def mnist_adamax_showdown(hidden_size = 300, n_epochs = 10, n_tests = 20):
dataset = get_mnist_dataset()
if is_test_mode():
dataset.shorten(200)
n_epochs = 0.1
n_tests = 3
make_mlp = lambda optimizer: GradientBasedPredictor(
function = MultiLayerPerceptron(
layer_sizes=[hidden_size, dataset.n_categories],
input_size = dataset.input_size,
hidden_activation='sig',
output_activation='lin',
w_init = normal_w_init(mag = 0.01, seed = 5)
),
cost_function = softmax_negative_log_likelihood,
optimizer = optimizer,
).compile()
return compare_predictors(
dataset=dataset,
online_predictors = {
'sgd': make_mlp(SimpleGradientDescent(eta = 0.1)),
'adamax': make_mlp(AdaMax(alpha = 1e-3)),
},
minibatch_size = 20,
test_epochs = sqrtspace(0, n_epochs, n_tests),
evaluation_function = percent_argmax_correct
)
def compare_example_predictors(
n_epochs = 5,
n_tests = 20,
minibatch_size = 10,
test_mode = False
):
"""
This demo shows how we can compare different online predictors. The demo trains both predictors on the dataset,
returning an object that contains the results.
:param test_mode: Set this to True to just run the demo quicky (but not to completion) to see that it doesn't break.
"""
dataset = get_mnist_dataset(flat = True)
# "Flatten" the 28x28 inputs to a 784-d vector
if test_mode:
# Shorten the dataset so we run through it quickly in test mode.
dataset = dataset.shorten(200)
n_epochs = 1
n_tests = 3
# Here we compare three predictors on MNIST - an MLP, a Perceptron, and a Random Forest.
# - The MLP is defined using Plato's interfaces - we create a Symbolic Predictor (GradientBasedPredictor) and
# then compile it into an IPredictor object
# - The Perceptron directly implements the IPredictor interface.
# - The Random Forest implements SciKit learn's predictor interface - that is, it has a fit(x, y) and a predict(x) method.
learning_curve_data = compare_predictors(
dataset = dataset,
online_predictors = {
'Perceptron': Perceptron(
w = np.zeros((dataset.input_size, dataset.n_categories)),
alpha = 0.001
).to_categorical(n_categories = dataset.n_categories), # .to_categorical allows the perceptron to be trained on integer labels.
'MLP': GradientBasedPredictor(
function = MultiLayerPerceptron(
layer_sizes=[500, dataset.n_categories],
input_size = dataset.input_size,
hidden_activation='sig', # Sigmoidal hidden units
output_activation='softmax', # Softmax output unit, since we're doing multinomial classification
w_init = normal_w_init(mag = 0.01, seed = 5)
),
cost_function = negative_log_likelihood_dangerous, # "Dangerous" because it doesn't check to see that output is normalized, but we know it is because it comes from softmax.
optimizer = SimpleGradientDescent(eta = 0.1),
).compile(), # .compile() returns an IPredictor
},
offline_predictors={
'RF': RandomForestClassifier(n_estimators = 40)
},
minibatch_size = minibatch_size,
test_epochs = sqrtspace(0, n_epochs, n_tests),
evaluation_function = percent_argmax_correct # Compares one-hot
)
# Results is a LearningCurveData object
return learning_curve_data
def mlp_normalization(hidden_size = 300, n_epochs = 30, n_tests = 50, minibatch_size=20):
"""
Compare mlps with different schemes for normalizing input.
regular: Regular vanilla MLP
normalize: Mean-subtract/normalize over minibatch
normalize and scale: Mean-subtract/normalize over minibatch AND multiply by a trainable
(per-unit) scale parameter.
Conclusions: No significant benefit to scale parameter. Normalizing gives
a head start but incurs a small cost later on. But really all classifiers are quite similar.
