def assess_prediction_functions(test_pairs, functions, costs, print_results=False, prediction_minibatches = None):
"""
:param test_pairs: A list<pair_name, (x, y)>, where x, y are equal-length vectors representing the samples in a dataset.
Eg. [('training', (x_train, y_train)), ('test', (x_test, y_test))]
:param functions: A list<function_name, function> of functions for computing the forward pass.
:param costs: A list<(cost_name, cost_function)> or dict<cost_name: cost_function> of cost functions, where cost_function has the form:
cost = cost_fcn(guess, y), where cost is a scalar, and guess is the output of the prediction function given one
of the inputs (x) in test_pairs.
:param prediction_minibatches: Size of minibatches to predict in.
:return: A ModelTestScore object
"""
if isinstance(test_pairs, DataSet):
test_pairs = _dataset_to_test_pair(test_pairs)
assert isinstance(test_pairs, list)
assert all(len(_)==2 for _ in test_pairs)
assert all(len(pair)==2 for name, pair in test_pairs)
if isinstance(functions, dict):
functions = functions.items()
if callable(functions):
functions = [(functions.__name__ if hasattr(functions, '__name__') else None, functions)]
else:
assert all(callable(f) for name, f in functions)
if callable(costs):
costs = [(costs.__name__, costs)]
elif isinstance(costs, string_types):
costs = [(costs, get_evaluation_function(costs))]
elif isinstance(costs, dict):
costs = costs.items()
else:
costs = [(cost, get_evaluation_function(cost)) if isinstance(cost, string_types) else (cost.__name__, cost) if callable(cost) else cost for cost in costs]
assert all(callable(cost) for name, cost in costs)
results = ModelTestScore()
for test_pair_name, (x, y) in test_pairs:
for function_name, function in functions:
if prediction_minibatches is None:
predictions = function(x)
else:
predictions = minibatch_process(function, minibatch_size=prediction_minibatches, mb_args=(x, ))
for cost_name, cost_function in costs:
results[test_pair_name, function_name, cost_name] = cost_function(predictions, y)
if print_results:
print(results.get_table())
return results
def assess_prediction_functions(test_pairs,
functions,
costs,
print_results=False,
prediction_minibatches=None):
"""
:param test_pairs: A list<pair_name, (x, y)>, where x, y are equal-length vectors representing the samples in a dataset.
Eg. [('training', (x_train, y_train)), ('test', (x_test, y_test))]
:param functions: A list<function_name, function> of functions for computing the forward pass.
:param costs: A list<(cost_name, cost_function)> or dict<cost_name: cost_function> of cost functions, where cost_function has the form:
cost = cost_fcn(guess, y), where cost is a scalar, and guess is the output of the prediction function given one
of the inputs (x) in test_pairs.
:param prediction_minibatches: Size of minibatches to predict in.
:return: A ModelTestScore object
"""
if isinstance(test_pairs, DataSet):
test_pairs = _dataset_to_test_pair(test_pairs)
assert isinstance(test_pairs, list)
assert all(len(_) == 2 for _ in test_pairs)
assert all(len(pair) == 2 for name, pair in test_pairs)
if isinstance(functions, dict):
functions = functions.items()
if callable(functions):
functions = [
(functions.__name__ if hasattr(functions, '__name__') else None,
functions)
]
else:
assert all(callable(f) for name, f in functions)
if callable(costs):
costs = [(costs.__name__, costs)]
elif isinstance(costs, string_types):
costs = [(costs, get_evaluation_function(costs))]
elif isinstance(costs, dict):
costs = costs.items()
else:
costs = [(cost, get_evaluation_function(cost)) if isinstance(
cost, string_types) else
(cost.__name__, cost) if callable(cost) else cost
for cost in costs]
assert all(callable(cost) for name, cost in costs)
results = ModelTestScore()
for test_pair_name, (x, y) in test_pairs:
for function_name, function in functions:
if prediction_minibatches is None:
predictions = function(x)
else:
predictions = minibatch_process(
function,
minibatch_size=prediction_minibatches,
mb_args=(x, ))
for cost_name, cost_function in costs:
results[test_pair_name, function_name,
cost_name] = cost_function(predictions, y)
if print_results:
print(results.get_table())
return results
def evaluate_predictor(predictor, test_set, evaluation_function):
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
output = predictor.predict(test_set.input)
score = evaluation_function(actual=output, target=test_set.target)
return score
def assess_online_predictor(predictor, dataset, evaluation_function, test_epochs, minibatch_size, test_on = 'training+test',
accumulator = None, report_test_scores=True, test_batch_size = None, test_callback = None):
"""
DEPRECATED: use assess_prediction_functions_on_generator in train_and_test_old.py
Train an online predictor and return the LearningCurveData.
:param predictor: An IPredictor object
:param dataset: A DataSet object
:param evaluation_function: A function of the form: score=fcn(actual_values, target_values)
:param test_epochs: List of epochs to test at. Eg. [0.5, 1, 2, 4]
:param minibatch_size: Number of samples per minibatch, or:
'full' to do full-batch.
'stretch': to stretch the size of each batch so that we make just one call to "train" between each test. Use
this, for instance, if your predictor trains on one sample at a time in sequence anyway.
:param report_test_scores: Print out the test scores as they're computed (T/F)
:param test_callback: A callback which takes the predictor, and is called every time a test
is done. This can be useful for plotting/debugging the state.
