Merge pull request #3476 from qingqing01/bp_test

Compare the gradient consistency between GPU and CPU calculations
revert-3824-remove_grad_op_type
qingqing01 8 years ago committed by GitHub
commit c68bfc30a8

@ -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());
}
};

@ -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());
}
};

@ -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)

File diff suppressed because it is too large Load Diff

@ -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()

@ -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()

Loading…
Cancel
Save