garage.misc.tensor_utils module

Utiliy functions for tensors.

concat_tensor_dict_list(tensor_dict_list)[source]

Concatenate dictionary of list of tensor.

Parameters:tensor_dict_list (dict[list]) – a list of dictionaries of {tensors or dictionary of tensors}.
Returns:
a dictionary of {stacked tensors or dictionary of
stacked tensors}
Return type:dict
discount_cumsum(x, discount)[source]

Discounted cumulative sum.

See https://docs.scipy.org/doc/scipy/reference/tutorial/signal.html#difference-equation-filtering # noqa: E501 Here, we have y[t] - discount*y[t+1] = x[t] or rev(y)[t] - discount*rev(y)[t-1] = rev(x)[t]

Parameters:
  • x (np.ndarrary) – Input.
  • discount (float) – Discount factor.
Returns:

Discounted cumulative sum.
Return type:

np.ndarrary
explained_variance_1d(ypred, y, valids=None)[source]

Explained variation for 1D inputs.

It is the proportion of the variance in one variable that is explained or predicted from another variable.

Parameters:
  • ypred (np.ndarray) – Sample data from the first variable. Shape: \((N, max_path_length)\).
  • y (np.ndarray) – Sample data from the second variable. Shape: \((N, max_path_length)\).
  • valids (np.ndarray) – Optional argument. Array indicating valid indices. If None, it assumes the entire input array are valid. Shape: \((N, max_path_length)\).
Returns:

The explained variance.
Return type:

float
flatten_tensors(tensors)[source]

Flatten a list of tensors.

Parameters:tensors (list[numpy.ndarray]) – List of tensors to be flattened.
Returns:Flattened tensors.
Return type:numpy.ndarray
normalize_pixel_batch(observations)[source]

Normalize the observations (images).

Normalize pixel values to be between [0, 1].

Parameters:observations (numpy.ndarray) – Observations from environment. obses should be unflattened and should contain pixel values.
Returns:Normalized observations.
Return type:numpy.ndarray
pad_tensor(x, max_len, mode='zero')[source]

Pad tensors.

Parameters:
  • x (numpy.ndarray) – Tensors to be padded.
  • max_len (int) – Maximum length.
  • mode (str) – If ‘last’, pad with the last element, otherwise pad with 0.
Returns:

Padded tensor.
Return type:

numpy.ndarray
pad_tensor_dict(tensor_dict, max_len, mode='zero')[source]

Pad dictionary of tensors.

Parameters:
  • tensor_dict (dict[numpy.ndarray]) – Tensors to be padded.
  • max_len (int) – Maximum length.
  • mode (str) – If ‘last’, pad with the last element, otherwise pad with 0.
Returns:

Padded tensor.
Return type:

dict[numpy.ndarray]
pad_tensor_n(xs, max_len)[source]

Pad array of tensors.

Parameters:
  • xs (numpy.ndarray) – Tensors to be padded.
  • max_len (int) – Maximum length.
Returns:

Padded tensor.
Return type:

numpy.ndarray
rrse(actual, predicted)[source]

Root Relative Squared Error.

Parameters:
  • actual (np.ndarray) – The actual value.
  • predicted (np.ndarray) – The predicted value.
Returns:

The root relative square error between the actual and the

predicted value.
Return type:

float
slice_nested_dict(dict_or_array, start, stop)[source]

Slice a dictionary containing arrays (or dictionaries).

This function is primarily intended for un-batching env_infos and action_infos.

Parameters:
  • dict_or_array (dict[str, dict or np.ndarray] or np.ndarray) – A nested dictionary should only contain dictionaries and numpy arrays (recursively).
  • start (int) – First index to be included in the slice.
  • stop (int) – First index to be excluded from the slice. In other words, these are typical python slice indices.
Returns:

The input, but sliced.
Return type:

dict or np.ndarray
sliding_window(t, window, smear=False)[source]

Create a sliding window over a tensor.

Parameters:
  • t (np.ndarray) – A tensor to create sliding window from, with shape \((N, D)\), where N is the length of a trajectory, D is the dimension of each step in trajectory.
  • window (int) – Window size, mush be less than N.
  • smear (bool) – If true, copy the last window so that N windows are generated.
Returns:

All windows generate over t, with shape \((M, W, D)\),

where W is the window size. If smear if False, M is \(N-W+1\), otherwise M is N.
Return type:

np.ndarray
Raises:
split_tensor_dict_list(tensor_dict)[source]

Split dictionary of list of tensor.

Parameters:tensor_dict (dict[numpy.ndarray]) – a dictionary of {tensors or dictionary of tensors}.
Returns:
a dictionary of {stacked tensors or dictionary of
stacked tensors}
Return type:dict
stack_and_pad_tensor_dict_list(tensor_dict_list, max_len)[source]

Stack and pad array of list of tensors.

Input paths are a list of N dicts, each with values of shape \((D, S^*)\). This function stack and pad the values with the input key with max_len, so output will be shape \((N, D, S^*)\).

Parameters:
  • tensor_dict_list (list[dict]) – List of dict to be stacked and padded. Value of each dict will be shape of \((D, S^*)\).
  • max_len (int) – Maximum length for padding.
Returns:

a dictionary of {stacked tensors or dictionary of

stacked tensors}. Shape: \((N, D, S^*)\) where N is the len of input paths.
Return type:

dict
stack_tensor_dict_list(tensor_dict_list)[source]

Stack a list of dictionaries of {tensors or dictionary of tensors}.

Parameters:tensor_dict_list (dict[list]) – a list of dictionaries of {tensors or dictionary of tensors}.
Returns:
a dictionary of {stacked tensors or dictionary of
stacked tensors}
Return type:dict
truncate_tensor_dict(tensor_dict, truncated_len)[source]

Truncate dictionary of list of tensor.

Parameters:
  • tensor_dict (dict[numpy.ndarray]) – a dictionary of {tensors or dictionary of tensors}.
  • truncated_len (int) – Length to truncate.
Returns:

a dictionary of {stacked tensors or dictionary of

stacked tensors}
Return type:

dict
unflatten_tensors(flattened, tensor_shapes)[source]

Unflatten a flattened tensors into a list of tensors.

Parameters:
  • flattened (numpy.ndarray) – Flattened tensors.
  • tensor_shapes (tuple) – Tensor shapes.
Returns:

Unflattened list of tensors.
Return type:

list[numpy.ndarray]