def conditional_mean(
self, tag: str, x: ep.Tensor, cond: ep.Tensor, step: int
) -> None:
cond_ = cond.numpy()
if ~cond_.any():
return
x_ = x.numpy()
x_ = x_[cond_]
self.writer.add_scalar(tag, x_.mean(axis=0).item(), step)
def test_2d(x2d: Tensor, p: float, axis: int, keepdims: bool) -> None:
assert isinstance(axis, int) # see test4d for the more general test
assert_allclose(
lp(x2d, p, axis=axis, keepdims=keepdims).numpy(),
norm(x2d.numpy(), ord=p, axis=axis, keepdims=keepdims),
rtol=1e-6,
)
if p not in norms:
return
assert_allclose(
norms[p](x2d, axis=axis, keepdims=keepdims).numpy(),
norm(x2d.numpy(), ord=p, axis=axis, keepdims=keepdims),
rtol=1e-6,
)
def clear(self, dims: ep.Tensor):
if self.tensor is None:
self.tensor = dims
dims = dims.numpy()
assert dims.shape == (self.N, )
assert dims.dtype == np.bool
self.data[:, dims] = np.nan
def append(self, x: ep.Tensor):
if self.tensor is None:
self.tensor = x
x = x.numpy()
assert x.shape == (self.N, )
self.data[self.next] = x
self.next = (self.next + 1) % self.maxlen
def histogram(
self, tag: str, x: ep.Tensor, step: int, *, first: bool = True
) -> None:
x = x.numpy()
self.writer.add_histogram(tag, x, step)
if first:
self.writer.add_scalar(tag + "/0", x[0].item(), step)
def test_4d(x4d: Tensor, p: float, axis: Optional[ep.types.AxisAxes],
keepdims: bool) -> None:
actual = lp(x4d, p, axis=axis, keepdims=keepdims).numpy()
# numpy does not support arbitrary axes (limited to vector and matrix norms)
if axis is None:
axes = tuple(range(x4d.ndim))
elif not isinstance(axis, tuple):
axes = (axis, )
else:
axes = axis
del axis
axes = tuple(i % x4d.ndim for i in axes)
x = x4d.numpy()
other = tuple(i for i in range(x.ndim) if i not in axes)
x = np.transpose(x, other + axes)
x = x.reshape(x.shape[:len(other)] + (-1, ))
desired = norm(x, ord=p, axis=-1)
if keepdims:
shape = tuple(1 if i in axes else x4d.shape[i]
for i in range(x4d.ndim))
desired = desired.reshape(shape)
assert_allclose(actual, desired, rtol=1e-6)
def test_numpy_inplace(t: Tensor) -> None:
copy = t + 0
a = t.numpy().copy()
a[:] += 1
assert (t == copy).all()
def test_numpy_readonly(t: Tensor) -> None:
a = t.numpy()
assert a.flags.writeable is False
with pytest.raises(ValueError, match="read-only"):
a[:] += 1
def test_1d(x1d: Tensor, p: float) -> None:
assert_allclose(lp(x1d, p).numpy(), norm(x1d.numpy(), ord=p))
assert_allclose(norms[p](x1d).numpy(), norm(x1d.numpy(), ord=p))