echoflow.core package

class echoflow.core.BatchNorm(input_dims: int, momentum: float = 0.0, eps: float = 1e-05)

Bases: echoflow.core.base.BaseFlow

Batch normalization from RealNVP.

Parameters

• input_dims – The number of input dimensions.

• momentum – The momentum used to compute the running mean/var.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.Coupling(input_dims: int, hidden_dims: int, input_mask: torch.Tensor, context_dims: int = 0)

Bases: echoflow.core.base.BaseFlow

Coupling layer from RealNVP.

The coupling layer partitions the input x into two parts, x1 and x2, and applies an invertible transform:

\[\begin{split}y1 &= x1 \\ y2 &= x2 * exp(s(x1)) + t(x1)\end{split}\]

which modifies only one of the partitions.

Parameters

• input_dims – The number of input dimensions.

• hidden_dims – The hidden size to use for the scale/translate nets.

• input_mask – A binary mask for the input.

• context_dims – The number of context dimensions. If specified, then the output is conditioned on context.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.MADE(input_dims, hidden_dims, context_dims=0)

Bases: echoflow.core.base.BaseFlow

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.OneHot(cardinality: List[int])

Bases: echoflow.core.base.BaseFlow

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.Reverse(input_dims: int)

Bases: echoflow.core.base.BaseFlow

Reversing layer from MADE.

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.SequentialFlow(*modules)

Bases: echoflow.core.base.BaseFlow

Apply a sequence of flows.

Parameters

*args – A list of BaseFlow layers to apply in order.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

Submodules

echoflow.core.base module

class echoflow.core.base.BaseFlow

Bases: torch.nn.modules.module.Module

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.batch_norm module

class echoflow.core.batch_norm.BatchNorm(input_dims: int, momentum: float = 0.0, eps: float = 1e-05)

Bases: echoflow.core.base.BaseFlow

Batch normalization from RealNVP.

Parameters

• input_dims – The number of input dimensions.

• momentum – The momentum used to compute the running mean/var.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.coupling module

class echoflow.core.coupling.Coupling(input_dims: int, hidden_dims: int, input_mask: torch.Tensor, context_dims: int = 0)

Bases: echoflow.core.base.BaseFlow

Coupling layer from RealNVP.

The coupling layer partitions the input x into two parts, x1 and x2, and applies an invertible transform:

\[\begin{split}y1 &= x1 \\ y2 &= x2 * exp(s(x1)) + t(x1)\end{split}\]

which modifies only one of the partitions.

Parameters

• input_dims – The number of input dimensions.

• hidden_dims – The hidden size to use for the scale/translate nets.

• input_mask – A binary mask for the input.

• context_dims – The number of context dimensions. If specified, then the output is conditioned on context.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.made module

class echoflow.core.made.MADE(input_dims, hidden_dims, context_dims=0)

Bases: echoflow.core.base.BaseFlow

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

class echoflow.core.made.MaskedLinear(input_dims: int, out_features: int, weight_mask: torch.Tensor, context_dims: int = 0)

Bases: echoflow.core.base.BaseFlow

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.made.get_mask(in_features: int, out_features: int, in_flow_features: int, mask_type=None)

echoflow.core.one_hot module

class echoflow.core.one_hot.OneHot(cardinality: List[int])

Bases: echoflow.core.base.BaseFlow

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.reverse module

class echoflow.core.reverse.Reverse(input_dims: int)

Bases: echoflow.core.base.BaseFlow

Reversing layer from MADE.

Normalizing flow layer.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None

echoflow.core.sequential module

class echoflow.core.sequential.SequentialFlow(*modules)

Bases: echoflow.core.base.BaseFlow

Apply a sequence of flows.

Parameters

*args – A list of BaseFlow layers to apply in order.

forward(inputs: torch.Tensor, contexts: Optional[torch.Tensor] = None, inverse: bool = False)

Transform a batch of data.

Parameters

• inputs – The input tensor.

• contexts – An optional context tensor (for conditional sampling).

• inverse – Whether to apply the direct or inverse transform.

Returns

• outputs (torch.Tensor) – The output tensor.

• logdet (torch.Tensor) – The log-determinant of the Jacobian.

training = None