model_search

Cell

Bases: nn.Module

A cell as defined in differentiable architecture search. A single cell corresponds to a computation graph with the number of input nodes defined by n_input_states and the number of hidden nodes defined by steps. Input nodes only project to hidden nodes and hidden nodes project to each other with an acyclic connectivity pattern. The output of a cell corresponds to the concatenation of all hidden nodes. Hidden nodes are computed by integrating transformed outputs from sending nodes. Outputs from sending nodes correspond to mixture operations, i.e. a weighted combination of pre-specified operations applied to the variable specified by the sending node (see MixedOp).

Attributes:

Name	Type	Description
_steps		number of hidden nodes
_n_input_states		number of input nodes
_ops		list of mixture operations (amounts to the list of edges in the cell)

Source code in src/autora/theorist/darts/model_search.py

class Cell(nn.Module):
    """
    A cell as defined in differentiable architecture search. A single cell corresponds
    to a computation graph with the number of input nodes defined by n_input_states and
    the number of hidden nodes defined by steps. Input nodes only project to hidden nodes and hidden
    nodes project to each other with an acyclic connectivity pattern. The output of a cell
    corresponds to the concatenation of all hidden nodes. Hidden nodes are computed by integrating
    transformed outputs from sending nodes. Outputs from sending nodes correspond to
    mixture operations, i.e. a weighted combination of pre-specified operations applied to the
    variable specified by the sending node (see MixedOp).

    Attributes:
        _steps: number of hidden nodes
        _n_input_states: number of input nodes
        _ops: list of mixture operations (amounts to the list of edges in the cell)
    """

    def __init__(
        self,
        steps: int = 2,
        n_input_states: int = 1,
        primitives: Sequence[str] = PRIMITIVES,
    ):
        """
        Initializes a cell based on the number of hidden nodes (steps)
        and the number of input nodes (n_input_states).

        Arguments:
            steps: number of hidden nodes
            n_input_states: number of input nodes
        """
        # The first and second nodes of cell k are set equal to the outputs of
        # cell k − 2 and cell k − 1, respectively, and 1 × 1 convolutions
        # (ReLUConvBN) are inserted as necessary
        super(Cell, self).__init__()

        # set parameters
        self._steps = steps  # hidden nodes
        self._n_input_states = n_input_states  # input nodes

        # EDIT 11/04/19 SM: adapting to new SimpleNet data (changed from
        # multiplier to steps)
        self._multiplier = steps

        # set operations according to number of modules (empty)
        self._ops = nn.ModuleList()
        # iterate over edges: edges between each hidden node and input nodes +
        # prev hidden nodes
        for i in range(self._steps):  # hidden nodes
            for j in range(self._n_input_states + i):  # 2 refers to the 2 input nodes
                # defines the stride for link between cells
                # adds a mixed operation (derived from architecture parameters alpha)
                # for 4 intermediate nodes, a total of 14 connections
                # (MixedOps) is added
                op = MixedOp(primitives)
                # appends cell with mixed operation
                self._ops.append(op)

    def forward(self, input_states: List, weights: torch.Tensor):
        """
        Computes the output of a cell given a list of input states
        (variables represented in input nodes) and a weight matrix specifying the weights of each
        operation for each edge.

        Arguments:
            input_states: list of input nodes
            weights: matrix specifying architecture weights, i.e. the weights associated
                with each operation for each edge
        """
        # initialize states (activities of each node in the cell)
        states = list()

        # add each input node to the number of states
        for input in input_states:
            states.append(input)

        offset = 0
        # this computes the states from intermediate nodes and adds them to the list of states
        # (values of nodes)
        # for each hidden node, compute edge between existing states (input
        # nodes / previous hidden) nodes and current node
        for i in range(
            self._steps
        ):  # compute the state for each hidden node, first hidden node is
            # sum of input nodes, second is sum of input and first hidden
            s = sum(
                self._ops[offset + j](h, weights[offset + j])
                for j, h in enumerate(states)
            )
            offset += len(states)
            states.append(s)

        # concatenates the states of the last n (self._multiplier) intermediate
        # nodes to get the output of a cell
        result = torch.cat(states[-self._multiplier :], dim=1)
        return result

__init__(steps=2, n_input_states=1, primitives=PRIMITIVES)

Initializes a cell based on the number of hidden nodes (steps) and the number of input nodes (n_input_states).

Parameters:

Name	Type	Description	Default
steps	int	number of hidden nodes	2
n_input_states	int	number of input nodes	1

Source code in src/autora/theorist/darts/model_search.py

def __init__(
    self,
    steps: int = 2,
    n_input_states: int = 1,
    primitives: Sequence[str] = PRIMITIVES,
):
    """
    Initializes a cell based on the number of hidden nodes (steps)
    and the number of input nodes (n_input_states).

    Arguments:
        steps: number of hidden nodes
        n_input_states: number of input nodes
    """
    # The first and second nodes of cell k are set equal to the outputs of
    # cell k − 2 and cell k − 1, respectively, and 1 × 1 convolutions
    # (ReLUConvBN) are inserted as necessary
    super(Cell, self).__init__()

    # set parameters
    self._steps = steps  # hidden nodes
    self._n_input_states = n_input_states  # input nodes

    # EDIT 11/04/19 SM: adapting to new SimpleNet data (changed from
    # multiplier to steps)
    self._multiplier = steps

    # set operations according to number of modules (empty)
    self._ops = nn.ModuleList()
    # iterate over edges: edges between each hidden node and input nodes +
    # prev hidden nodes
    for i in range(self._steps):  # hidden nodes
        for j in range(self._n_input_states + i):  # 2 refers to the 2 input nodes
            # defines the stride for link between cells
            # adds a mixed operation (derived from architecture parameters alpha)
            # for 4 intermediate nodes, a total of 14 connections
            # (MixedOps) is added
            op = MixedOp(primitives)
            # appends cell with mixed operation
            self._ops.append(op)

forward(input_states, weights)

Computes the output of a cell given a list of input states (variables represented in input nodes) and a weight matrix specifying the weights of each operation for each edge.

Parameters:

Name	Type	Description	Default
input_states	List	list of input nodes	required
weights	torch.Tensor	matrix specifying architecture weights, i.e. the weights associated with each operation for each edge	required

Source code in src/autora/theorist/darts/model_search.py

def forward(self, input_states: List, weights: torch.Tensor):
    """
    Computes the output of a cell given a list of input states
    (variables represented in input nodes) and a weight matrix specifying the weights of each
    operation for each edge.

