def testLrnGrad(self):
# Test for LRNGrad that compares against the CPU implementation.
shape = [1, 2, 3, 4]
total_size = np.prod(shape)
in_image_vals = np.arange(1, total_size + 1, dtype=np.float32)
out_image_vals = np.arange(1, total_size + 1, dtype=np.float32)
out_grads_vals = np.arange(1, total_size + 1, dtype=np.float32)
depth_radius = np.random.randint(1, shape[3])
bias = 1.0 + np.random.rand()
alpha = 1.0 * np.random.rand()
beta = 1.0 * np.random.rand()
with self.test_session():
in_image = constant_op.constant(in_image_vals, shape=shape)
out_image = constant_op.constant(out_image_vals, shape=shape)
out_grads = constant_op.constant(out_grads_vals, shape=shape)
with ops.device(CPU_DEVICE):
expected = gen_nn_ops._lrn_grad(out_grads, in_image, out_image,
depth_radius, bias, alpha, beta)
with self.test_scope():
actual = gen_nn_ops._lrn_grad(out_grads, in_image, out_image,
depth_radius, bias, alpha, beta)
expected_val = expected.eval()
actual_val = actual.eval()
self.assertAllClose(actual_val, expected_val, rtol=1e-3)
def _LRNGrad(op, grad):
depth_radius = op.get_attr("depth_radius")
bias = op.get_attr("bias")
alpha = op.get_attr("alpha")
beta = op.get_attr("beta")
return [gen_nn_ops._lrn_grad(grad, op.inputs[0], op.outputs[0], depth_radius,
bias, alpha, beta)]
