def __call__(self, inp: to.Tensor) -> to.Tensor:
"""
Evaluate the features and normalize them.
.. note::
Only processing of 1-dim input (e.g., no images)! The input can be batched along the first dimension.
:param inp: input i.e. observations in the RL setting
:return: 1-dim vector of all feature values given the observations
"""
if inp.ndimension() > 2:
raise pyrado.ShapeErr(
msg='RBF class can only handle 1-dim or 2-dim input!')
inp = atleast_2D(
inp
) # first dim is the batch size, the second dim it the actual input dimension
inp = inp.reshape(inp.shape[0], 1,
inp.shape[1]).repeat(1, self.centers.shape[0],
1) # reshape explicitly
exp_sq_dist = to.exp(-self.scale * to.pow(inp - self.centers, 2))
feat_val = to.empty(inp.shape[0], self.num_feat)
for i, sample in enumerate(exp_sq_dist):
if self._state_wise_norm:
# Normalize the features such that the activation for every state dimension sums up to one
feat_val[i, :] = normalize(sample, axis=0,
order=1).t().reshape(-1, )
else:
# Turn the features into a vector and normalize over all of them
feat_val[i, :] = normalize(sample.t().reshape(-1, ),
axis=-1,
order=1)
return feat_val
def _normalize_and_reshape(self, inp: to.Tensor) -> to.Tensor:
"""
Normalize (depending on `state_wise_norm`) and reshape the input.
:param inp: input tensor of exponentiated squared distances
:return: feature value
"""
if self._state_wise_norm:
# Normalize the features such that the activation for every state dimension sums up to one
return normalize(inp, axis=0, order=1).t().reshape(-1)
else:
# Turn the features into a vector and normalize over all of them
return normalize(inp.t().reshape(-1), axis=-1, order=1)
def test_normalize(dtype, axis):
for _ in range(10):
x = to.rand(5, 3) if dtype == 'torch' else np.random.rand(5, 3)
x_norm = normalize(x, axis=axis, order=1)
if isinstance(x_norm, to.Tensor):
x_norm = x_norm.numpy() # for easier checking with pytest.approx
assert np.sum(x_norm, axis=axis) == pytest.approx(1.)
def cosine_similarity(x: to.Tensor, y: to.Tensor) -> to.Tensor:
r"""
Compute the cosine similarity between two tensors $D_cos(x,y) = \frac{x^T y}{|x| \cdot |y|}$.
:param x: input tensor
:param y: input tensor
:return: cosine similarity value
"""
if not isinstance(x, to.Tensor):
raise pyrado.TypeErr(given=x, expected_type=to.Tensor)
if not isinstance(y, to.Tensor):
raise pyrado.TypeErr(given=y, expected_type=to.Tensor)
x_normed = normalize(x, order=2)
y_normed = x_normed if y is x else normalize(y, order=2)
return x_normed.dot(y_normed)
def derivative(self, inp: to.Tensor) -> to.Tensor:
"""
Compute the derivative of the features w.r.t. the inputs.
.. note::
Only processing of 1-dim input (e.g., no images)! The input can be batched along the first dimension.
:param inp: input i.e. observations in the RL setting
:return: value of all features derivatives given the observations
"""
if inp.ndimension() > 2:
raise pyrado.ShapeErr(
msg="RBF class can only handle 1-dim or 2-dim input!")
inp = to.atleast_2d(
inp
) # first dim is the batch size, the second dim it the actual input dimension
inp = inp.reshape(inp.shape[0], 1,
inp.shape[1]).repeat(1, self.centers.shape[0],
1) # reshape explicitly
exp_sq_dist = to.exp(-self.scale * to.pow(inp - self.centers, 2))
exp_sq_dist_d = -2 * self.scale * (inp - self.centers)
feat_val = to.empty(inp.shape[0], self.num_feat)
feat_val_dot = to.empty(inp.shape[0], self.num_feat)
for i, (sample, sample_d) in enumerate(zip(exp_sq_dist,
exp_sq_dist_d)):
if self._state_wise_norm:
# Normalize the features such that the activation for every state dimension sums up to one
feat_val[i, :] = normalize(sample, axis=0, order=1).reshape(-1)
else:
# Turn the features into a vector and normalize over all of them
feat_val[i, :] = normalize(
sample.t().reshape(-1),
axis=-1,
order=1,
)
feat_val_dot[
i, :] = sample_d.reshape(-1) * feat_val[i, :] - feat_val[
i, :] * sum(sample_d.reshape(-1) * feat_val[i, :])
return feat_val_dot