    Arguments:
        input_states: list of input nodes
        weights: matrix specifying architecture weights, i.e. the weights associated
            with each operation for each edge
    """
    # initialize states (activities of each node in the cell)
    states = list()

    # add each input node to the number of states
    for input in input_states:
        states.append(input)

    offset = 0
    # this computes the states from intermediate nodes and adds them to the list of states
    # (values of nodes)
    # for each hidden node, compute edge between existing states (input
    # nodes / previous hidden) nodes and current node
    for i in range(
        self._steps
    ):  # compute the state for each hidden node, first hidden node is
        # sum of input nodes, second is sum of input and first hidden
        s = sum(
            self._ops[offset + j](h, weights[offset + j])
            for j, h in enumerate(states)
        )
        offset += len(states)
        states.append(s)

    # concatenates the states of the last n (self._multiplier) intermediate
    # nodes to get the output of a cell
    result = torch.cat(states[-self._multiplier :], dim=1)
    return result

DARTSType

Bases: str, Enum

Enumerator that indexes different variants of DARTS.

Source code in src/autora/theorist/darts/model_search.py

class DARTSType(str, Enum):
    """
    Enumerator that indexes different variants of DARTS.
    """

    # Liu, Simonyan & Yang (2018). Darts: Differentiable architecture search
    ORIGINAL = "original"

    # Chu, Zhou, Zhang & Li (2020). Fair darts: Eliminating unfair advantages
    # in differentiable architecture search
    FAIR = "fair"

MixedOp

Bases: nn.Module

Mixture operation as applied in Differentiable Architecture Search (DARTS). A mixture operation amounts to a weighted mixture of a pre-defined set of operations that is applied to an input variable.

Source code in src/autora/theorist/darts/model_search.py

class MixedOp(nn.Module):
    """
    Mixture operation as applied in Differentiable Architecture Search (DARTS).
    A mixture operation amounts to a weighted mixture of a pre-defined set of operations
    that is applied to an input variable.
    """

    def __init__(self, primitives: Sequence[str] = PRIMITIVES):
        """
        Initializes a mixture operation based on a pre-specified set of primitive operations.

        Arguments:
            primitives: list of primitives to be used in the mixture operation
        """
        super(MixedOp, self).__init__()
        self._ops = nn.ModuleList()
        # loop through all the 8 primitive operations
        for primitive in primitives:
            # OPS returns an nn module for a given primitive (defines as a string)
            op = operation_factory(primitive)

            # add the operation
            self._ops.append(op)

    def forward(self, x: torch.Tensor, weights: torch.Tensor) -> float:
        """
        Computes a mixture operation as a weighted sum of all primitive operations.

        Arguments:
            x: input to the mixture operations
            weights: weight vector containing the weights associated with each operation

        Returns:
            y: result of the weighted mixture operation
        """
        # there are 8 weights for all the eight primitives. then it returns the
        # weighted sum of all operations performed on a given input
        return sum(w * op(x) for w, op in zip(weights, self._ops))

__init__(primitives=PRIMITIVES)

Initializes a mixture operation based on a pre-specified set of primitive operations.

Parameters:

Name	Type	Description	Default
primitives	Sequence[str]	list of primitives to be used in the mixture operation	PRIMITIVES

Source code in src/autora/theorist/darts/model_search.py

def __init__(self, primitives: Sequence[str] = PRIMITIVES):
    """
    Initializes a mixture operation based on a pre-specified set of primitive operations.

    Arguments:
        primitives: list of primitives to be used in the mixture operation
    """
    super(MixedOp, self).__init__()
    self._ops = nn.ModuleList()
    # loop through all the 8 primitive operations
    for primitive in primitives:
        # OPS returns an nn module for a given primitive (defines as a string)
        op = operation_factory(primitive)

        # add the operation
        self._ops.append(op)

forward(x, weights)

Computes a mixture operation as a weighted sum of all primitive operations.

Parameters:

Name	Type	Description	Default
x	torch.Tensor	input to the mixture operations	required
weights	torch.Tensor	weight vector containing the weights associated with each operation	required

Returns:

Name	Type	Description
y	float	result of the weighted mixture operation

Source code in src/autora/theorist/darts/model_search.py

def forward(self, x: torch.Tensor, weights: torch.Tensor) -> float:
    """
    Computes a mixture operation as a weighted sum of all primitive operations.

    Arguments:
        x: input to the mixture operations
        weights: weight vector containing the weights associated with each operation

    Returns:
        y: result of the weighted mixture operation
    """
    # there are 8 weights for all the eight primitives. then it returns the
    # weighted sum of all operations performed on a given input
    return sum(w * op(x) for w, op in zip(weights, self._ops))

Network

Bases: nn.Module

A PyTorch computation graph according to DARTS. It consists of a single computation cell which transforms an input vector (containing all input variable) into an output vector, by applying a set of mixture operations which are defined by the architecture weights (labeled "alphas" of the network).

The network flow looks as follows: An input vector (with _n_input_states elements) is split into _n_input_states separate input nodes (one node per element). The input nodes are then passed through a computation cell with _steps hidden nodes (see Cell). The output of the computation cell corresponds to the concatenation of its hidden nodes (a single vector). The final output corresponds to a (trained) affine transformation of this concatenation (labeled "classifier").

Attributes:

Name	Type	Description
_n_input_states		length of input vector (translates to number of input nodes)
_num_classes		length of output vector
_criterion		optimization criterion used to define the loss
_steps		number of hidden nodes in the cell
_architecture_fixed		specifies whether the architecture weights shall remain fixed (not trained)
_classifier_weight_decay		a weight decay applied to the classifier

Source code in src/autora/theorist/darts/model_search.py

class Network(nn.Module):
    """
    A PyTorch computation graph according to DARTS.
    It consists of a single computation cell which transforms an
    input vector (containing all input variable) into an output vector, by applying a set of
    mixture operations which are defined by the architecture weights (labeled "alphas" of the
    network).

    The network flow looks as follows: An input vector (with _n_input_states elements) is split into
    _n_input_states separate input nodes (one node per element). The input nodes are then passed
    through a computation cell with _steps hidden nodes (see Cell). The output of the computation
    cell corresponds to the concatenation of its hidden nodes (a single vector). The final output
    corresponds to a (trained) affine transformation of this concatenation (labeled "classifier").

