Polish unit test for xe, generate probablities (#5096)

* Cross Entropy Wrong

* Fix XE

* Polish gradient check for xe

* Fix compile

paddle/operators/cross_entropy_op.cc

@@ -162,6 +162,8 @@ or not. But the output only shares the LoD with input `X`.
 namespace ops = paddle::operators;
 REGISTER_OP(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
             cross_entropy_grad, ops::CrossEntropyGradientOp);
-REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<float>);
+REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<float>,
+                       ops::CrossEntropyOpKernel<double>);
 REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
-                       ops::CrossEntropyGradientOpKernel<float>);
+                       ops::CrossEntropyGradientOpKernel<float>,
+                       ops::CrossEntropyGradientOpKernel<double>);

paddle/operators/cross_entropy_op.cu

@@ -108,6 +108,8 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>,
+                       ops::CrossEntropyOpCUDAKernel<double>);
 REGISTER_OP_GPU_KERNEL(cross_entropy_grad,
-                       ops::CrossEntropyGradientOpCUDAKernel<float>);
+                       ops::CrossEntropyGradientOpCUDAKernel<float>,
+                       ops::CrossEntropyGradientOpCUDAKernel<double>);

paddle/operators/math/cross_entropy.cc

@@ -54,6 +54,7 @@ class CrossEntropyFunctor<platform::CPUPlace, T> {
 };
 
 template class CrossEntropyFunctor<platform::CPUPlace, float>;
+template class CrossEntropyFunctor<platform::CPUPlace, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/cross_entropy.cu

@@ -39,11 +39,36 @@ __device__ __forceinline__ T sum_single_warp(T val) {
   return val;
 }
 
+// CUDA do not support dynamic arrary in template
+// https://stackoverflow.com/questions/20497209
+template <typename T>
+struct SharedMemory {
+  // Ensure that we won't compile any un-specialized types
+  __device__ T* GetPointer() { return NULL; }
+};
+
+template <>
+struct SharedMemory<float> {
+  __device__ float* GetPointer() {
+    extern __shared__ float s_float[];
+    return s_float;
+  }
+};
+
+template <>
+struct SharedMemory<double> {
+  __device__ double* GetPointer() {
+    extern __shared__ double s_double[];
+    return s_double;
+  }
+};
+
 template <typename T>
 __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
                                        const int class_num) {
   int tid = threadIdx.x;
-  extern __shared__ T d_sum[];
+  SharedMemory<T> d_sum_shared;
+  T* d_sum = d_sum_shared.GetPointer();
   d_sum[tid] = 0;
 
   int cur_idx = tid;
@@ -102,6 +127,7 @@ class CrossEntropyFunctor<platform::GPUPlace, T> {
 };
 
 template class CrossEntropyFunctor<platform::GPUPlace, float>;
+template class CrossEntropyFunctor<platform::GPUPlace, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

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

@@ -8,6 +8,15 @@ from paddle.v2.framework.executor import Executor
 from paddle.v2.framework.framework import Program, OpProtoHolder
 
 
+def randomize_probability(batch_size, class_num, dtype='float32'):
+    prob = np.random.uniform(
+        0.1, 1.0, size=(batch_size, class_num)).astype(dtype)
+    prob_sum = prob.sum(axis=1)
+    for i in xrange(len(prob)):
+        prob[i] /= prob_sum[i]
+    return prob
+
+
 def grad_var_name(var_name):
     return var_name + "@GRAD"
 
@@ -379,9 +388,9 @@ class OpTest(unittest.TestCase):
             def err_msg():
                 offset = np.argmax(diff_mat > max_relative_error)
                 return ("%s Variable %s max gradient diff %f over limit %f, "
-                        "the first error element is %d") % (
+                        "the first error element is %d, %f, %f") % (
                             msg_prefix, name, max_diff, max_relative_error,
-                            offset)
+                            offset, a.flatten()[offset], b.flatten()[offset])
 
             self.assertLessEqual(max_diff, max_relative_error, err_msg())
 
@@ -389,6 +398,7 @@ class OpTest(unittest.TestCase):
                    inputs_to_check,
                    output_names,
                    no_grad_set=None,
+                   numeric_grad_delta=0.005,
                    in_place=False,
                    max_relative_error=0.005,
                    user_defined_grads=None):
@@ -411,6 +421,7 @@ class OpTest(unittest.TestCase):
                 self.inputs,
                 input_to_check,
                 output_names,
+                delta=numeric_grad_delta,
                 in_place=in_place) for input_to_check in inputs_to_check
         ]
         grad_names = [

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

 import unittest
 import numpy as np
-from op_test import OpTest
+from op_test import OpTest, randomize_probability
 
 
 class TestCrossEntropyOp1(OpTest):
@@ -12,12 +12,12 @@ class TestCrossEntropyOp1(OpTest):
         batch_size = 30
         class_num = 10
 
-        X = np.random.uniform(0.1, 1.0,
-                              [batch_size, class_num]).astype("float32")
+        X = randomize_probability(batch_size, class_num, dtype='float64')
+
         label = np.random.randint(0, class_num, (batch_size, 1), dtype="int32")
         cross_entropy = np.asmatrix(
             [[-np.log(X[i][label[i][0]])] for i in range(X.shape[0])],
-            dtype="float32")
+            dtype="float64")
 
         self.inputs = {"X": X, "Label": label}
         self.outputs = {"Y": cross_entropy}
@@ -27,7 +27,7 @@ class TestCrossEntropyOp1(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(["X"], "Y")
+        self.check_grad(["X"], "Y", numeric_grad_delta=0.001)
 
 
 class TestCrossEntropyOp2(OpTest):
@@ -39,8 +39,7 @@ class TestCrossEntropyOp2(OpTest):
         batch_size = 5
         class_num = 37
 
-        X = np.random.uniform(0.1, 1.0,
-                              [batch_size, class_num]).astype("float32")
+        X = randomize_probability(batch_size, class_num)
         label = np.random.uniform(0.1, 1.0,
                                   [batch_size, class_num]).astype("float32")
         label /= label.sum(axis=1, keepdims=True)
@@ -55,7 +54,8 @@ class TestCrossEntropyOp2(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(["X"], "Y", max_relative_error=0.05)
+        self.check_grad(
+            ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001)
 
 
 class TestCrossEntropyOp3(OpTest):
@@ -67,8 +67,7 @@ class TestCrossEntropyOp3(OpTest):
         batch_size = 5
         class_num = 17
 
-        X = np.random.uniform(0.1, 1.0,
-                              [batch_size, class_num]).astype("float32")
+        X = randomize_probability(batch_size, class_num)
         label_index = np.random.randint(
             0, class_num, (batch_size), dtype="int32")
         label = np.zeros(X.shape)
@@ -88,7 +87,8 @@ class TestCrossEntropyOp3(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(["X"], "Y", max_relative_error=0.05)
+        self.check_grad(
+            ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001)
 
 
 if __name__ == "__main__":