backward(model, f, y_batch)
model.w = model.w - learning_rate * grad
def main(_):
"""High level pipeline.
This script performs the trainsing, evaling and testing state of the model.
"""
# learning_rate = FLAGS.learning_rate
# feature_type = FLAGS.feature_type
# model_type = FLAGS.model_type
# num_steps = FLAGS.num_steps
feature_type = 'default'
model_type = 'svm'
# Load dataset.
data = read_dataset('data/train_lab.txt', 'data/image_data')
# Data Processing.
data = preprocess_data(data, 'default')
print("Finish preprocessing...")
# Initialize model.
ndim = data['image'].shape[1]
if model_type == 'linear':
model = LinearRegression(ndim, 'uniform')
elif model_type == 'logistic':
model = LogisticRegression(ndim, 'uniform')
elif model_type == 'svm':
model = SupportVectorMachine(ndim, 'uniform')
# Train Model.
print("Start to train the model...")
model = train_model(data, model)
# Eval Model.
print("Start to evaluate the model...")
data_val = read_dataset('data/val_lab.txt', 'data/image_data')
data_val = preprocess_data(data_val, feature_type)
loss, acc = eval_model(data_val, model)
print(loss, acc)
# Test Model.
print("Start doing the test")
data_test = read_dataset('data/test_lab.txt', 'data/image_data')
print("Start preprocess testing data")
data_test = preprocess_data(data_test, feature_type)
print("Making predictions")
data_test['label'] = model.predict(model.forward(data_test['image']))
print("Output the results to csv file")
write_dataset('data/test_lab.txt', data_test)
# Generate Kaggle output.
print("Finished!")
def test_input_output(self):
model = LinearRegression(10)
x = np.zeros([4, 10])
y = np.zeros([4, ])
# Check forward shape.
f = model.forward(x)
self.assertEqual(f.shape, (4,))
# Check backward shape.
gradient = model.backward(f, y)
self.assertEqual(gradient.shape, (11,))
# Check loss shape.
loss = model.loss(f, y)
self.assertEqual(loss.shape, ())
