def getTheta(self):
if self.use_tensor:
return utils.vectorizeParams(
self.Ws, self.L, self.W, self.V)
else:
return utils.vectorizeParams(
self.Ws, self.L, self.W)
def calculateTotalGradient(self, model):
grad = np.zeros(model.num_parameters)
self.dJ_dWs += len(self.trees_train) * model.lambda_Ws * model.Ws
self.dJ_dL += len(self.trees_train) * model.lambda_L * model.L
self.dJ_dW += len(self.trees_train) * model.lambda_W * model.W
if model.use_tensor:
self.dJ_dV += len(self.trees_train) * model.lambda_V * model.V
grad = utils.vectorizeParams(self.dJ_dWs, self.dJ_dL, self.dJ_dW,
self.dJ_dV)
else:
grad = utils.vectorizeParams(self.dJ_dWs, self.dJ_dL, self.dJ_dW)
return grad
def calculateTotalGradient(self, model):
grad = np.zeros(model.num_parameters)
# add regularizer gradients
self.dJ_dWs += len(self.trees_train) * model.lambda_Ws * model.Ws
self.dJ_dL += len(self.trees_train) * model.lambda_L * model.L
self.dJ_dW += len(self.trees_train) * model.lambda_W * model.W
if model.use_tensor:
self.dJ_dV += len(self.trees_train)*model.lambda_V * model.V
grad = utils.vectorizeParams(
self.dJ_dWs, self.dJ_dL, self.dJ_dW, self.dJ_dV)
else:
grad = utils.vectorizeParams(
self.dJ_dWs, self.dJ_dL, self.dJ_dW)
return grad
def calculateTotalGradient(self, model, scaler):
grad = np.zeros(model.num_parameters)
# average the gradient by dividing the minibatch size
self.dJ_dWs *= scaler
self.dJ_dL *= scaler
self.dJ_dW *= scaler
# add regularizer gradients
self.dJ_dWs += model.lambda_Ws * model.Ws
self.dJ_dL += model.lambda_L * model.L
self.dJ_dW += model.lambda_W * model.W
if model.use_tensor:
self.dJ_dV *= scaler
self.dJ_dV += model.lambda_V * model.V
grad = utils.vectorizeParams(
self.dJ_dWs, self.dJ_dL, self.dJ_dW, self.dJ_dV)
else:
grad = utils.vectorizeParams(
self.dJ_dWs, self.dJ_dL, self.dJ_dW)
return grad
def getTheta(self):
if self.use_tensor:
return utils.vectorizeParams(self.Ws, self.L, self.W, self.V)
else:
return utils.vectorizeParams(self.Ws, self.L, self.W)