learning.normalizer#

Normalizer module.

Enables distributed normalizers to keep preprocessing consistent across all nodes. The normalizer is based on the implementation in https://github.com/openai/baselines.

Module Contents#

class learning.normalizer.Normalizer(size: int, eps: float = 0.01)#

Normalizer to maintain an estimate on the current data mean and variance.

Used to normalize input data to zero mean and unit variance. .. py:method:: __call__(x: torch.FloatTensor) -> torch.FloatTensor

Alias for self.normalize.

Parameters:

  • size (int) –

  • eps (float) –

normalize(x: torch.FloatTensor) → torch.FloatTensor#

Normalize the input data with the current mean and variance estimate.

Parameters:

x (torch.FloatTensor) – Input data array.

Returns:

The normalized data.

Return type:

torch.FloatTensor

update(x: torch.FloatTensor) → None#

Update the mean and variance estimate with new data.

Parameters:

x (torch.FloatTensor) – New input data. Expects a 3D array of shape (episodes, timestep, data dimension).

Raises:

AssertionError – Shape check failed.

Return type:

None

wrap_obs(states: torch.FloatTensor, goals: torch.FloatTensor) → torch.FloatTensor#

Wrap states and goals into a contingent input tensor.

Parameters:

  • states (torch.FloatTensor) – Input states array.

  • goals (torch.FloatTensor) – Input goals array.

Returns:

A fused state goal tensor.

Return type:

torch.FloatTensor

load(checkpoint: Any) → None#

Load data for the state_norm.

Parameters:

checkpoint (Any) – dict containing loaded data.

Return type:

None

save(f: io.BufferedWriter) → None#

Save data for the state_norm.

Parameters:

f (io.BufferedWriter) –

Return type:

None