maml.models.dl package

Deep learning module.

class maml.models.dl.AtomSets(describer: BaseDescriber | None = None, input_dim: int | None = None, is_embedding: bool = True, n_neurons: Sequence[int] = (64, 64), n_neurons_final: Sequence[int] = (64, 64), n_targets: int = 1, activation: str = ‘relu’, embedding_vcal: int = 95, embedding_dim: int = 32, symmetry_func: list[str] | str = ‘mean’, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = (), is_classification: bool = False, **symmetry_func_kwargs)

Bases: KerasModel

This class implements the DeepSets models.

_get_data_generator(features, targets, batch_size=128, is_shuffle=True)

_predict(features: ndarray, **kwargs)

Predict the values given a set of inputs based on fitted models.

• Parameters features (np.ndarray) – array-like input features.

• Returns List of output objects.

evaluate(eval_objs, eval_targets, is_feature: bool = False, batch_size: int = 16)

Evaluate objs, targets.

• Parameters

  • eval_objs (list) – objs for evaluation

  • eval_targets (list) – target list for the corresponding objects

  • is_feature (bool) – whether x is feature matrix

  • batch_size (int) – evaluation batch size

fit(features: list | np.ndarray, targets: list | np.ndarray | None = None, val_features: list | np.ndarray | None = None, val_targets: list | np.ndarray | None = None, **kwargs)

• Parameters

  • features (list* or *np.ndarray) – Numerical input feature list or numpy array with dim (m, n) where m is the number of data and n is the feature dimension.

  • targets (list* or *np.ndarray) – Numerical output target list, or numpy array with dim (m, ).

  • val_features (list* or *np.ndarray) – validation features

  • val_targets (list* or *np.ndarray) – validation targets.

classmethod from_dir(dirname: str)

Load the models from file :param dirname: directory name :type dirname: str

Returns: object instance.

save(dirname: str)

Save the models and describers.

• Parameters dirname (str) – dirname for save

class maml.models.dl.MLP(input_dim: int | None = None, describer: BaseDescriber | None = None, n_neurons: tuple = (64, 64), activation: str = ‘relu’, n_targets: int = 1, is_classification: bool = False, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = (), **kwargs)

Bases: KerasModel

This class implements the multi-layer perceptron models.

class maml.models.dl.WeightedAverageLayer(*args, **kwargs)

Bases: Layer

Weight average the features using weights.

result= sum{w_i^a * value_i} / sum{w_i^a}

build(input_shape: Sequence)

Build the layer.

• Parameters input_shape (tuple) – input shape tuple

call(inputs: Sequence, mask: tf.Tensor | None = None)

Core logic of the layer.

• Parameters

  • inputs (tuple) – input tuple of length 3

  • mask (tf.Tensor) – not used here

static compute_output_shape(input_shape: Sequence)

Compute output shape from input shape :param input_shape: input shape list. :type input_shape: tuple/list

get_config()

Get layer configuration.

static reduce_sum(prop: Tensor, index: Tensor, perm: Sequence)

Reduce sum the tensors using index.

• Parameters

  • prop (tf.Tensor) – tensor with shape [1, n, …]

  • index (tf.Tensor) – integer tensor with shape [1, n]

  • perm (list) – permutation for transpose.

class maml.models.dl.WeightedSet2Set(*args, **kwargs)

Bases: Layer

This is a modified version from megnet.layers.readout.Set2Set. Here, in addition to taking features and indices as inputs, we also take a weight tensor. The input to the core logic is [features, weights, indices].

_lstm(h, c)

build(input_shape: Sequence)

Build the output shape from input shapes :param input_shape: input shape list. :type input_shape: tuple/list

call(inputs, mask=None)

Core logic of the layer.

• Parameters

  • inputs (tuple) – input tuple of length 3

  • mask (tf.Tensor) – not used here

compute_output_shape(input_shape)

Compute output shapes from input shapes :param input_shape: input shapes. :type input_shape: sequence of tuple

Returns: sequence of tuples output shapes

get_config()

Part of keras layer interface, where the signature is converted into a dict

• Returns configurational dictionary.

maml.models.dl._atomsets module

neural network models.

class maml.models.dl._atomsets.AtomSets(describer: BaseDescriber | None = None, input_dim: int | None = None, is_embedding: bool = True, n_neurons: Sequence[int] = (64, 64), n_neurons_final: Sequence[int] = (64, 64), n_targets: int = 1, activation: str = ‘relu’, embedding_vcal: int = 95, embedding_dim: int = 32, symmetry_func: list[str] | str = ‘mean’, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = (), is_classification: bool = False, **symmetry_func_kwargs)

Bases: KerasModel

This class implements the DeepSets models.

_get_data_generator(features, targets, batch_size=128, is_shuffle=True)

_predict(features: ndarray, **kwargs)

Predict the values given a set of inputs based on fitted models.

• Parameters features (np.ndarray) – array-like input features.

• Returns List of output objects.

evaluate(eval_objs, eval_targets, is_feature: bool = False, batch_size: int = 16)

Evaluate objs, targets.

• Parameters

  • eval_objs (list) – objs for evaluation

  • eval_targets (list) – target list for the corresponding objects

  • is_feature (bool) – whether x is feature matrix

  • batch_size (int) – evaluation batch size

fit(features: list | np.ndarray, targets: list | np.ndarray | None = None, val_features: list | np.ndarray | None = None, val_targets: list | np.ndarray | None = None, **kwargs)

• Parameters

  • features (list* or *np.ndarray) – Numerical input feature list or numpy array with dim (m, n) where m is the number of data and n is the feature dimension.

