Clean code and update doc.

paddle/operators/lstm_op.cc

@@ -126,11 +126,11 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
              " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.")
         .AsDispensable();
     AddOutput("Hidden",
-              "(LoDTensor) the hidden state lod tensor of LSTM operator. "
-              "The shape and lod is the same with the `Input`.");
+              "(LoDTensor) the hidden state of LSTM operator. "
+              "The shape is (T x D), and lod is the same with the `Input`.");
     AddOutput("Cell",
-              "(LoDTensor) the cell state lod tensor of LSTM operator. "
-              "The shape and lod is the same with the `Input`.");
+              "(LoDTensor) the cell state of LSTM operator. "
+              "The shape is (T x D), and lod is the same with the `Input`.");
     AddOutput("BatchGate",
               "(LoDTensor) This LoDTensor contains input gate, forget gate "
               "and output gate after the nonlinear computation. This "
@@ -141,7 +141,7 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
               "in the raw input.")
         .AsIntermediate();
     AddOutput("BatchCellPreAct",
-              "(LoDTensor) This LoDTensor is get in the forward and used "
+              "(LoDTensor) This LoDTensor is got in the forward and used "
               "in the backward.")
         .AsIntermediate();
     AddAttr<bool>("usePeepholes",

paddle/operators/lstm_op.h

@@ -155,7 +155,6 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     auto* batch_cell_pre_act = ctx.Input<LoDTensor>("BatchCellPreAct");
 
     auto* hidden_g = ctx.Input<LoDTensor>(framework::GradVarName("Hidden"));
-    // auto* cell_g = ctx.Input<LoDTensor>(framework::GradVarName("Cell"));
 
     auto* in_g = ctx.Output<LoDTensor>(framework::GradVarName("Input"));
     auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
@@ -251,7 +250,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       lstm_grad.gateGrad = gate_g.data<T>();
       lstm_grad.outputGrad = out_g.data<T>();
 
-      if (n != 0) {
+      if (n) {
         int bstart_pre = static_cast<int>(batch_starts[n - 1]);
         Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
         Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
@@ -292,17 +291,6 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     }
     if (bias && bias_g) {
       /* backward bias */
-      // Following Eigen computation failed for double type on GPU device.
-      // bias_g->mutable_data<T>(ctx.GetPlace());
-      // Tensor bias_mat;
-      // bias_mat.ShareDataWith(*bias_g);
-      // bias_mat.Resize({1, 4 * frame_size});
-
-      // auto bias_g_e = EigenVector<T>::Flatten(bias_mat);
-      // auto gate_g_e = EigenMatrix<T>::From(batch_gate_g);
-      // Eigen::array<int, 1> dims{{0}};
-      // bias_g_e.device(ctx.GetEigenDevice<Place>()) = gate_g_e.sum(dims);
-
       int m = static_cast<int>(batch_gate_g.dims()[0]);
       int n = static_cast<int>(batch_gate_g.dims()[1]);
 

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

@@ -157,15 +157,17 @@ class TestLstmOp(OpTest):
         }
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(atol=1e-8)
 
     #TODO(qingqing) add more unit testing case
     def test_check_grad(self):
-        # TODO(qingqing) remove folowing two lines after the check_grad is refined.
-        self.outputs['BatchGate'] = None
-        self.outputs['BatchCellPreAct'] = None
+        # TODO(qingqing) remove folowing lines after the check_grad is refined.
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
         self.check_grad(
-            ['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=0.02)
+            ['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=5e-4)
 
 
 class TestLstmOpHasNoInitial(TestLstmOp):