    Attributes:
        _n_input_states: length of input vector (translates to number of input nodes)
        _num_classes: length of output vector
        _criterion: optimization criterion used to define the loss
        _steps: number of hidden nodes in the cell
        _architecture_fixed: specifies whether the architecture weights shall remain fixed
            (not trained)
        _classifier_weight_decay: a weight decay applied to the classifier

    """

    def __init__(
        self,
        num_classes: int,
        criterion: Callable,
        steps: int = 2,
        n_input_states: int = 2,
        architecture_fixed: bool = False,
        train_classifier_coefficients: bool = False,
        train_classifier_bias: bool = False,
        classifier_weight_decay: float = 0,
        darts_type: DARTSType = DARTSType.ORIGINAL,
        primitives: Sequence[str] = PRIMITIVES,
    ):
        """
        Initializes the network.

        Arguments:
            num_classes: length of output vector
            criterion: optimization criterion used to define the loss
            steps: number of hidden nodes in the cell
            n_input_states: length of input vector (translates to number of input nodes)
            architecture_fixed: specifies whether the architecture weights shall remain fixed
            train_classifier_coefficients: specifies whether the classifier coefficients shall be
                trained
            train_classifier_bias: specifies whether the classifier bias shall be trained
            classifier_weight_decay: a weight decay applied to the classifier
            darts_type: variant of DARTS (regular or fair) that is applied for training
        """
        super(Network, self).__init__()

        # set parameters
        self._num_classes = num_classes  # number of output classes
        self._criterion = criterion  # optimization criterion (e.g., softmax)
        self._steps = steps  # the number of intermediate nodes (e.g., 2)
        self._n_input_states = n_input_states  # number of input nodes
        self.DARTS_type = darts_type  # darts variant
        self._multiplier = (
            1  # the number of internal nodes that get concatenated to the output
        )
        self.primitives = primitives

        # set parameters
        self._dim_output = self._steps
        self._architecture_fixed = architecture_fixed
        self._classifier_weight_decay = classifier_weight_decay

        # input nodes
        self.stem = nn.Sequential(Fan_Out(self._n_input_states))

        self.cells = (
            nn.ModuleList()
        )  # get list of all current modules (should be empty)

        # generate a cell that undergoes architecture search
        self.cells = Cell(steps, self._n_input_states, self.primitives)

        # last layer is a linear classifier (e.g. with 10 CIFAR classes)
        self.classifier = nn.Linear(
            self._dim_output, num_classes
        )  # make this the number of input states

        # initialize classifier weights
        if train_classifier_coefficients is False:
            self.classifier.weight.data.fill_(1)
            self.classifier.weight.requires_grad = False

        if train_classifier_bias is False:
            self.classifier.bias.data.fill_(0)
            self.classifier.bias.requires_grad = False

        # initializes weights of the architecture
        self._initialize_alphas()

    # function for copying the network
    def new(self) -> nn.Module:
        """
        Returns a copy of the network.

        Returns:
            a copy of the network

        """

        model_new = Network(
            # self._C, self._num_classes, self._criterion, steps=self._steps
            num_classes=self._num_classes,
            criterion=self._criterion,
            steps=self._steps,
            n_input_states=self._n_input_states,
            architecture_fixed=self._architecture_fixed,
            classifier_weight_decay=self._classifier_weight_decay,
            darts_type=self.DARTS_type,
            primitives=self.primitives,
        )

        for x, y in zip(model_new.arch_parameters(), self.arch_parameters()):
            x.data.copy_(y.data)
        return model_new

    # computes forward pass for full network
    def forward(self, x: torch.Tensor):
        """
        Computes output of the network.

        Arguments:
            x: input to the network
        """

        # compute stem first
        input_states = self.stem(x)

        # get architecture weights
        if self._architecture_fixed:
            weights = self.alphas_normal
        else:
            if self.DARTS_type == DARTSType.ORIGINAL:
                weights = F.softmax(self.alphas_normal, dim=-1)
            elif self.DARTS_type == DARTSType.FAIR:
                weights = torch.sigmoid(self.alphas_normal)
            else:
                raise Exception(
                    "DARTS Type " + str(self.DARTS_type) + " not implemented"
                )

        # then apply cell with weights
        cell_output = self.cells(input_states, weights)

        # compute logits
        logits = self.classifier(cell_output.view(cell_output.size(0), -1))
        # just gets output to have only 2 dimensions (batch_size x num units in
        # output layer)

        return logits

    def _loss(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
        """
        Computes the loss of the network for the specified criterion.

        Arguments:
            input: input patterns
            target: target patterns

        Returns:
            loss
        """
        logits = self(input)
        return self._criterion(logits, target)  # returns cross entropy by default

    # regularization
    def apply_weight_decay_to_classifier(self, lr: float):
        """
        Applies a weight decay to the weights projecting from the cell to the final output layer.

        Arguments:
            lr: learning rate
        """
        # weight decay proportional to degrees of freedom
        for p in self.classifier.parameters():
            if p.requires_grad is False:
                continue
            p.data.sub_(
                self._classifier_weight_decay
                * lr
                * torch.sign(p.data)
                * (torch.abs(p.data))
            )  # weight decay

    def _initialize_alphas(self):
        """
        Initializes the architecture weights.
        """
        # compute the number of possible connections between nodes
        k = sum(1 for i in range(self._steps) for n in range(self._n_input_states + i))
        # number of available primitive operations (8 different types for a
        # conv net)
        num_ops = len(self.primitives)

        # e.g., generate 14 (number of available edges) by 8 (operations)
        # weight matrix for normal alphas of the architecture
        self.alphas_normal = Variable(
            1e-3 * torch.randn(k, num_ops), requires_grad=True
        )
        # those are all the parameters of the architecture
        self._arch_parameters = [self.alphas_normal]

    # provide back the architecture as a parameter
    def arch_parameters(self) -> List:
        """
        Returns architecture weights.

        Returns:
            _arch_parameters: architecture weights.
        """
        return self._arch_parameters

    # fixes architecture
    def fix_architecture(
        self, switch: bool, new_weights: Optional[torch.Tensor] = None
    ):
        """
        Freezes or unfreezes the architecture weights.

        Arguments:
            switch: set true to freeze architecture weights or false unfreeze
            new_weights: new set of architecture weights
        """
        self._architecture_fixed = switch
        if new_weights is not None:
            self.alphas_normal = new_weights
        return

    def sample_alphas_normal(
        self, sample_amp: float = 1, fair_darts_weight_threshold: float = 0
    ) -> torch.Tensor:
        """
        Samples an architecture from the mixed operations from a probability distribution that is
        defined by the (softmaxed) architecture weights.
        This amounts to selecting one operation per edge (i.e., setting the architecture
        weight of that operation to one while setting the others to zero).

