Design of IfThenElse operator and layer #3119
Description
After talking with @zchen0211 and @Canpio , here is the very early draft of the idea of IfThenElse operator and layer:
A Simple Example
The Python Application Program
cost = paddle.layer.mse( paddel.layer.fc( paddle.layer.data(name="features"), output_size = 100), paddle.layer.data(name="label")) parameters = paddle.train(cost, minist_reader(), {"features":0, "label":1}) print paddle.inf(cost, parameters, inference_input_reader())FC Layer as a Python Function
def fc(input): output = paddle.framework.tensor() W = paddle.framework.tensor(label="parameter") b = paddle.framework.tensor(label="parameter") paddle.operator.fc(input, W, b, output) return outputFC Operator as a C++ Class
class FC : public framework:Operator {...};The IfThenElse Layer and Operator
Another Example Using IfThenElse Layer
def a_network(input): A = paddle.layer.fc(input, parameter_name="A") B = paddle.layer.fc(input, parameter_name="B") return paddle.layer.if_then_else( paddle.layer.less_than(A, B), paddle.layer.fc( paddle.layer.fc( paddle.layer.fc paddle.layer.fc(A, output_size = 100))), paddle.layer.fc( paddle.layer.fc( paddle.layer.fc(B, output_size = 100))))The IfThenElse Layer
def if_then_else(cond, left, right) : ret = paddle.framework.tensor() paddle.operator.if_then_else(cond, left, right, ret) return retThe IfThenElseOp C++ Class
class IfThenElseOp : public framework::Operator { // Check that cond returns a minibatch of boolean values. // Check that left and right outputs are of the same size. };