:return: LearningCurveData containing the score on the test sets
"""
# TODO: Remove this class, as it is deprecated
record = LearningCurveData()
testing_sets = dataset_to_testing_sets(dataset, test_on)
if accumulator is None:
prediction_functions = {k: predictor.predict for k in testing_sets}
else:
accum_constructor = {'avg': RunningAverage}[accumulator] \
if isinstance(accumulator, str) else accumulator
accumulators = {k: accum_constructor() for k in testing_sets}
prediction_functions = {k: lambda inp, kp=k: accumulators[kp](predictor.predict(inp)) for k in testing_sets}
# Bewate the in-loop lambda - but I think we're ok here.
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
def do_test(current_epoch):
scores = [(k, evaluation_function(process_in_batches(prediction_functions[k], x, test_batch_size), y)) for k, (x, y) in testing_sets.items()]
if report_test_scores:
print('Scores at Epoch %s: %s, after %.2fs' % (current_epoch, ', '.join('%s: %.3f' % (set_name, score) for set_name, score in scores), time.time()-start_time))
record.add(current_epoch, scores)
if test_callback is not None:
record.add(current_epoch, ('callback', test_callback(predictor)))
start_time = time.time()
if minibatch_size == 'stretch':
test_samples = (np.array(test_epochs) * dataset.training_set.n_samples).astype(int)
i=0
if test_samples[0] == 0:
do_test(i)
i += 1
for indices in checkpoint_minibatch_index_generator(n_samples=dataset.training_set.n_samples, checkpoints=test_samples, slice_when_possible=True):
predictor.train(dataset.training_set.input[indices], dataset.training_set.target[indices])
do_test(test_epochs[i])
i += 1
else:
checker = CheckPointCounter(test_epochs)
last_n_samples_seen = 0
for (n_samples_seen, input_minibatch, target_minibatch) in \
dataset.training_set.minibatch_iterator(minibatch_size = minibatch_size, epochs = float('inf'), single_channel = True):
current_epoch = (float(last_n_samples_seen))/dataset.training_set.n_samples
last_n_samples_seen = n_samples_seen
time_for_a_test, done = checker.check(current_epoch)
if time_for_a_test:
do_test(current_epoch)
if done:
break
predictor.train(input_minibatch, target_minibatch)
return record
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 assess_online_predictor(predictor,
dataset,
evaluation_function,
test_epochs,
minibatch_size,
test_on='training+test',
accumulator=None,
report_test_scores=True,
test_batch_size=None,
test_callback=None):
"""
DEPRECATED: use assess_prediction_functions_on_generator in train_and_test_old.py
Train an online predictor and return the LearningCurveData.
:param predictor: An IPredictor object
:param dataset: A DataSet object
:param evaluation_function: A function of the form: score=fcn(actual_values, target_values)
:param test_epochs: List of epochs to test at. Eg. [0.5, 1, 2, 4]
:param minibatch_size: Number of samples per minibatch, or:
'full' to do full-batch.
'stretch': to stretch the size of each batch so that we make just one call to "train" between each test. Use
this, for instance, if your predictor trains on one sample at a time in sequence anyway.
:param report_test_scores: Print out the test scores as they're computed (T/F)
:param test_callback: A callback which takes the predictor, and is called every time a test
is done. This can be useful for plotting/debugging the state.
:return: LearningCurveData containing the score on the test sets
"""
# TODO: Remove this class, as it is deprecated
record = LearningCurveData()
testing_sets = dataset_to_testing_sets(dataset, test_on)
if accumulator is None:
prediction_functions = {k: predictor.predict for k in testing_sets}
else:
accum_constructor = {'avg': RunningAverage}[accumulator] \
if isinstance(accumulator, str) else accumulator
accumulators = {k: accum_constructor() for k in testing_sets}
prediction_functions = {
k: lambda inp, kp=k: accumulators[kp](predictor.predict(inp))
for k in testing_sets
}
# Bewate the in-loop lambda - but I think we're ok here.
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
def do_test(current_epoch):
scores = [(k,
evaluation_function(
process_in_batches(prediction_functions[k], x,
test_batch_size), y))
for k, (x, y) in testing_sets.items()]
if report_test_scores:
print('Scores at Epoch %s: %s, after %.2fs' %
(current_epoch, ', '.join('%s: %.3f' % (set_name, score)
for set_name, score in scores),
time.time() - start_time))
record.add(current_epoch, scores)
if test_callback is not None:
record.add(current_epoch, ('callback', test_callback(predictor)))
start_time = time.time()
if minibatch_size == 'stretch':
test_samples = (np.array(test_epochs) *
dataset.training_set.n_samples).astype(int)
i = 0
if test_samples[0] == 0:
do_test(i)
i += 1
for indices in checkpoint_minibatch_index_generator(
n_samples=dataset.training_set.n_samples,
checkpoints=test_samples,
slice_when_possible=True):
predictor.train(dataset.training_set.input[indices],
dataset.training_set.target[indices])
do_test(test_epochs[i])
i += 1
else:
checker = CheckPointCounter(test_epochs)
last_n_samples_seen = 0
for (n_samples_seen, input_minibatch, target_minibatch) in \
dataset.training_set.minibatch_iterator(minibatch_size = minibatch_size, epochs = float('inf'), single_channel = True):
current_epoch = (
float(last_n_samples_seen)) / dataset.training_set.n_samples
last_n_samples_seen = n_samples_seen
time_for_a_test, done = checker.check(current_epoch)
if time_for_a_test:
do_test(current_epoch)
if done:
break
predictor.train(input_minibatch, target_minibatch)
return record
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 evaluate_predictor(predictor, test_set, evaluation_function):
if isinstance(evaluation_function, str):
evaluation_function = get_evaluation_function(evaluation_function)
output = predictor.predict(test_set.input)
score = evaluation_function(actual = output, target = test_set.target)
return score