        Arguments:
            sample_amp: temperature that is applied before passing the weights through a softmax
            fair_darts_weight_threshold: used in fair DARTS. If an architecture weight is below
                this value then it is set to zero.

        Returns:
            alphas_normal_sample: sampled architecture weights.
        """

        alphas_normal = self.alphas_normal.clone()
        alphas_normal_sample = Variable(torch.zeros(alphas_normal.data.shape))

        for edge in range(alphas_normal.data.shape[0]):
            if self.DARTS_type == DARTSType.ORIGINAL:
                W_soft = F.softmax(alphas_normal[edge] * sample_amp, dim=0)
            elif self.DARTS_type == DARTSType.FAIR:
                transformed_alphas_normal = alphas_normal[edge]
                above_threshold = False
                for idx in range(len(transformed_alphas_normal.data)):
                    if (
                        torch.sigmoid(transformed_alphas_normal).data[idx]
                        > fair_darts_weight_threshold
                    ):
                        above_threshold = True
                        break
                if above_threshold:
                    W_soft = F.softmax(transformed_alphas_normal * sample_amp, dim=0)
                else:
                    W_soft = Variable(torch.zeros(alphas_normal[edge].shape))
                    W_soft[self.primitives.index("none")] = 1

            else:
                raise Exception(
                    "DARTS Type " + str(self.DARTS_type) + " not implemented"
                )

            if torch.any(W_soft != W_soft):
                warnings.warn(
                    "Cannot properly sample from architecture weights due to nan entries."
                )
                k_sample = random.randrange(len(W_soft))
            else:
                k_sample = np.random.choice(range(len(W_soft)), p=W_soft.data.numpy())
            alphas_normal_sample[edge, k_sample] = 1

        return alphas_normal_sample

    def max_alphas_normal(self) -> torch.Tensor:
        """
        Samples an architecture from the mixed operations by selecting, for each edge,
        the operation with the largest architecture weight.

        Returns:
            alphas_normal_sample: sampled architecture weights.
        """
        alphas_normal = self.alphas_normal.clone()
        alphas_normal_sample = Variable(torch.zeros(alphas_normal.data.shape))

        for edge in range(alphas_normal.data.shape[0]):
            row = alphas_normal[edge]
            max_idx = np.argmax(row.data)
            alphas_normal_sample[edge, max_idx] = 1

        return alphas_normal_sample

    # returns the genotype of the model
    def genotype(self, sample: bool = False) -> Genotype:
        """
        Computes a genotype of the model which specifies the current computation graph based on
        the largest architecture weight for each edge, or based on a sample.
        The genotype can be used for parsing or plotting the computation graph.

        Arguments:
            sample: if set to true, the architecture will be determined by sampling
                from a probability distribution that is determined by the
                softmaxed architecture weights. If set to false (default), the architecture will be
                determined based on the largest architecture weight per edge.

        Returns:
            genotype: genotype describing the current (sampled) architecture
        """
        # this function uses the architecture weights to retrieve the
        # operations with the highest weights
        def _parse(weights):
            gene = []
            n = (
                self._n_input_states
            )  # 2 ... changed this to adapt to number of input states
            start = 0
            for i in range(self._steps):
                end = start + n
                W = weights[start:end].copy()
                # first get all the edges for a given node, edges are sorted according to their
                # highest (non-none) weight, starting from the edge with the smallest heighest
                # weight

                if "none" in self.primitives:
                    none_index = self.primitives.index("none")
                else:
                    none_index = -1

                edges = sorted(
                    range(n),
                    key=lambda x: -max(
                        W[x][k] for k in range(len(W[x])) if k != none_index
                    ),
                )
                # for each edge, figure out which is the primitive with the
                # highest
                for (
                    j
                ) in edges:  # looping through all the edges for the current node (i)
                    if sample:
                        W_soft = F.softmax(Variable(torch.from_numpy(W[j])))
                        k_best = np.random.choice(
                            range(len(W[j])), p=W_soft.data.numpy()
                        )
                    else:
                        k_best = None
                        # looping through all the primitives
                        for k in range(len(W[j])):
                            # choose the primitive with the highest weight
                            # if k != self.primitives.index('none'):
                            # EDIT SM 01/13: commented to include "none"
                            # weights in genotype
                            if k_best is None or W[j][k] > W[j][k_best]:
                                k_best = k
                        # add gene (primitive, edge number)
                    gene.append((self.primitives[k_best], j))
                start = end
                n += 1
            return gene

        if self._architecture_fixed:
            gene_normal = _parse(self.alphas_normal.data.cpu().numpy())
        else:
            gene_normal = _parse(
                F.softmax(self.alphas_normal, dim=-1).data.cpu().numpy()
            )

        concat = range(2 + self._steps - self._multiplier, self._steps + 2)
        genotype = Genotype(
            normal=gene_normal,
            normal_concat=concat,
        )
        return genotype

    def count_parameters(self, print_parameters: bool = False) -> Tuple[int, int, list]:
        """
        Counts and returns the parameters (coefficients) of the architecture defined by the
        highest architecture weights.

        Arguments:
            print_parameters: if set to true, the function will print all parameters.

        Returns:
            n_params_total: total number of parameters
            n_params_base: number of parameters determined by the classifier
            param_list: list of parameters specifying the corresponding edge (operation)
                and value
        """

        # counts only parameters of operations with the highest architecture weight
        n_params_total = 0

        # count classifier
        for parameter in self.classifier.parameters():
            if parameter.requires_grad is True:
                n_params_total += parameter.data.numel()

        # count stem
        for parameter in self.stem.parameters():
            if parameter.requires_grad is True:
                n_params_total += parameter.data.numel()

        n_params_base = (
            n_params_total  # number of parameters, excluding individual cells
        )

        param_list = list()
        # now count number of parameters for cells that have highest
        # probability
        for idx, op in enumerate(self.cells._ops):
            # pick most operation with highest likelihood
            values = self.alphas_normal[idx, :].data.numpy()
            maxIdx = np.where(values == max(values))

            tmp_param_list = list()
            if isiterable(op._ops[maxIdx[0].item(0)]):  # Zero is not iterable

                for subop in op._ops[maxIdx[0].item(0)]:

                    for parameter in subop.parameters():
                        tmp_param_list.append(parameter.data.numpy().squeeze())
                        if parameter.requires_grad is True:
                            n_params_total += parameter.data.numel()

            if print_parameters:
                print(
                    "Edge ("
                    + str(idx)
                    + "): "
                    + get_operation_label(
                        self.primitives[maxIdx[0].item(0)], tmp_param_list
                    )
                )
            param_list.append(tmp_param_list)

