wingbeats.modelling package

Submodules

wingbeats.modelling.builds module

Library for factory functions to define model architectures within the Functional API

wingbeats.modelling.builds.build_embedder(in_shape, out_shape, f_extractor, reg_param=0.001, input_name='input_signal', model_name='Embedder', training=None)[source]

Build model for learning hierarchical class embeddings.

Architectures

f_extractor(Layer) + Dense + L2

Outputs

predicted embedding

Parameters

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Embedder’.

  • training (bool, optional) – Whether to run model in training mode. Particularly relevant for layers such as Batch Normalization and Dropout. Defaults to None.

Returns

Embedder

wingbeats.modelling.builds.build_hiera_classifier(in_shape, out_shapes, f_extractor, reg_param=0.001, taxonomic_levels=['genus', 'species'], parallel=True, input_name='input_signal', model_name='Hiera_Classifier', training=None)[source]

Build model for classifying signals according to more than one taxonomic level.

Architectures (layers in brackets are branched out)

  • (series) f_extractor(Layer) + Dense(+ Softmax) + DenseBlock(+ Softmax) … + DenseBlock + Softmax

  • (parallel) f_extractor(Layer) (+ Dense + Softmax) (+ Dense + Softmax) …

Outputs

predicted class probabilities

Parameters

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. The model only predicts one but multiple superior levels can be inferred from the predicted one and penalized in the loss. Defaults to [‘genus’, ‘species’].

  • parallel (bool) – Whether to attach parallel Dense layers for every prediction. Otherwise, they are connected one after another. Defaults to True.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hiera_Classifier’.

  • training (bool, optional) – Whether to run model in training mode. Particularly relevant for layers such as Batch Normalization and Dropout. Defaults to None.

Returns

Hierarchical Classifier

wingbeats.modelling.builds.build_hiera_embedder_classifier(in_shape, out_shapes, f_extractor, reg_param=0.001, taxonomic_levels=['genus', 'species'], parallel=True, input_name='input_signal', model_name='Hiera_Embedder_Classifier', training=None)[source]

Build model for learning embeddings and classifying signals according to more than one taxonomic level.

Architectures (layers in brackets are branched out)

  • (series) f_extractor(Layer) + Dense (+ L2) + DenseBlock(+ Softmax) + … + DenseBlock + Softmax

  • (parallel) f_extractor(Layer) + Dense (+ L2) (+ DenseBlock + Softmax) …

Outputs

predicted embedding and class probabilities

Parameters

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. The model only predicts one but multiple superior levels can be inferred from the predicted one and penalized in the loss. Defaults to [‘genus’, ‘species’].

  • parallel (bool) – Whether to attach parallel Dense layers for every prediction. Otherwise, they are connected one after another. Defaults to True.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hiera_Embedder_Classifier’.

  • training (bool, optional) – Whether to run model in training mode. Particularly relevant for layers such as Batch Normalization and Dropout. Defaults to None.

Returns

Hierarchical Embedder Classifier

wingbeats.modelling.builds.build_simple_classifier(in_shape, out_shape, f_extractor, reg_param=0.001, taxonomic_levels=['species'], input_name='input_signal', model_name='Simple_Classifier', training=None)[source]

Build model for classifying signals according to only one taxonomic level i.e. genus or species. It is possible to extend the loss function to penalize the model for getting wrong higher hierarchies, as well (just add them to the tax_levels list).

Architectures

f_extractor(Layer) + Dense + Softmax

Outputs

predicted class probabilities

Parameters

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. The model only predicts one but multiple superior levels can be inferred from the predicted one and penalized in the loss. Defaults to [‘species’].

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Simple_Classifier’.

  • training (bool, optional) – Whether to run model in training mode. Particularly relevant for layers such as Batch Normalization and Dropout. Defaults to None.

