def main(): DB_MODE = settings.DB_MODE EMF_MODE = settings.EMF_MODE hndl_Log = EMF_Logging_Handle(mode=EMF_MODE) hndl_DB = get_DB_Handle(EMF_mode=EMF_MODE, DB_mode=DB_MODE) hndl_WrdSlct = EMF_WordSelector_Handle(hndl_DB) hndl_WrdSlct.pred_trns_ptrns = PredictorTransformationKeys hndl_WrdSlct.resp_trns_ptrns = ResponseTransformationKeys hndl_WrdSlct.resp_trns_ptrns = ['None', 'Futr_Lvl', 'Futr_Change'] hndl_WrdSlct.resp_data_tickers = ['SP500_RealPrice'] hndl_WordSet = EMF_WordSet_Handle(hndl_DB) hndl_WordSet.resp_words = hndl_WrdSlct.select_resp_words_random() hndl_WordSet.pred_words = hndl_WrdSlct.select_pred_words_random()
def __init__(self, hndl_DB): self._hndl_DB = hndl_DB self.hndl_WordSet = EMF_WordSet_Handle(self._hndl_DB) self.hndl_WrdSlct = EMF_WordSelector_Handle(self._hndl_DB) self.hndl_Res = EMF_Results_Handle() self.models = []
class EMF_Model_Runner(object):
def __init__(self, hndl_DB):
self._hndl_DB = hndl_DB
self.hndl_WordSet = EMF_WordSet_Handle(self._hndl_DB)
self.hndl_WrdSlct = EMF_WordSelector_Handle(self._hndl_DB)
self.hndl_Res = EMF_Results_Handle()
self.models = []
def set_model_from_template(self, template):
# Set Response Ticker in WordSelector
self.hndl_WrdSlct.resp_data_tickers = template['responseTicker']
if 'responseTrns' in template:
self.hndl_WrdSlct.resp_trns_ptrns = template['responseTrns']
else:
self.hndl_WrdSlct.resp_trns_ptrns = ResponseTransformationKeys
if 'responseKwargs' in template:
self.hndl_WrdSlct.resp_trns_kwargs = template['responseKwargs']
if 'responseCanPredict' in template:
self.hndl_WrdSlct.resp_can_predict = template['responseCanPredict']
if 'predictorKwargs' in template:
self.hndl_WrdSlct.pred_trns_kwargs = template['predictorKwargs']
if 'predictorCriteria' in template:
predCrit = template['predictorCriteria']
if 'periodicity' in predCrit:
self.hndl_WrdSlct.pred_data_periodicity = predCrit['periodicity']
if 'categorical' in predCrit:
self.hndl_WrdSlct.pred_data_is_categorical = predCrit['categorical']
if 'minDate' in predCrit:
self.hndl_WrdSlct.pred_data_min_date = predCrit['minDate']
if 'maxDate' in predCrit:
self.hndl_WrdSlct.pred_data_max_date = predCrit['maxDate']
if 'matchRespPeriodicity' in predCrit:
raise NotImplementedError
if 'predictorTrns' in template:
self.hndl_WrdSlct.pred_trns_ptrns = template['predictorTrns']
else:
self.hndl_WrdSlct.pred_trns_ptrns = PredictorTransformationKeys
# Set Models
for model in template['models']:
self.models.append(AvailableModels[model])
def train_model_batch(self):
'''
'''
# Choose Model at Random
self.hndl_Model = choice(self.models)(self.hndl_WordSet)
# Choose Response Word at Random and Add Response Word to WordSet
self.hndl_WordSet.resp_words = self.hndl_WrdSlct.select_resp_words_random()
# Add Response Word to Result Handle
self.hndl_Res.set_response_words(self.hndl_WordSet.resp_words)
# Run Batches
for i in xrange(randint(MIN_BATCH_SIZE, MAX_BATCH_SIZE)):
# Add Predictive Words to WordSet
# self.hndl_WordSet.pred_words = self.hndl_WrdSlct.select_pred_words_random()
self.hndl_WordSet.pred_words = self.hndl_WrdSlct.select_pred_words_effectiveness()
# Make Sure WordSet is Prepared
self.__evaluate_run_readiness()
# Run Model
(pred_array, resp_array) = self.hndl_WordSet.get_word_arrays(mode=TRAINING, bootstrap=True)
for (j, hndl_Word) in enumerate(self.hndl_WordSet.resp_words):
self.hndl_Model.train_model(pred_array, resp_array[:,j])
self.__save_model_results(hndl_Word, j)
# Prepare Predictive Words for Next Run (Safety First)
del self.hndl_WrdSlct.pred_words
del self.hndl_WordSet.pred_words
# Prepare Response Words and Model for Next Run (Safety First)
del self.hndl_Model
del self.hndl_WrdSlct.resp_words
del self.hndl_WrdSlct.resp_words
return self.hndl_Res
def __save_model_results(self, hndl_Word, idx):