        # # get parameters from final linear classifier
        # tmp_param_list = list()
        # for parameter in self.classifier.parameters():
        #   for subparameter in parameter:
        #     tmp_param_list.append(subparameter.data.numpy().squeeze())

        # get parameters from final linear for each edge
        for edge in range(self._steps):
            tmp_param_list = list()
            # add weight
            tmp_param_list.append(
                self.classifier._parameters["weight"].data[:, edge].numpy()
            )
            # add partial bias (bias of classifier units will be divided by
            # number of edges)
            if "bias" in self.classifier._parameters.keys() and edge == 0:
                tmp_param_list.append(self.classifier._parameters["bias"].data.numpy())
            param_list.append(tmp_param_list)

            if print_parameters:
                print(
                    "Classifier from Node "
                    + str(edge)
                    + ": "
                    + get_operation_label("classifier_concat", tmp_param_list)
                )

        return (n_params_total, n_params_base, param_list)

    def architecture_to_str_list(
        self,
        input_labels: Sequence[str],
        output_labels: Sequence[str],
        output_function_label: str = "",
        decimals_to_display: int = 2,
        output_format: Literal["latex", "console"] = "console",
    ) -> List:
        """
        Returns a list of strings representing the model.

        Arguments:
            input_labels: list of strings representing the input states.
            output_labels: list of strings representing the output states.
            output_function_label: string representing the output function.
            decimals_to_display: number of decimals to display.
            output_format: if set to `"console"`, returns equations formatted for the command line,
                if set to `"latex"`, returns equations in latex format


        Returns:
            list of strings representing the model
        """
        (n_params_total, n_params_base, param_list) = self.count_parameters(
            print_parameters=False
        )
        genotype = self.genotype().normal
        steps = self._steps
        edge_list = list()

        n = len(input_labels)
        start = 0
        for i in range(steps):  # for every node
            end = start + n
            # for k in [2*i, 2*i + 1]:

            edge_operations_list = list()
            op_list = list()

            for k in range(start, end):
                if (
                    output_format == "latex"
                ):  # for every edge projecting to current node
                    v = "k_" + str(i + 1)
                else:
                    v = "k" + str(i + 1)
                op, j = genotype[k]
                if j < len(input_labels):
                    u = input_labels[j]
                else:
                    if output_format == "latex":
                        u = "k_" + str(j - len(input_labels) + 1)
                    else:
                        u = "k" + str(j - len(input_labels) + 1)
                if op != "none":
                    op_label = op
                    params = param_list[
                        start + j
                    ]  # note: genotype order and param list order don't align
                    op_label = get_operation_label(
                        op,
                        params,
                        decimals=decimals_to_display,
                        input_var=u,
                        output_format=output_format,
                    )
                    op_list.append(op)
                    edge_operations_list.append(op_label)

            if len(edge_operations_list) == 0:
                edge_str = v + " = 0"
            else:
                edge_str = ""
            for i, edge_operation in enumerate(edge_operations_list):
                if i == 0:
                    edge_str += v + " = " + edge_operation
                if i > 0:
                    if (
                        op_list[i] != "add"
                        and op_list[i] != "subtract"
                        and op_list[i] != "none"
                    ):
                        edge_str += " +"
                    edge_str += " " + edge_operation

            edge_list.append(edge_str)
            start = end
            n += 1

        # TODO: extend to multiple outputs
        if output_format == "latex":
            classifier_str = output_labels[0] + " = " + output_function_label
            if output_function_label != "":
                classifier_str += "\\left("
        else:
            classifier_str = output_labels[0] + " = " + output_function_label
            if output_function_label != "":
                classifier_str += "("

        bias = None
        for i in range(steps):
            param_idx = len(param_list) - steps + i
            tmp_param_list = param_list[param_idx]
            if i == 0 and len(tmp_param_list) == 2:
                bias = tmp_param_list[1]
            if i > 0:
                classifier_str += " + "

            if output_format == "latex":
                input_var = "k_" + str(i + 1)
            else:
                input_var = "k" + str(i + 1)

            classifier_str += get_operation_label(
                "classifier",
                tmp_param_list[0],
                decimals=decimals_to_display,
                input_var=input_var,
            )

            if i == steps - 1 and bias is not None:
                classifier_str += " + " + str(bias[0])

            if i == steps - 1:
                if output_function_label != "":
                    if output_format == "latex":
                        classifier_str += "\\right)"
                    else:
                        classifier_str += ")"

        edge_list.append(classifier_str)

        return edge_list

__init__(num_classes, criterion, steps=2, n_input_states=2, architecture_fixed=False, train_classifier_coefficients=False, train_classifier_bias=False, classifier_weight_decay=0, darts_type=DARTSType.ORIGINAL, primitives=PRIMITIVES)

Initializes the network.

Parameters:

Name	Type	Description	Default
num_classes	int	length of output vector	required
criterion	Callable	optimization criterion used to define the loss	required
steps	int	number of hidden nodes in the cell	2
n_input_states	int	length of input vector (translates to number of input nodes)	2
architecture_fixed	bool	specifies whether the architecture weights shall remain fixed	False
train_classifier_coefficients	bool	specifies whether the classifier coefficients shall be trained	False
train_classifier_bias	bool	specifies whether the classifier bias shall be trained	False
classifier_weight_decay	float	a weight decay applied to the classifier	0
darts_type	DARTSType	variant of DARTS (regular or fair) that is applied for training	DARTSType.ORIGINAL

Source code in src/autora/theorist/darts/model_search.py

def __init__(
    self,
    num_classes: int,
    criterion: Callable,
    steps: int = 2,
    n_input_states: int = 2,
    architecture_fixed: bool = False,
    train_classifier_coefficients: bool = False,
    train_classifier_bias: bool = False,
    classifier_weight_decay: float = 0,
    darts_type: DARTSType = DARTSType.ORIGINAL,
    primitives: Sequence[str] = PRIMITIVES,
):
    """
    Initializes the network.

