add negative clipping for softmax.

paddle/operators/math/softmax.h

@@ -25,6 +25,14 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
+template <typename T>
+struct ValueClip {
+  HOSTDEVICE T operator()(const T& x) const {
+    const T kThreshold = -64.;
+    return x < kThreshold ? kThreshold : x;
+  }
+};
+
 template <typename Place, typename T>
 class SoftmaxFunctor {
  public:
@@ -47,7 +55,8 @@ class SoftmaxFunctor {
                            logits.maximum(along_class)
                                .eval()
                                .reshape(batch_by_one)
-                               .broadcast(one_by_class));
+                               .broadcast(one_by_class))
+                              .unaryExpr(ValueClip<T>());
 
     softmax.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
     softmax.device(context.GetEigenDevice<Place>()) =

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

@@ -5,7 +5,7 @@ from op_test import OpTest
 
 def stable_softmax(x):
     """Compute the softmax of vector x in a numerically stable way."""
-    shiftx = x - np.max(x)
+    shiftx = x - np.max(x).clip(-64.)
     exps = np.exp(shiftx)
     return exps / np.sum(exps)