:param hidden_size: Size of hidden layer
"""
dataset = get_mnist_dataset()
if is_test_mode():
dataset.shorten(200)
n_epochs = 0.1
n_tests = 3
make_mlp = lambda normalize, scale: GradientBasedPredictor(
function = MultiLayerPerceptron(
layer_sizes=[hidden_size, dataset.n_categories],
input_size = dataset.input_size,
hidden_activation='sig',
output_activation='lin',
normalize_minibatch=normalize,
scale_param=scale,
w_init = normal_w_init(mag = 0.01, seed = 5)
),
cost_function = softmax_negative_log_likelihood,
optimizer = SimpleGradientDescent(eta = 0.1),
).compile()
return compare_predictors(
dataset=dataset,
online_predictors = {
'regular': make_mlp(normalize = False, scale = False),
'normalize': make_mlp(normalize=True, scale = False),
'normalize and scale': make_mlp(normalize=True, scale = True),
},
minibatch_size = minibatch_size,
test_epochs = sqrtspace(0, n_epochs, n_tests),
evaluation_function = percent_argmax_correct
)
def test_mlp():
assert_online_predictor_not_broken(
predictor_constructor = lambda n_dim_in, n_dim_out:
GradientBasedPredictor(
function = MultiLayerPerceptron(
layer_sizes = [100, n_dim_out],
input_size = n_dim_in,
output_activation='softmax',
w_init = lambda n_in, n_out, rng = np.random.RandomState(3252): 0.1*rng.randn(n_in, n_out)
),
cost_function=negative_log_likelihood_dangerous,
optimizer=SimpleGradientDescent(eta = 0.1),
).compile(),
categorical_target=True,
minibatch_size=10,
n_epochs=2
)
def demo_mnist_mlp(test_mode = False):
"""
Train an MLP on MNIST and print the test scores as training progresses.
"""
if test_mode:
test_period = 200
minibatch_size = 5
n_epochs = 0.01
dataset = get_mnist_dataset(n_training_samples=30, n_test_samples=30)
else:
test_period = 1000
minibatch_size = 20
n_epochs = 10
dataset = get_mnist_dataset()
# Setup the training and test functions
classifier = MultiLayerPerceptron(
layer_sizes=[500, 10],
input_size = 784,
hidden_activation='sig',
output_activation='softmax',
w_init = normal_w_init(mag = 0.01)
)
training_cost_function = normalized_negative_log_likelihood
optimizer = SimpleGradientDescent(eta = 0.1)
training_function = SupervisedTrainingFunction(classifier, training_cost_function, optimizer).compile()
test_cost_function = percent_correct
test_function = SupervisedTestFunction(classifier, test_cost_function).compile()
def report_test(i):
training_cost = test_function(dataset.training_set.input, dataset.training_set.target)
print 'Training score at iteration %s: %s' % (i, training_cost)
test_cost = test_function(dataset.test_set.input, dataset.test_set.target)
print 'Test score at iteration %s: %s' % (i, test_cost)
# Train and periodically report the test score.
print 'Running MLP on MNIST Dataset...'
for i, (_, image_minibatch, label_minibatch) in enumerate(dataset.training_set.minibatch_iterator(minibatch_size = minibatch_size, epochs = n_epochs, single_channel = True)):
if i % test_period == 0:
report_test(i)
training_function(image_minibatch, label_minibatch)
report_test('Final')
print '...Done.'
def test_param_serialization():
"""
Pros -
:return:
"""
dataset = get_synthetic_clusters_dataset()
predictor_constructor = lambda: GradientBasedPredictor(
function=MultiLayerPerceptron(layer_sizes=[100, dataset.n_categories],
input_size=dataset.input_shape[0],
output_activation='softmax',
w_init=lambda n_in, n_out, rng=np.random.
RandomState(3252): 0.1 * rng.randn(
n_in, n_out)),
cost_function=negative_log_likelihood_dangerous,
optimizer=SimpleGradientDescent(eta=0.1),
).compile()
evaluate = lambda pred: evaluate_predictor(pred, dataset.test_set,
percent_argmax_correct)
# Train up predictor and save params
predictor = predictor_constructor()
pre_training_score = evaluate(predictor)
assert pre_training_score < 35
train_online_predictor(predictor,
dataset.training_set,
minibatch_size=20,
n_epochs=3)
post_training_score = evaluate(predictor)
assert post_training_score > 95
trained_param_string = dumps_params(predictor)
# Instantiate new predictor and load params
new_predictor = predictor_constructor()
new_pre_training_score = evaluate(new_predictor)
assert new_pre_training_score < 35
loads_params(new_predictor, trained_param_string)
loaded_score = evaluate(new_predictor)
assert loaded_score == post_training_score > 95