    Arguments:
        num_classes: length of output vector
        criterion: optimization criterion used to define the loss
        steps: number of hidden nodes in the cell
        n_input_states: length of input vector (translates to number of input nodes)
        architecture_fixed: specifies whether the architecture weights shall remain fixed
        train_classifier_coefficients: specifies whether the classifier coefficients shall be
            trained
        train_classifier_bias: specifies whether the classifier bias shall be trained
        classifier_weight_decay: a weight decay applied to the classifier
        darts_type: variant of DARTS (regular or fair) that is applied for training
    """
    super(Network, self).__init__()

    # set parameters
    self._num_classes = num_classes  # number of output classes
    self._criterion = criterion  # optimization criterion (e.g., softmax)
    self._steps = steps  # the number of intermediate nodes (e.g., 2)
    self._n_input_states = n_input_states  # number of input nodes
    self.DARTS_type = darts_type  # darts variant
    self._multiplier = (
        1  # the number of internal nodes that get concatenated to the output
    )
    self.primitives = primitives

    # set parameters
    self._dim_output = self._steps
    self._architecture_fixed = architecture_fixed
    self._classifier_weight_decay = classifier_weight_decay

    # input nodes
    self.stem = nn.Sequential(Fan_Out(self._n_input_states))

    self.cells = (
        nn.ModuleList()
    )  # get list of all current modules (should be empty)

    # generate a cell that undergoes architecture search
    self.cells = Cell(steps, self._n_input_states, self.primitives)

    # last layer is a linear classifier (e.g. with 10 CIFAR classes)
    self.classifier = nn.Linear(
        self._dim_output, num_classes
    )  # make this the number of input states

    # initialize classifier weights
    if train_classifier_coefficients is False:
        self.classifier.weight.data.fill_(1)
        self.classifier.weight.requires_grad = False

    if train_classifier_bias is False:
        self.classifier.bias.data.fill_(0)
        self.classifier.bias.requires_grad = False

    # initializes weights of the architecture
    self._initialize_alphas()

apply_weight_decay_to_classifier(lr)

Applies a weight decay to the weights projecting from the cell to the final output layer.

Parameters:

Name	Type	Description	Default
lr	float	learning rate	required

Source code in src/autora/theorist/darts/model_search.py

def apply_weight_decay_to_classifier(self, lr: float):
    """
    Applies a weight decay to the weights projecting from the cell to the final output layer.

    Arguments:
        lr: learning rate
    """
    # weight decay proportional to degrees of freedom
    for p in self.classifier.parameters():
        if p.requires_grad is False:
            continue
        p.data.sub_(
            self._classifier_weight_decay
            * lr
            * torch.sign(p.data)
            * (torch.abs(p.data))
        )  # weight decay

arch_parameters()

Returns architecture weights.

Returns:

Name	Type	Description
_arch_parameters	List	architecture weights.

Source code in src/autora/theorist/darts/model_search.py

def arch_parameters(self) -> List:
    """
    Returns architecture weights.

    Returns:
        _arch_parameters: architecture weights.
    """
    return self._arch_parameters

architecture_to_str_list(input_labels, output_labels, output_function_label='', decimals_to_display=2, output_format='console')

Returns a list of strings representing the model.

Parameters:

Name	Type	Description	Default
input_labels	Sequence[str]	list of strings representing the input states.	required
output_labels	Sequence[str]	list of strings representing the output states.	required
output_function_label	str	string representing the output function.	''
decimals_to_display	int	number of decimals to display.	2
output_format	Literal['latex', 'console']	if set to "console", returns equations formatted for the command line, if set to "latex", returns equations in latex format	'console'

Returns:

Type	Description
List	list of strings representing the model

Source code in src/autora/theorist/darts/model_search.py

def architecture_to_str_list(
    self,
    input_labels: Sequence[str],
    output_labels: Sequence[str],
    output_function_label: str = "",
    decimals_to_display: int = 2,
    output_format: Literal["latex", "console"] = "console",
) -> List:
    """
    Returns a list of strings representing the model.

    Arguments:
        input_labels: list of strings representing the input states.
        output_labels: list of strings representing the output states.
        output_function_label: string representing the output function.
        decimals_to_display: number of decimals to display.
        output_format: if set to `"console"`, returns equations formatted for the command line,
            if set to `"latex"`, returns equations in latex format


    Returns:
        list of strings representing the model
    """
    (n_params_total, n_params_base, param_list) = self.count_parameters(
        print_parameters=False
    )
    genotype = self.genotype().normal
    steps = self._steps
    edge_list = list()

    n = len(input_labels)
    start = 0
    for i in range(steps):  # for every node
        end = start + n
        # for k in [2*i, 2*i + 1]:

        edge_operations_list = list()
        op_list = list()

        for k in range(start, end):
            if (
                output_format == "latex"
            ):  # for every edge projecting to current node
                v = "k_" + str(i + 1)
            else:
                v = "k" + str(i + 1)
            op, j = genotype[k]
            if j < len(input_labels):
                u = input_labels[j]
            else:
                if output_format == "latex":
                    u = "k_" + str(j - len(input_labels) + 1)
                else:
                    u = "k" + str(j - len(input_labels) + 1)
            if op != "none":
                op_label = op
                params = param_list[
                    start + j
                ]  # note: genotype order and param list order don't align
                op_label = get_operation_label(
                    op,
                    params,
                    decimals=decimals_to_display,
                    input_var=u,
                    output_format=output_format,
                )
                op_list.append(op)
                edge_operations_list.append(op_label)

        if len(edge_operations_list) == 0:
            edge_str = v + " = 0"
        else:
            edge_str = ""
        for i, edge_operation in enumerate(edge_operations_list):
            if i == 0:
                edge_str += v + " = " + edge_operation
            if i > 0:
                if (
                    op_list[i] != "add"
                    and op_list[i] != "subtract"
                    and op_list[i] != "none"
                ):
                    edge_str += " +"
                edge_str += " " + edge_operation

        edge_list.append(edge_str)
        start = end
        n += 1

    # TODO: extend to multiple outputs
    if output_format == "latex":
        classifier_str = output_labels[0] + " = " + output_function_label
        if output_function_label != "":
            classifier_str += "\\left("
    else:
        classifier_str = output_labels[0] + " = " + output_function_label
        if output_function_label != "":
            classifier_str += "("

    bias = None
    for i in range(steps):
        param_idx = len(param_list) - steps + i
        tmp_param_list = param_list[param_idx]
        if i == 0 and len(tmp_param_list) == 2:
            bias = tmp_param_list[1]
        if i > 0:
            classifier_str += " + "

        if output_format == "latex":
            input_var = "k_" + str(i + 1)
        else:
            input_var = "k" + str(i + 1)

        classifier_str += get_operation_label(
            "classifier",
            tmp_param_list[0],
            decimals=decimals_to_display,
            input_var=input_var,
        )

        if i == steps - 1 and bias is not None:
            classifier_str += " + " + str(bias[0])

        if i == steps - 1:
            if output_function_label != "":
                if output_format == "latex":
                    classifier_str += "\\right)"
                else:
                    classifier_str += ")"

    edge_list.append(classifier_str)

    return edge_list

count_parameters(print_parameters=False)

Counts and returns the parameters (coefficients) of the architecture defined by the highest architecture weights.

