def test_compare_predictors(hang_plot=False):
from sklearn.svm import SVC
dataset = get_synthetic_clusters_dataset()
w_constructor = lambda rng=np.random.RandomState(45): .1 * rng.randn(
dataset.input_shape[0], dataset.n_categories)
records = compare_predictors(
dataset=dataset,
offline_predictors={'SVM': SVC()},
online_predictors={
'fast-perceptron':
Perceptron(alpha=0.1, w=w_constructor()).to_categorical(),
'slow-perceptron':
Perceptron(alpha=0.001, w=w_constructor()).to_categorical()
},
minibatch_size=10,
test_epochs=sqrtspace(0, 10, 20),
evaluation_function='percent_correct')
assert 99 < records['SVM'].get_scores('Test') <= 100
assert 20 < records['slow-perceptron'].get_scores(
'Test')[0] < 40 and 95 < records['slow-perceptron'].get_scores(
'Test')[-1] <= 100
assert 20 < records['fast-perceptron'].get_scores(
'Test')[0] < 40 and 98 < records['fast-perceptron'].get_scores(
'Test')[-1] <= 100
plot_learning_curves(records, hang=hang_plot)
def test_compare_predictors(hang_plot = False):
from sklearn.svm import SVC
dataset = get_synthetic_clusters_dataset()
w_constructor = lambda rng = np.random.RandomState(45): .1*rng.randn(dataset.input_shape[0], dataset.n_categories)
records = compare_predictors(
dataset = dataset,
offline_predictors={
'SVM': SVC()
},
online_predictors={
'fast-perceptron': Perceptron(alpha = 0.1, w = w_constructor()).to_categorical(),
'slow-perceptron': Perceptron(alpha = 0.001, w = w_constructor()).to_categorical()
},
minibatch_size = 10,
test_epochs = sqrtspace(0, 10, 20),
evaluation_function='percent_correct'
)
assert 99 < records['SVM'].get_scores('Test') <= 100
assert 20 < records['slow-perceptron'].get_scores('Test')[0] < 40 and 95 < records['slow-perceptron'].get_scores('Test')[-1] <= 100
assert 20 < records['fast-perceptron'].get_scores('Test')[0] < 40 and 98 < records['fast-perceptron'].get_scores('Test')[-1] <= 100
plot_learning_curves(records, hang = hang_plot)
def compare_predictors(dataset, online_predictors={}, offline_predictors={}, minibatch_size = 'full',
evaluation_function = 'mse', test_epochs = sqrtspace(0, 1, 10), report_test_scores = True,
test_on = 'training+test', test_batch_size = None, accumulators = None, online_test_callbacks = {}):
"""
DEPRECATED: use train_and_test_online_predictor instead.
Compare a set of predictors by running them on a dataset, and return the learning curves for each predictor.
:param dataset: A DataSet object
:param online_predictors: A dict<str:IPredictor> of online predictors. An online predictor is
sequentially fed minibatches of data and updates its parameters with each minibatch.
:param offline_predictors: A dict<str:object> of offline predictors. Offline predictors obey sklearn's
Estimator/Predictor interfaces - ie they methods
estimator = object.fit(data, targets) and
prediction = object.predict(data)
:param minibatch_size: Size of the minibatches to use for online predictors. Can be:
An int, in which case it represents the minibatch size for all classifiers.
A dict<str: int>, in which case you can set the minibatch size per-classifier.
In place of the int, you can put 'all' if you want to train on the whole dataset in each iteration.
:param test_epochs: Test points to use for online predictors. Can be:
A list of integers - in which case the classifier is after seeing this many samples.
A list of floats - in which case the classifier is tested after seeing this many epochs.
'always' - In which case a test is performed after every training step
The final test point determines the end of training.
:param evaluation_function: Function used to evaluate output of predictors
:param report_test_scores: Boolean indicating whether you'd like to report results online.
:param test_on: 'training', 'test', 'training+test'
:param test_batch_size: When the test set is too large to process in one step, use this to break it
up into chunks.
:param accumulators: A dict<str: accum_fcn>, where accum_fcn is a stateful-function of the form:
accmulated_output = accum_fcn(this_output)
Special case: accum_fcn can be 'avg' to make a running average.
:param online_test_callbacks: A dict<str: fcn> where fcn is a callback that takes an online
predictor as an argument. Useful for logging/plotting/debugging progress during training.
:return: An OrderedDict<LearningCurveData>
"""
all_keys = online_predictors.keys()+offline_predictors.keys()
assert len(all_keys) > 0, 'You have to give at least one predictor. Is that too much to ask?'
assert len(all_keys) == len(np.unique(all_keys)), "You have multiple predictors using the same names. Change that."