'''
TODO:
Remove Spaghetti code of pushing around idx (create run_model instance)
'''
self.hndl_WordSet.save_predictions(self.hndl_Model, hndl_Word)
# self.hndl_WordSet.plot_values(hndl_Word) # TEST: DELETE
score = self.__evaluate_model(hndl_Word, idx)
if score >= MODEL_RETENTION_THRESHHOLD:
log.info('MODEL: Model accepted with score {0}'.format(score))
(dates, values) = self.hndl_WordSet.get_prediction_series(hndl_Word)
self.hndl_Res.add_model(self.hndl_Model, dates, values, score)
self.__save_model_attributes(hndl_Word)
else:
log.info('MODEL: Model rejected with score {0}'.format(score))
# Put Word Statistics In DB
self.hndl_WrdSlct.save_word_statistics(score*self.hndl_Model.adjusted_feat_scores, hndl_Word)
# Put Data Statistics In DB
self.hndl_WrdSlct.save_data_statistics(score*self.hndl_Model.adjusted_feat_scores, hndl_Word)
# Put Model Statistics In DB
# Not Implemented
def __save_model_metadata(self, hndl_Word, idx):
'''
We save metadata for use in content attribute selections
'''
# Linear Series Attributes
(smoothed, pred, raw, res) = self.hndl_WordSet.get_residual_data(hndl_Word)
# Linear Series Attributes / Mean, Std (for various time periods)
# Linear Series Attributes / Biggest Jumps (Top 1, 2, 3)
# Linear Series Attributes / Start Time, End Time
# Linear Series Attributes / Model Attributes
# Linear Series Attributes / Categories, Geographies, etc.
raise NotImplementedError
def __save_model_attributes(self, hndl_Word):
self.hndl_Model.save_model_attributes(self.hndl_WordSet.pred_words, hndl_Word)
def __evaluate_model(self, hndl_Word, idx):
'''
TODO:
Remove Spaghetti code of pushing around idx (create run_model instance)
'''
(smoothed, pred, raw, res) = self.hndl_WordSet.get_residual_data(hndl_Word)
(pred_array, resp_array) = self.hndl_WordSet.get_word_arrays(mode=TRAINING, bootstrap=False)
score = self.hndl_Model.test_model(pred_array, resp_array[:,idx]) # Need this for word statistics
# Compare Modeled Levels
score_3 = abs(utl_RnM.level_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_three)-.003)
score_2 = abs(utl_RnM.level_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_two)-.05)
score_1 = abs(utl_RnM.level_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_one)-.32)
score -= 2*(score_3+score_2+score_1)
# Compare Modeled Changes
score_3 = abs(utl_RnM.change_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_three)-.003)
score_2 = abs(utl_RnM.change_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_two)-.05)
score_1 = abs(utl_RnM.change_outlier_proportion(smoothed, raw, check_fn=utl_RnM.more_than_one)-.32)
score -= 1*(score_3+score_2+score_1)
# Prefer Smooth Model Levels
score_3 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_three)-.003)
score_2 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_two)-.05)
score_1 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_one)-.32)
score -= .3*(score_3+score_2+score_1)
# Prefer Smooth Model Changes
score_3 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_three)-.003)
score_2 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_two)-.05)
score_1 = abs(utl_RnM.level_outlier_proportion(smoothed, pred, check_fn=utl_RnM.more_than_one)-.32)
score -= .7*(score_3+score_2+score_1)
return score
def __evaluate_run_readiness(self):
'''
Checks to make sure the model is compatible with the chosen words
'''
if not self.hndl_WordSet.predictor_words_are_set():
raise AttributeError
if not self.hndl_WordSet.response_word_is_set():
raise AttributeError
for type_ in self.hndl_WordSet.get_response_word_types():
if type_ not in self.hndl_Model.allowed_respVar_types:
raise AttributeError
for type_ in self.hndl_WordSet.get_predictor_word_types():
if type_ not in self.hndl_Model.allowed_predVar_types:
raise AttributeError
