Merge remote-tracking branch 'origin/rnn-backward-python' into feature/recurrent_op_backward_fix

paddle/operators/recurrent_op.cc

@@ -80,7 +80,6 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
   // Now all variables in scope must be created outside of op.
   PADDLE_ENFORCE_NOT_NULL(stepnet_);
   PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs");
-  PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs");
 
   if (seq_len_ > step_scopes->size()) {
     for (size_t i = step_scopes->size(); i < seq_len_; ++i) {

paddle/pybind/pybind.cc

@@ -311,6 +311,15 @@ All parameter, weight, gradient are variables in Paddle.
              self.set_falsenet(net.Clone());
            });
 
+  rnn.def("backward",
+          [](const operators::RecurrentOp &forwardOp,
+             const std::unordered_set<std::string> &no_grad_vars) {
+            const auto &op = *static_cast<const OperatorBase *>(&forwardOp);
+            return Backward(op, no_grad_vars);
+          });
+
+  ExposeOperator(rnn);
+
   m.def("unique_integer", UniqueIntegerGenerator);
 
   m.def("is_compile_gpu", IsCompileGPU);

python/paddle/v2/framework/tests/test_recurrent_op.py

@@ -3,6 +3,7 @@ import paddle.v2.framework.core as core
 import unittest
 import numpy as np
 from paddle.v2.framework.op import Operator, RecurrentOp
+from gradient_checker import GradientChecker
 
 
 def py_sigmoid(x):
@@ -68,7 +69,7 @@ def create_tensor(scope, name, shape, np_data):
     return tensor
 
 
-class TestRecurrentOp(unittest.TestCase):
+class RecurrentOpTest(unittest.TestCase):
     '''
     Test RNNOp
 
@@ -160,5 +161,42 @@ class TestRecurrentOp(unittest.TestCase):
         self.assertEqual(pd_output.shape, py_output.shape)
 
 
+class RecurrentGradientOpTest(unittest.TestCase):
+    def create_forward_op(self):
+        self.forward_op = RecurrentOp(
+            # inputs
+            inlinks=["x"],
+            boot_memories=["h_boot"],
+            step_net="stepnet",
+            # outputs
+            outlinks=["h"],
+            step_scopes="step_scopes",
+            # attributes
+            inlink_alias=["x@alias"],
+            outlink_alias=["h@alias"],
+            pre_memories=["h@pre"],
+            memories=["h@alias"])
+
+        # create a stepnet for RNN
+        stepnet = core.Net.create()
+        x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
+        h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
+        sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
+        sig_op = Operator("sigmoid", X="sum", Y="h@alias")
+
+        for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
+            stepnet.add_op(op)
+        stepnet.complete_add_op(True)
+        self.forward_op.set_stepnet(stepnet)
+
+    def create_gradient_op(self):
+        a = set()
+        backward_op = core.RecurrentOp.backward(self.forward_op, a)
+
+    def test_grad(self):
+        self.create_forward_op()
+        self.create_gradient_op()
+
+
 if __name__ == '__main__':
     unittest.main()