def _testGradient(self, x, use_gpu=False, **kwargs):
with self.cached_session(use_gpu=use_gpu):
x = ops.convert_to_tensor(x, dtype=dtypes.float64)
grad_naive_theoretical, _ = gradient_checker_v2.compute_gradient(
lambda y: math_ops.cumsum(math_ops.exp(y), **kwargs), [x])
grad_fused_theoretical, _ = gradient_checker_v2.compute_gradient(
lambda y: math_ops.exp(math_ops.cumulative_logsumexp(y, **kwargs)),
[x])
self.assertAllClose(grad_fused_theoretical, grad_naive_theoretical)
def test1DLarge(self):
# This test ensures that the operation is correct even when the naive
# implementation would overflow.
x_np = np.arange(20) * 20.0
for use_gpu in (True, False):
with self.cached_session(use_gpu=use_gpu):
x_tf = ops.convert_to_tensor(x_np, dtype=dtypes.float32)
result_fused = self.evaluate(math_ops.cumulative_logsumexp(x_tf))
result_map = self.evaluate(self._logSumExpMap(x_tf))
self.assertAllClose(result_fused, result_map)
def _computeLogSumExp(self, x, **kwargs): result_naive = math_ops.cumsum(math_ops.exp(x), **kwargs) result_fused = math_ops.exp(math_ops.cumulative_logsumexp(x, **kwargs)) return result_naive, result_fused