type_constructor_dict = OrderedDict(
[(k, ('offline', offline_predictors[k])) for k in sorted(offline_predictors.keys())] +
[(k, ('online', online_predictors[k])) for k in sorted(online_predictors.keys())]
)
minibatch_size = _pack_into_dict(minibatch_size, expected_keys=online_predictors.keys())
accumulators = _pack_into_dict(accumulators, expected_keys=online_predictors.keys())
online_test_callbacks = _pack_into_dict(online_test_callbacks, expected_keys=online_predictors.keys(), allow_subset=True)
test_epochs = np.array(test_epochs)
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
records = OrderedDict()
# Run the offline predictors
for predictor_name, (predictor_type, predictor) in type_constructor_dict.items():
print('%s\nRunning predictor %s\n%s' % ('='*20, predictor_name, '-'*20))
records[predictor_name] = \
assess_offline_predictor(
predictor=predictor,
dataset = dataset,
evaluation_function = evaluation_function,
report_test_scores = report_test_scores,
test_on = test_on,
test_batch_size = test_batch_size
) if predictor_type == 'offline' else \
assess_online_predictor(
predictor=predictor,
dataset = dataset,
evaluation_function = evaluation_function,
test_epochs = test_epochs,
accumulator = accumulators[predictor_name],
minibatch_size = minibatch_size[predictor_name],
report_test_scores = report_test_scores,
test_on = test_on,
test_batch_size = test_batch_size,
test_callback=online_test_callbacks[predictor_name] if predictor_name in online_test_callbacks else None
) if predictor_type == 'online' else \
bad_value(predictor_type)
print('Done!')
return records
def compare_predictors(dataset,
online_predictors={},
offline_predictors={},
minibatch_size='full',
evaluation_function='mse',
test_epochs=sqrtspace(0, 1, 10),
report_test_scores=True,
test_on='training+test',
test_batch_size=None,
accumulators=None,
online_test_callbacks={}):
"""
DEPRECATED: use train_and_test_online_predictor instead.
Compare a set of predictors by running them on a dataset, and return the learning curves for each predictor.
:param dataset: A DataSet object
:param online_predictors: A dict<str:IPredictor> of online predictors. An online predictor is
sequentially fed minibatches of data and updates its parameters with each minibatch.
:param offline_predictors: A dict<str:object> of offline predictors. Offline predictors obey sklearn's
Estimator/Predictor interfaces - ie they methods
estimator = object.fit(data, targets) and
prediction = object.predict(data)
:param minibatch_size: Size of the minibatches to use for online predictors. Can be:
An int, in which case it represents the minibatch size for all classifiers.
A dict<str: int>, in which case you can set the minibatch size per-classifier.
In place of the int, you can put 'all' if you want to train on the whole dataset in each iteration.
:param test_epochs: Test points to use for online predictors. Can be:
A list of integers - in which case the classifier is after seeing this many samples.
A list of floats - in which case the classifier is tested after seeing this many epochs.
'always' - In which case a test is performed after every training step
The final test point determines the end of training.
:param evaluation_function: Function used to evaluate output of predictors
:param report_test_scores: Boolean indicating whether you'd like to report results online.
:param test_on: 'training', 'test', 'training+test'
:param test_batch_size: When the test set is too large to process in one step, use this to break it
up into chunks.
:param accumulators: A dict<str: accum_fcn>, where accum_fcn is a stateful-function of the form:
accmulated_output = accum_fcn(this_output)
Special case: accum_fcn can be 'avg' to make a running average.
:param online_test_callbacks: A dict<str: fcn> where fcn is a callback that takes an online
predictor as an argument. Useful for logging/plotting/debugging progress during training.
:return: An OrderedDict<LearningCurveData>
"""
all_keys = online_predictors.keys() + offline_predictors.keys()
assert len(
all_keys
) > 0, 'You have to give at least one predictor. Is that too much to ask?'
assert len(all_keys) == len(
np.unique(all_keys)
), "You have multiple predictors using the same names. Change that."
type_constructor_dict = OrderedDict(
[(k, ('offline', offline_predictors[k]))
for k in sorted(offline_predictors.keys())] +
[(k, ('online', online_predictors[k]))
for k in sorted(online_predictors.keys())])
minibatch_size = _pack_into_dict(minibatch_size,
expected_keys=online_predictors.keys())
accumulators = _pack_into_dict(accumulators,
expected_keys=online_predictors.keys())
online_test_callbacks = _pack_into_dict(
online_test_callbacks,
expected_keys=online_predictors.keys(),
allow_subset=True)
test_epochs = np.array(test_epochs)
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
records = OrderedDict()
# Run the offline predictors
for predictor_name, (predictor_type,
predictor) in type_constructor_dict.items():
print('%s\nRunning predictor %s\n%s' %
('=' * 20, predictor_name, '-' * 20))
records[predictor_name] = \
assess_offline_predictor(
predictor=predictor,
dataset = dataset,
evaluation_function = evaluation_function,
report_test_scores = report_test_scores,
test_on = test_on,
test_batch_size = test_batch_size
) if predictor_type == 'offline' else \
assess_online_predictor(
predictor=predictor,
dataset = dataset,
evaluation_function = evaluation_function,
test_epochs = test_epochs,
accumulator = accumulators[predictor_name],
minibatch_size = minibatch_size[predictor_name],
report_test_scores = report_test_scores,
test_on = test_on,
test_batch_size = test_batch_size,
test_callback=online_test_callbacks[predictor_name] if predictor_name in online_test_callbacks else None
) if predictor_type == 'online' else \
bad_value(predictor_type)
print('Done!')
return records