Returns

Simple Classifier

wingbeats.modelling.builds.build_simple_embedder_classifier(in_shape, out_shapes, f_extractor, reg_param=0.001, taxonomic_levels=['species'], input_name='input_signal', model_name='Simple_Embedder_Classifier', training=None)[source]

Build model for learning the embedding of one taxonomic level and classifying signals according to only one taxonomic level (does not have to coincide to the embedding). It is possible to extend the loss function to penalize the model for getting wrong higher hierarchies, as well (just add them to the tax_levels list).

Architectures (layers in brackets are branched out)

f_extractor(Layer) + Dense(+ L2) + DenseBlock + Softmax

Outputs

predicted embedding and class probabilities

Parameters

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. The model only predicts one but multiple superior levels can be inferred from the predicted one and penalized in the loss. Defaults to [‘species’].

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Simple_Embedder_Classifier’.

  • training (bool, optional) – Whether to run model in training mode. Particularly relevant for layers such as Batch Normalization and Dropout. Defaults to None.

Returns

Simple Embedder Classifier

wingbeats.modelling.builds.embed(x, out_shape, reg_param=0.001, apply_l2=False)[source]

Compute embedding of a vector by passing it through a Dense layer and l2-normalizing it (meant to be used as module in models with embedding layers).

Parameters

  • x (Tensor) – Input features.

  • out_shape (int) – Dense layer output shape.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • apply_l2 (bool) – Whether to l2-normalize the features output by the Dense layer. Defaults to False.

Returns

Predicted embedding.

wingbeats.modelling.builds.predict_prob(x, out_shape, reg_param=0.001, taxonomic_level='species', add_softmax=True, as_block=False)[source]

Compute (normalized) probabilities of signal belonging to different classes of specified taxonomic level.

Parameters

  • x (Tensor) – Input features.

  • out_shape (int) – Dense layer output shape.

  • reg_param (float) – Regularization parameter for the weights in the Dense layer. Defaults to 0.001.

  • taxonomic_level (str) – Predicted taxonomic level. Defaults to ‘species’.

  • add_softmax (bool) – Whether to normalize the probabilities with a Softmax layer. Defaults to True.

  • as_block (bool) – Whether to pass x through a simple Dense layer or a Dense block. See wingbeats.modelling.layers. Defaults to False.

Returns

Predicted probability vector.

wingbeats.modelling.callbacks module

Library for model callbacks to be used during training

wingbeats.modelling.callbacks.lr_exp(epoch, max_lr, max_ep=30)[source]

Define an exponentially decaying learning rate schedule.

Parameters

  • epoch (int) – Current epoch.

  • max_lr (float) – Maximum learning rate.

  • max_ep (int) – How many epochs the pattern should repeat before the learning rate stays constant. Defaults to 30.

Returns

Learning rate

Return type

float

wingbeats.modelling.callbacks.lr_triangle(epoch, max_lr, max_ep=30, step_size=15)[source]

Define a decreasing triangular learning rate schedule. Each isosceles triangle represents a new cycle with length 2 * step_size.

Parameters

  • epoch (int) – Current epoch.

  • max_lr (float) – Maximum learning rate.

  • max_ep (int) – How many epochs the pattern should repeat before the learning rate stays constant. Defaults to 30.

  • step_size (int) – Number of epochs until monotony reverses (half of triangle). Defaults to 15.

Returns

Learning rate

Return type

float

wingbeats.modelling.hypertuning module

Library for hyperparameter optimization functions

wingbeats.modelling.hypertuning.build_hyper_embedder(hp, in_shape, out_shape, f_extractor, lr_values, reg_values, emb_matrix, input_name='input_signal', model_name='Hyper_Embedder', strategy=None)[source]

Build Embedder for Hyperband-optimization.

Parameters

  • hp (kerastuner.hyperband) – Hyperband object.

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • lr_values (list) – Discrete learning rate values.

  • reg_values (list) – Discrete regularization parameter values.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Defaults to None.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hyper_Embedder’.

  • strategy (Strategy from tf.distribute, optional) – Distribution strategy (CPU, GPU, TPU). Defaults to None (CPU).