Parameters:

Name	Type	Description	Default
print_parameters	bool	if set to true, the function will print all parameters.	False

Returns:

Name	Type	Description
n_params_total	int	total number of parameters
n_params_base	int	number of parameters determined by the classifier
param_list	list	list of parameters specifying the corresponding edge (operation) and value

Source code in src/autora/theorist/darts/model_search.py

def count_parameters(self, print_parameters: bool = False) -> Tuple[int, int, list]:
    """
    Counts and returns the parameters (coefficients) of the architecture defined by the
    highest architecture weights.

    Arguments:
        print_parameters: if set to true, the function will print all parameters.

    Returns:
        n_params_total: total number of parameters
        n_params_base: number of parameters determined by the classifier
        param_list: list of parameters specifying the corresponding edge (operation)
            and value
    """

    # counts only parameters of operations with the highest architecture weight
    n_params_total = 0

    # count classifier
    for parameter in self.classifier.parameters():
        if parameter.requires_grad is True:
            n_params_total += parameter.data.numel()

    # count stem
    for parameter in self.stem.parameters():
        if parameter.requires_grad is True:
            n_params_total += parameter.data.numel()

    n_params_base = (
        n_params_total  # number of parameters, excluding individual cells
    )

    param_list = list()
    # now count number of parameters for cells that have highest
    # probability
    for idx, op in enumerate(self.cells._ops):
        # pick most operation with highest likelihood
        values = self.alphas_normal[idx, :].data.numpy()
        maxIdx = np.where(values == max(values))

        tmp_param_list = list()
        if isiterable(op._ops[maxIdx[0].item(0)]):  # Zero is not iterable

            for subop in op._ops[maxIdx[0].item(0)]:

                for parameter in subop.parameters():
                    tmp_param_list.append(parameter.data.numpy().squeeze())
                    if parameter.requires_grad is True:
                        n_params_total += parameter.data.numel()

        if print_parameters:
            print(
                "Edge ("
                + str(idx)
                + "): "
                + get_operation_label(
                    self.primitives[maxIdx[0].item(0)], tmp_param_list
                )
            )
        param_list.append(tmp_param_list)

    # # get parameters from final linear classifier
    # tmp_param_list = list()
    # for parameter in self.classifier.parameters():
    #   for subparameter in parameter:
    #     tmp_param_list.append(subparameter.data.numpy().squeeze())

    # get parameters from final linear for each edge
    for edge in range(self._steps):
        tmp_param_list = list()
        # add weight
        tmp_param_list.append(
            self.classifier._parameters["weight"].data[:, edge].numpy()
        )
        # add partial bias (bias of classifier units will be divided by
        # number of edges)
        if "bias" in self.classifier._parameters.keys() and edge == 0:
            tmp_param_list.append(self.classifier._parameters["bias"].data.numpy())
        param_list.append(tmp_param_list)

        if print_parameters:
            print(
                "Classifier from Node "
                + str(edge)
                + ": "
                + get_operation_label("classifier_concat", tmp_param_list)
            )

    return (n_params_total, n_params_base, param_list)

fix_architecture(switch, new_weights=None)

Freezes or unfreezes the architecture weights.

Parameters:

Name	Type	Description	Default
switch	bool	set true to freeze architecture weights or false unfreeze	required
new_weights	Optional[torch.Tensor]	new set of architecture weights	None

Source code in src/autora/theorist/darts/model_search.py

def fix_architecture(
    self, switch: bool, new_weights: Optional[torch.Tensor] = None
):
    """
    Freezes or unfreezes the architecture weights.

    Arguments:
        switch: set true to freeze architecture weights or false unfreeze
        new_weights: new set of architecture weights
    """
    self._architecture_fixed = switch
    if new_weights is not None:
        self.alphas_normal = new_weights
    return

forward(x)

Computes output of the network.

Parameters:

Name	Type	Description	Default
x	torch.Tensor	input to the network	required

Source code in src/autora/theorist/darts/model_search.py

def forward(self, x: torch.Tensor):
    """
    Computes output of the network.

    Arguments:
        x: input to the network
    """

    # compute stem first
    input_states = self.stem(x)

    # get architecture weights
    if self._architecture_fixed:
        weights = self.alphas_normal
    else:
        if self.DARTS_type == DARTSType.ORIGINAL:
            weights = F.softmax(self.alphas_normal, dim=-1)
        elif self.DARTS_type == DARTSType.FAIR:
            weights = torch.sigmoid(self.alphas_normal)
        else:
            raise Exception(
                "DARTS Type " + str(self.DARTS_type) + " not implemented"
            )

    # then apply cell with weights
    cell_output = self.cells(input_states, weights)

    # compute logits
    logits = self.classifier(cell_output.view(cell_output.size(0), -1))
    # just gets output to have only 2 dimensions (batch_size x num units in
    # output layer)

    return logits

genotype(sample=False)

Computes a genotype of the model which specifies the current computation graph based on the largest architecture weight for each edge, or based on a sample. The genotype can be used for parsing or plotting the computation graph.

Parameters:

Name	Type	Description	Default
sample	bool	if set to true, the architecture will be determined by sampling from a probability distribution that is determined by the softmaxed architecture weights. If set to false (default), the architecture will be determined based on the largest architecture weight per edge.	False

Returns:

Name	Type	Description
genotype	Genotype	genotype describing the current (sampled) architecture

Source code in src/autora/theorist/darts/model_search.py

def genotype(self, sample: bool = False) -> Genotype:
    """
    Computes a genotype of the model which specifies the current computation graph based on
    the largest architecture weight for each edge, or based on a sample.
    The genotype can be used for parsing or plotting the computation graph.

    Arguments:
        sample: if set to true, the architecture will be determined by sampling
            from a probability distribution that is determined by the
            softmaxed architecture weights. If set to false (default), the architecture will be
            determined based on the largest architecture weight per edge.