  • targets (list* or *np.ndarray) – Numerical output target list, or numpy array with dim (m, ).

  • val_features (list* or *np.ndarray) – validation features

  • val_targets (list* or *np.ndarray) – validation targets.

classmethod from_dir(dirname: str)

Load the models from file :param dirname: directory name :type dirname: str

Returns: object instance.

save(dirname: str)

Save the models and describers.

• Parameters dirname (str) – dirname for save

maml.models.dl._atomsets.construct_atom_sets(input_dim: int | None = None, is_embedding: bool = True, n_neurons: Sequence[int] = (64, 64), n_neurons_final: Sequence[int] = (64, 64), n_targets: int = 1, activation: str = ‘relu’, embedding_vcal: int = 95, embedding_dim: int = 32, symmetry_func: list[str] | str = ‘mean’, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = (), is_classification: bool = False, **symmetry_func_kwargs)

f(X) = rho(sum_{x in X} phi(x)), where X is a set. phi is implemented as a neural network and rho is a symmetry function.

todo: implement attention mechanism

• Parameters

  • input_dim (int) – input dimension, if None, then integer inputs + embedding are assumed.

  • is_embedding (bool) – whether the input should be embedded

  • n_neurons (tuple) – number of hidden-layer neurons before passing to symmetry function

  • n_neurons_final (tuple) – number of hidden-layer neurons after symmetry function

  • n_targets (int) – number of output targets

  • activation (str) – activation function

  • embedding_vcal (int) – int, embedding vocabulary

  • embedding_dim (int) – int, embedding dimension

  • symmetry_func (str) – symmetry function, choose from [‘set2set’, ‘sum’, ‘mean’, ‘max’, ‘min’, ‘prod’]

  • optimizer (str) – optimizer for the models

  • loss (str) – loss function for the models

  • compile_metrics (tuple) – metrics for validation

  • symmetry_func_kwargs (dict) – kwargs for symmetry function

maml.models.dl._keras_utils module

Keras utils for deserialize activations and otpimizers.

maml.models.dl._keras_utils.deserialize_keras_activation(activation)

serialize keras activation.

• Parameters activation (str*, dict, or *Activation object) – activation to serialize

Returns: Activation object

maml.models.dl._keras_utils.deserialize_keras_optimizer(optimizer)

serialize keras activation.

• Parameters optimizer (str*, dict, or *optimizer object) – optimizer to serialize

Returns: Activation object

maml.models.dl._layers module

Deep learning layers.

class maml.models.dl._layers.WeightedAverageLayer(*args, **kwargs)

Bases: Layer

Weight average the features using weights.

result= sum{w_i^a * value_i} / sum{w_i^a}

build(input_shape: Sequence)

Build the layer.

• Parameters input_shape (tuple) – input shape tuple

call(inputs: Sequence, mask: tf.Tensor | None = None)

Core logic of the layer.

• Parameters

  • inputs (tuple) – input tuple of length 3

  • mask (tf.Tensor) – not used here

static compute_output_shape(input_shape: Sequence)

Compute output shape from input shape :param input_shape: input shape list. :type input_shape: tuple/list

get_config()

Get layer configuration.

static reduce_sum(prop: Tensor, index: Tensor, perm: Sequence)

Reduce sum the tensors using index.

• Parameters

  • prop (tf.Tensor) – tensor with shape [1, n, …]

  • index (tf.Tensor) – integer tensor with shape [1, n]

  • perm (list) – permutation for transpose.

class maml.models.dl._layers.WeightedSet2Set(*args, **kwargs)

Bases: Layer

This is a modified version from megnet.layers.readout.Set2Set. Here, in addition to taking features and indices as inputs, we also take a weight tensor. The input to the core logic is [features, weights, indices].

_lstm(h, c)

build(input_shape: Sequence)

Build the output shape from input shapes :param input_shape: input shape list. :type input_shape: tuple/list

call(inputs, mask=None)

Core logic of the layer.

• Parameters

  • inputs (tuple) – input tuple of length 3

  • mask (tf.Tensor) – not used here

compute_output_shape(input_shape)

Compute output shapes from input shapes :param input_shape: input shapes. :type input_shape: sequence of tuple

Returns: sequence of tuples output shapes

get_config()

Part of keras layer interface, where the signature is converted into a dict

• Returns configurational dictionary.

maml.models.dl._mlp module

Multi-layer perceptron models.

class maml.models.dl._mlp.MLP(input_dim: int | None = None, describer: BaseDescriber | None = None, n_neurons: tuple = (64, 64), activation: str = ‘relu’, n_targets: int = 1, is_classification: bool = False, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = (), **kwargs)

Bases: KerasModel

This class implements the multi-layer perceptron models.

maml.models.dl._mlp.construct_mlp(input_dim: int, n_neurons: tuple = (64, 64), activation: str = ‘relu’, n_targets: int = 1, is_classification: bool = False, optimizer: str = ‘adam’, loss: str = ‘mse’, compile_metrics: tuple = ())

Constructor for multi-layer perceptron models.

• Parameters

  • input_dim (int) – input dimension, i.e., feature dimension

  • n_neurons (tuple) – list of hidden neuron sizes

  • activation (str) – activation function

  • n_targets (int) – number of targets

  • is_classification (bool) – whether the target is a classification problem

  • optimizer (str) – optimizer

  • loss (str) – loss function

  • compile_metrics (tuple) – metrics to evaluate during epochs