def get_importances(self, target_examples, update=False):
print('\t in get importances ...')
api = VWAPI()
print('\t getting targets ...')
#utils.debug_print(str(target_examples))
# assume target_examples are in perserved index order
examples = [example['meta']['features'] for example in target_examples]
print('\t calling my battle heavy get_bulk_responses...')
answers = api.get_bulk_responses(examples)
#utils.debug_print(str(answers))
print('\t closing/shutting down api')
api.vw.close() # del doesn't seemt to close socket :-/
del api
print('\t returning importances, apply np.argsort')
importances = [answer.importance for answer in answers]
ordered_importances = np.argsort(importances)
if update:
butler.algorithms.set(key='importances', value=ordered_importances)
# ordered by importance on indices into target_examples
return ordered_importances
def teach(self, butler, target_index, target_label):
api = VWAPI()
example = butler.targets.get_target_item(butler.exp_uid, target_index)
vw_example = api.to_vw_examples([example])
api.vw.send_example(target_label, features=vw_example)
api.vw.close()
del api
def getModel(self, butler):
precision = random.random()
num_reported_answers = butler.experiment.get(
key='num_reported_answers')
butler.algorithms.set(key='precision', value=precision)
ret = butler.algorithms.get(key=['precision', 'num_reported_answers'])
#print ret
num_reported_answers = ret[
'num_reported_answers'] # to prove that we can make it non None
print('num reported', num_reported_answers, '<')
print type(num_reported_answers), type(precision)
# Debug area for getting hold out accuracy
hold_out = None
if butler.experiment.get(key='hold_out'):
hold_out = set(butler.experiment.get(key='hold_out'))
print('\t ***', hold_out, ' is hold out')
if hold_out: # during inital queries it can be null beause nothing was held out
print('\t *** len of hold out: ', len(hold_out)) # have length
# so here we have a hold out and we need to test with it
hold_out_features = [butler.targets.get_target_item(butler.exp_uid, value[0])['meta']['features']\
for value in hold_out]
print(len(hold_out_features))
api = VWAPI()
answers = api.get_bulk_responses(hold_out_features)
api.vw.close() # del doesn't seemt to close socket :-/
del api
# Here we compare signs for the predicted answer and the held example
# a cheap way to threshold the linear regression to a categorical variable
scores = [(answer.prediction < 0) == (held_out_example[1] < 0)
for answer, held_out_example in zip(answers, hold_out)]
precision = sum(scores) / len(scores)
print(scores, precision)
# this return is identical to get()
#return butler.algorithms.get(key=['precision', 'num_reported_answers'])
return {
'precision': precision,
'num_reported_answers': num_reported_answers
}
def get_importances(self, butler=None, target_examples=None, update=False):
print('\t in get importances ...')
api = VWAPI()
print('\t getting targets ...')
# assume target_examples are in perserved index order
examples = [example['meta']['features'] for example in target_examples]
answers = api.get_bulk_responses(examples)
print('\t closing/shutting down api')
api.vw.close() # del doesn't seemt to close socket :-/
del api
importances = [answer.importance for answer in answers]
ordered_importances = np.argsort(importances)
if update:
butler.algorithms.set(key='importances', value=ordered_importances)
# ordered by importance on indices into target_examples
return ordered_importances
def teach(self, butler, args):
#print(args)
args = json.loads(args)
target_label = args['args']['target_label']
example = args['args']['example']
print('\t*** about to teach ...')
api = VWAPI()
# products are represented as a bag of words (and maybe syntax too)
vw_example = api.to_vw_examples([example])
#print("\t*** vw_example: ", str(vw_example))
api.vw.add_namespaces(vw_example)
# important to send namespaces, not features, because
# to_vw_examples creates namespaces
api.vw.send_example(response=target_label)
api.vw.close()
del api
print('\t*** ... taught')