def _run_step(self, model, train_data, validate_data, feature_mask,
n_model, model_mask):
if self.direction == 'forward' and self.n_step == 0:
if self.role == consts.GUEST:
loss, ic_val = self.get_ic_val_guest(model, train_data)
LOGGER.info("step {} n_model {}: ic_val {}".format(
self.n_step, n_model, ic_val))
model_info = ModelInfo(self.n_step, n_model, ic_val, loss,
self.step_direction)
self.models_trained[model_mask] = model_info
model_key = model_info.get_key()
self._put_model(model_key, model)
else:
model_info = ModelInfo(self.n_step, n_model, None, None,
self.step_direction)
self.models_trained[model_mask] = model_info
model_key = model_info.get_key()
self._put_model(model_key, model)
return
curr_step = Step()
curr_step.set_step_info((self.n_step, n_model))
trained_model = curr_step.run(model, train_data, validate_data,
feature_mask)
loss, ic_val = self.get_ic_val(trained_model, model_mask)
LOGGER.info("step {} n_model {}: ic_val {}".format(
self.n_step, n_model, ic_val))
model_info = ModelInfo(self.n_step, n_model, ic_val, loss,
self.step_direction)
self.models_trained[model_mask] = model_info
model_key = model_info.get_key()
self._put_model(model_key, trained_model)
def predict(data_instances, model):
if data_instances is None:
return
d_header = data_instances.schema.get("header")
best_feature = [d_header.index(x) for x in model.header]
best_mask = np.zeros(len(d_header), dtype=bool)
np.put(best_mask, best_feature, 1)
new_data = data_instances.mapValues(lambda v: Step.slice_data_instance(v, best_mask))
pred_result = model.predict(new_data)
return pred_result