    Returns:
        genotype: genotype describing the current (sampled) architecture
    """
    # this function uses the architecture weights to retrieve the
    # operations with the highest weights
    def _parse(weights):
        gene = []
        n = (
            self._n_input_states
        )  # 2 ... changed this to adapt to number of input states
        start = 0
        for i in range(self._steps):
            end = start + n
            W = weights[start:end].copy()
            # first get all the edges for a given node, edges are sorted according to their
            # highest (non-none) weight, starting from the edge with the smallest heighest
            # weight

            if "none" in self.primitives:
                none_index = self.primitives.index("none")
            else:
                none_index = -1

            edges = sorted(
                range(n),
                key=lambda x: -max(
                    W[x][k] for k in range(len(W[x])) if k != none_index
                ),
            )
            # for each edge, figure out which is the primitive with the
            # highest
            for (
                j
            ) in edges:  # looping through all the edges for the current node (i)
                if sample:
                    W_soft = F.softmax(Variable(torch.from_numpy(W[j])))
                    k_best = np.random.choice(
                        range(len(W[j])), p=W_soft.data.numpy()
                    )
                else:
                    k_best = None
                    # looping through all the primitives
                    for k in range(len(W[j])):
                        # choose the primitive with the highest weight
                        # if k != self.primitives.index('none'):
                        # EDIT SM 01/13: commented to include "none"
                        # weights in genotype
                        if k_best is None or W[j][k] > W[j][k_best]:
                            k_best = k
                    # add gene (primitive, edge number)
                gene.append((self.primitives[k_best], j))
            start = end
            n += 1
        return gene

    if self._architecture_fixed:
        gene_normal = _parse(self.alphas_normal.data.cpu().numpy())
    else:
        gene_normal = _parse(
            F.softmax(self.alphas_normal, dim=-1).data.cpu().numpy()
        )

    concat = range(2 + self._steps - self._multiplier, self._steps + 2)
    genotype = Genotype(
        normal=gene_normal,
        normal_concat=concat,
    )
    return genotype

max_alphas_normal()

Samples an architecture from the mixed operations by selecting, for each edge, the operation with the largest architecture weight.

Returns:

Name	Type	Description
alphas_normal_sample	torch.Tensor	sampled architecture weights.

Source code in src/autora/theorist/darts/model_search.py

def max_alphas_normal(self) -> torch.Tensor:
    """
    Samples an architecture from the mixed operations by selecting, for each edge,
    the operation with the largest architecture weight.

    Returns:
        alphas_normal_sample: sampled architecture weights.
    """
    alphas_normal = self.alphas_normal.clone()
    alphas_normal_sample = Variable(torch.zeros(alphas_normal.data.shape))

    for edge in range(alphas_normal.data.shape[0]):
        row = alphas_normal[edge]
        max_idx = np.argmax(row.data)
        alphas_normal_sample[edge, max_idx] = 1

    return alphas_normal_sample

new()

Returns a copy of the network.

Returns:

Type	Description
nn.Module	a copy of the network

Source code in src/autora/theorist/darts/model_search.py

def new(self) -> nn.Module:
    """
    Returns a copy of the network.

    Returns:
        a copy of the network

    """

    model_new = Network(
        # self._C, self._num_classes, self._criterion, steps=self._steps
        num_classes=self._num_classes,
        criterion=self._criterion,
        steps=self._steps,
        n_input_states=self._n_input_states,
        architecture_fixed=self._architecture_fixed,
        classifier_weight_decay=self._classifier_weight_decay,
        darts_type=self.DARTS_type,
        primitives=self.primitives,
    )

    for x, y in zip(model_new.arch_parameters(), self.arch_parameters()):
        x.data.copy_(y.data)
    return model_new

sample_alphas_normal(sample_amp=1, fair_darts_weight_threshold=0)

Samples an architecture from the mixed operations from a probability distribution that is defined by the (softmaxed) architecture weights. This amounts to selecting one operation per edge (i.e., setting the architecture weight of that operation to one while setting the others to zero).

Parameters:

Name	Type	Description	Default
sample_amp	float	temperature that is applied before passing the weights through a softmax	1
fair_darts_weight_threshold	float	used in fair DARTS. If an architecture weight is below this value then it is set to zero.	0

Returns:

Name	Type	Description
alphas_normal_sample	torch.Tensor	sampled architecture weights.

Source code in src/autora/theorist/darts/model_search.py

def sample_alphas_normal(
    self, sample_amp: float = 1, fair_darts_weight_threshold: float = 0
) -> torch.Tensor:
    """
    Samples an architecture from the mixed operations from a probability distribution that is
    defined by the (softmaxed) architecture weights.
    This amounts to selecting one operation per edge (i.e., setting the architecture
    weight of that operation to one while setting the others to zero).

    Arguments:
        sample_amp: temperature that is applied before passing the weights through a softmax
        fair_darts_weight_threshold: used in fair DARTS. If an architecture weight is below
            this value then it is set to zero.

    Returns:
        alphas_normal_sample: sampled architecture weights.
    """

    alphas_normal = self.alphas_normal.clone()
    alphas_normal_sample = Variable(torch.zeros(alphas_normal.data.shape))

    for edge in range(alphas_normal.data.shape[0]):
        if self.DARTS_type == DARTSType.ORIGINAL:
            W_soft = F.softmax(alphas_normal[edge] * sample_amp, dim=0)
        elif self.DARTS_type == DARTSType.FAIR:
            transformed_alphas_normal = alphas_normal[edge]
            above_threshold = False
            for idx in range(len(transformed_alphas_normal.data)):
                if (
                    torch.sigmoid(transformed_alphas_normal).data[idx]
                    > fair_darts_weight_threshold
                ):
                    above_threshold = True
                    break
            if above_threshold:
                W_soft = F.softmax(transformed_alphas_normal * sample_amp, dim=0)
            else:
                W_soft = Variable(torch.zeros(alphas_normal[edge].shape))
                W_soft[self.primitives.index("none")] = 1

        else:
            raise Exception(
                "DARTS Type " + str(self.DARTS_type) + " not implemented"
            )

        if torch.any(W_soft != W_soft):
            warnings.warn(
                "Cannot properly sample from architecture weights due to nan entries."
            )
            k_sample = random.randrange(len(W_soft))
        else:
            k_sample = np.random.choice(range(len(W_soft)), p=W_soft.data.numpy())
        alphas_normal_sample[edge, k_sample] = 1

    return alphas_normal_sample