use double in unittest.

Author: lcy-seso

paddle/operators/linear_chain_crf_op.cc

@@ -195,8 +195,6 @@ class LinearChainCRFOp : public framework::OperatorWithKernel {
     // is the sequence number in a mini-batch. The dimension set here should be
     // resized to its correct size in the function Compute.
     ctx->SetOutputDim("LogLikelihood", {emission_dims[0], 1});
-
-    ctx->ShareLoD("Emission", /*->*/ "EmissionExps");
   }
 
  protected:
@@ -402,7 +400,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel {
   // operator is determined by its input "EmissionExps".
   framework::DataType IndicateDataType(
       const framework::ExecutionContext& ctx) const override {
-    return framework::ToDataType(ctx.Input<LoDTensor>("EmissionExps")->type());
+    return framework::ToDataType(ctx.Input<LoDTensor>("LogLikelihood")->type());
   }
 };
 
@@ -562,7 +560,9 @@ REGISTER_OP(linear_chain_crf, ops::LinearChainCRFOp, ops::LinearChainCRFOpMaker,
             linear_chain_crf_grad, ops::LinearChainCRFGradOp);
 REGISTER_OP_CPU_KERNEL(
     linear_chain_crf,
-    ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, float>);
+    ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, float>,
+    ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     linear_chain_crf_grad,
-    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, float>);
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, float>,
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, double>);

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

@@ -32,7 +32,7 @@ def __init__(self, seq_start_positions, emission_weights, emission_row_max,
         # alpha is a memo table in dynamic programming to caculate
         # nomalization factor.
         self.alpha = np.zeros(
-            (seq_start_positions[-1], self.tag_num), dtype="float32")
+            (seq_start_positions[-1], self.tag_num), dtype="float64")
         self.log_likelihood = np.zeros((self.seq_num, 1))
 
     def _l1_norm(self, x):
@@ -92,12 +92,12 @@ def set_test_data(self):
         for i in range(SEQ_NUM):
             lod[-1].append(lod[-1][-1] + random.randint(1, MAX_SEQ_LEN))
         emission = np.random.uniform(-1, 1,
-                                     [lod[-1][-1], TAG_NUM]).astype("float32")
+                                     [lod[-1][-1], TAG_NUM]).astype("float64")
         emission_row_max = np.amax(emission, axis=1, keepdims=True)
         emission_exps = np.exp(emission - emission_row_max)
 
         transition = np.random.uniform(-0.5, 0.5,
-                                       [TAG_NUM + 2, TAG_NUM]).astype("float32")
+                                       [TAG_NUM + 2, TAG_NUM]).astype("float64")
         transition_exps = np.exp(transition)
 
         labels = np.random.randint(
@@ -128,17 +128,11 @@ def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(
-            ["Emission", "Transition"],
-            "LogLikelihood",
-            max_relative_error=0.05)
+        self.check_grad(["Emission", "Transition"], "LogLikelihood")
 
     def test_check_grad_ignore_transition(self):
         self.check_grad(
-            ["Emission"],
-            "LogLikelihood",
-            max_relative_error=0.05,
-            no_grad_set=set("Transition"))
+            ["Emission"], "LogLikelihood", no_grad_set=set("Transition"))
 
 
 if __name__ == "__main__":