Merge pull request #3476 from qingqing01/bp_test

Compare the gradient consistency between GPU and CPU calculations

paddle/operators/sigmoid_op.cc

@@ -44,7 +44,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>(framework::GradVarName("X"))
+        ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
 

paddle/operators/sigmoid_op.h

@@ -37,7 +37,7 @@ class SigmoidKernel : public framework::OpKernel {
     auto Y = EigenVector<T>::Flatten(*output);
     auto place = context.GetEigenDevice<Place>();
 
-    Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp());
+    Y.device(place) = 1. / (1. + (-X).exp());
   }
 };
 

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -25,3 +25,4 @@ py_test(test_operator SRCS test_operator.py)
 # py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
 py_test(test_recurrent_op SRCS test_recurrent_op.py)
+py_test(test_gradient_checker SRCS test_gradient_checker.py)

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

@@ -0,0 +1,43 @@
+import unittest
+import numpy
+from paddle.v2.framework.op import Operator
+from gradient_checker import GradientChecker
+from gradient_checker import get_numeric_gradient
+
+
+class GetNumericGradientTest(unittest.TestCase):
+    def test_add_op(self):
+        add_op = Operator('add_two', X="X", Y="Y", Out="Z")
+        x = numpy.random.random((10, 1)).astype("float32")
+        y = numpy.random.random((10, 1)).astype("float32")
+
+        arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
+        self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
+
+    def test_softmax_op(self):
+        def stable_softmax(x):
+            """Compute the softmax of vector x in a numerically stable way."""
+            shiftx = x - numpy.max(x)
+            exps = numpy.exp(shiftx)
+            return exps / numpy.sum(exps)
+
+        def label_softmax_grad(Y, dY):
+            dX = Y * 0.0
+            for i in range(Y.shape[0]):
+                d = numpy.dot(Y[i, :], dY[i, :])
+                dX[i, :] = Y[i, :] * (dY[i, :] - d)
+            return dX
+
+        softmax_op = Operator("softmax", X="X", Y="Y")
+
+        X = numpy.random.random((2, 2)).astype("float32")
+        Y = numpy.apply_along_axis(stable_softmax, 1, X)
+        dY = numpy.ones(Y.shape)
+        dX = label_softmax_grad(Y, dY)
+
+        arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
+        numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -1,6 +1,7 @@
 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 
 
 class TestSigmoidOp(unittest.TestCase):
@@ -8,12 +9,20 @@ class TestSigmoidOp(unittest.TestCase):
 
     def setUp(self):
         self.type = "sigmoid"
-        self.inputs = {'X': np.random.random((32, 100)).astype("float32")}
+        self.inputs = {'X': np.random.random((15, 31)).astype("float32")}
         self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
 
 
-#class TestSigmoidGradOp(unittest.TestCase):
-#TODO(qingqing) add unit test
+class TestSigmoidGradOp(GradientChecker):
+    def test_grad(self):
+        op = create_op("sigmoid")
+        inputs = {"X": np.random.uniform(0.1, 1, [11, 17]).astype("float32")}
+        # compare gpu and cpu results for backward op.
+        # this test will be skiped if only compiling CPU version.
+        self.compare_grad(op, inputs)
+        # check gradients 
+        self.check_grad(op, inputs, set("X"), "Y", max_relative_error=0.007)
+
 
 if __name__ == '__main__':
     unittest.main()