linna.network#

Classes

Network(torch_model)

Wraps a sequential PyTorch model.

NetworkLayer(torch_model, layer_idx)

class linna.network.Network(torch_model: torch.nn.Sequential)#

Wraps a sequential PyTorch model. A layer consists of linear layer and possibly a non-linear function. Suppose a sequential PyTorch model with 6 layers has the following structure:

nn.Linear, nn.ReLU, nn.Linear, nn.ReLU, nn.Linear, nn.Softmax

In our setting, we have 3 layers that are organized as follows:

Layer 0: nn.Linear, nn.ReLU Layer 1: nn.Linear, nn.ReLU Layer 2: nn.Linear, nn.Softmax

Throughout the class neurons are uniquely identified by their index, i.e. corresponding row index in the weight matrix.

reset()#
get_num_neurons()#

Returns the number of neurons

Returns:

Number of neurons

Return type:

int

forward(X: torch.Tensor, layer_idx: Optional[int] = None, grad: bool = False)#

Computes the forward pass until the given layer (inclusive)

Parameters:

  • X (torch.Tensor) –

  • layer_idx (int, optional) –

  • grad (bool, default=False) –

Returns:

Output of the neural network

Return type:

torch.Tensor

classify(X: torch.Tensor)#

Classifies the given inputs

Parameters:

X (torch.Tensor) – Inputs to be classified

Returns:

Classification result

Return type:

torch.Tensor

delete_neuron(layer_idx: int, neuron: int)#

Removes a neuron from a given layer

Parameters:

  • layer_idx (int) – Layer in which neuron should be deleted

  • neuron (int) – Neuron to be deleted

restore_neuron(layer_idx: int, neuron: int)#

Restores the neuron in the given layer

Parameters:

  • layer_idx (int) – Layer in which neuron should be restored

  • neuron (int) – Neuron to be restored

set_basis(layer_idx: int, basis: List[int])#

Sets the basis for the specific layer. Note that due to

Parameters:

  • layer_idx (int) – Layer

  • basis (List[int]) – List of basis neurons

readjust_weights(layer_idx: int, neuron: int, coef: torch.Tensor)#

Readjust the outgoing weights for the neuron in the given layer. Effectively, the weight matrix of layer_idx + 1 is modified.

Parameters:

  • layer_idx (int) – Layer whose outgoing weights are adjusted

  • neuron (int) – Neuron whose weight is adjusted

  • coef (torch.Tensor) – Linear coefficients

get_io_matrix(layer_idx: int, loader, size=1000) → ndarray#

Computes the IO matrix for the given layer

Parameters:

  • layer_idx (int) – Layer for which IO matrix should be computed

  • loader (DataLoader) –

  • size (int) – Number of images to be considered

Returns:

IO matrix

Return type:

torch.Tensor

export_to_nnet(filename: str)#
update_torch_model()#
class linna.network.NetworkLayer(torch_model: torch.nn.Sequential, layer_idx: int)#
get_weight()#

Return the weight of the layer

Returns:

Weight of the layer

Return type:

torch.Tensor

get_bias()#

Returns the bias of the layer

Returns:

Return bias of layer

Return type:

torch.Tensor

set_weight(weight: torch.Tensor)#

Sets the weight of the layer

Parameters:

weight (torch.Tensor) – Weight

set_bias(bias: torch.Tensor)#

Sets the bias of the layer

Parameters:

bias (torch.Tensor) –

get_input_weight(neuron: int)#

Returns the input weight of the neuron

Parameters:

neuron (int) – Neuron

Returns:

Input weights of the neuron

Return type:

torch.Tensor

delete_output(neuron: int)#

Deletes the neuron from the layer

Parameters:

neuron (int) – Neuron to be deleted

delete_input(neuron: int)#

Deletes the input neuron

Parameters:

neuron (int) – Neuron

get_neuron_index(neuron: int)#

Return the index of the neuron

Parameters:

neuron (int) – Neuron

Returns:

Index of the neuron

Return type:

int

readjust_weights(neuron: int, coef: torch.Tensor)#

Readjust

Parameters:

  • neuron (int) – Neuron

  • coef (torch.Tensor) – A tensor containing coefficients of linear combination

restore_neuron(neuron: int)#

Restores a given neuron

Parameters:

neuron (int) – Neuron to be restored

restore_input(neuron)#

Restores an input neuron to the layer

Parameters:

neuron (int) – Input neuron

reset()#

Resets the layer

restore_weights(neuron)#

Restores the weights of the layer w.r.t a previously removed neuron

Parameters:

neuron (int) – Neuron whose weights should be restored