def train(config):
'''The config object required for training is in train_config.py'''
config['hyperparameters']['seed'] = np.random.randint(1e5)
if config['dataset_name'] == 'mnist':
dataset = load_mnist(config['mnist_sample_shape'])
elif config['dataset_name'] == 'cifar10':
dataset = load_cifar10(flatten_input=(config['nn_type'] == 'mlp'))
else:
raise ValueError(err_msg)
nn = NN(data=dataset, **config['hyperparameters'])
start = timeit.default_timer()
train_logs = nn.train_loop(eval_each_epoch=config['eval_while_training'])
elapsed = round(timeit.default_timer()-start,4)
print(f'training runtime: {elapsed} seconds')
exp = utils.ExperimentResults()
exp.save(train_logs, 'train_logs')
exp.save(config, 'config')
exp.save(nn, 'neural_network')
exp.save(elapsed,'training_runtime')
test_results = nn.evaluate()
exp.save(test_results, 'test_results')
def mnist_net():
mndata = MNIST("mnist", return_type="numpy")
print("Loading images...")
images, labels = mndata.load_training()
features = images.T / 255
z = np.zeros((60000, 10))
z[np.arange(60000), labels] = 1
Y = z.T
nn = NN([784, 100, 30, 10])
nn.set_hyperparameters(learning_rate=0.5)
t = time()
nn.initialize_parameters()
print("Start Training...")
nn.minimize({"features": features, "labels": Y}, 20)
print("Finish Training.")
print("Training time: {0} seconds".format(round(time() - t, 2)))
print("Start Testing...")
t = time()
test_images, test_labels = mndata.load_testing()
test_features = test_images.T / 255
z = np.zeros((10000, 10))
z[np.arange(10000), test_labels] = 1
test_Y = z.T
print("Testing accuracy: {}".format(
round(nn.evaluate({
"features": test_features,
"labels": test_Y
}), 4)))
print("Testing time: {0} seconds".format(round(time() - t, 2)))
def train(hyperparameters, sample_shape, nn_type):
hyperparameters['seed'] = np.random.randint(1e5)
cifar10 = load_cifar10(flatten_input=(
nn_type == 'mlp')) # because mlp only processes 1D input
nn = NN(data=cifar10, **hyperparameters)
perform_evaluation = False
if nn_type == 'mlp':
perform_evaluation = True
elif hyperparameters['n_epochs'] == 1:
perform_evaluation = True
train_logs = nn.train_loop(eval_each_epoch=perform_evaluation)
exp = utils.ExperimentResults()
exp.save(train_logs, 'train_logs')
exp.save(hyperparameters, 'hyperparams')
exp.save(nn, 'neural_network')
test_results = nn.evaluate()
exp.save(test_results, 'test_results')
def xor_net():
a = np.array([[0, 0, 0], [0, 1, 1], [1, 0, 1], [1, 1, 0]])
features = a[:, 0:2].T
labels = a[:, 2].reshape((1, 4))
z = np.zeros((4, 2))
z[np.arange(4), labels] = 1
labels = z.T
nn = NN([2, 4, 3, 2])
nn.set_hyperparameters(batch_size=4, learning_rate=0.75)
t = time()
nn.initialize_parameters()
print("Start Training...")
nn.minimize({"features": features, "labels": labels}, 10000)
print("Finish Training.")
print("Training time: {0} seconds".format(round(time() - t, 2)))
print("Start Testing...")
t = time()
print("Testing accuracy: {}".format(
round(nn.evaluate({
"features": features,
"labels": labels
}), 4)))
print("Testing time: {0} seconds".format(round(time() - t, 2)))