Returns

Hyperband-optimizable Embedder

wingbeats.modelling.hypertuning.build_hyper_hiera_classifier(hp, in_shape, out_shape, f_extractor, lr_values, reg_values, taxonomic_levels=['genus', 'species'], parallel=True, input_name='input_signal', model_name='Hyper_Hiera_Classifier', strategy=None)[source]

Build Hierarchical Classifier for Hyperband-optimization.

Parameters

  • hp (kerastuner.hyperband) – Hyperband object.

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • lr_values (list) – Discrete learning rate values.

  • reg_values (list) – Discrete regularization parameter values.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. Defaults to [‘genus’, ‘species’].

  • parallel (bool) – Whether to attach parallel Dense layers for every prediction. Otherwise, they are connected one after another. Defaults to True.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hyper_Hiera_Classifier’.

  • strategy (Strategy from tf.distribute, optional) – Distribution strategy (CPU, GPU, TPU). Defaults to None (CPU).

Returns

Hyperband-optimizable Hierarchical Classifier

wingbeats.modelling.hypertuning.build_hyper_hiera_embedder_classifier(hp, in_shape, out_shape, f_extractor, lr_values, reg_values, emb_matrix, taxonomic_levels=['genus', 'species'], parallel=True, input_name='input_signal', model_name='Hyper_Hiera_Embedder_Classifier', strategy=None)[source]

Build Hierarchical Embedder-Classifier for Hyperband-optimization.

Parameters

  • hp (kerastuner.hyperband) – Hyperband object.

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • lr_values (list) – Discrete learning rate values.

  • reg_values (list) – Discrete regularization parameter values.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Defaults to None.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. Defaults to [‘genus’, ‘species’].

  • parallel (bool) – Whether to attach parallel Dense layers for every prediction. Otherwise, they are connected one after another. Defaults to True.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hyper_Hiera_Embedder_Classifier’.

  • strategy (Strategy from tf.distribute, optional) – Distribution strategy (CPU, GPU, TPU). Defaults to None (CPU).

Returns

Hyperband-optimizable Hierarchical Embedder-Classifier

wingbeats.modelling.hypertuning.build_hyper_simple_classifier(hp, in_shape, out_shape, f_extractor, lr_values, reg_values, taxonomic_levels=['species'], input_name='input_signal', model_name='Hyper_Simple_Classifier', strategy=None)[source]

Build Simple Classifier for Hyperband-optimization.

Parameters

  • hp (kerastuner.hyperband) – Hyperband object.

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • lr_values (list) – Discrete learning rate values.

  • reg_values (list) – Discrete regularization parameter values.

  • taxonomic_levels (list) – Taxonomic levels to include in the loss function. Defaults to [‘species’].

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hyper_Simple_Classifier’.

  • strategy (Strategy from tf.distribute, optional) – Distribution strategy (CPU, GPU, TPU). Defaults to None (CPU).

Returns

Hyperband-optimizable Simple Classifier

wingbeats.modelling.hypertuning.build_hyper_simple_embedder_classifier(hp, in_shape, out_shape, f_extractor, lr_values, reg_values, emb_matrix, taxonomic_levels=['species'], input_name='input_signal', model_name='Hyper_Simple_Embedder_Classifier', strategy=None)[source]

Build Simple Embedder Classifier for Hyperband-optimization.

Parameters

  • hp (kerastuner.hyperband) – Hyperband object.

  • in_shape (tuple) – Input shape. No need to specify batch dimension.

  • out_shape (tuple) – Model output shape (here equal to the size of embedded taxonomic level).

  • f_extractor (tf.Layer) – Feature extractor.

  • lr_values (list) – Discrete learning rate values.

  • reg_values (list) – Discrete regularization parameter values.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Defaults to None.

  • input_name (str) – Name of the input. Defaults to ‘input_signal’.

  • model_name (str) – Name of the architecture. Defaults to ‘Hyper_Simple_Embedder_Classifier’.

  • strategy (Strategy from tf.distribute, optional) – Distribution strategy (CPU, GPU, TPU). Defaults to None (CPU).

Returns

Hyperband-optimizable Simple Embedder Classifier

wingbeats.modelling.hypertuning.kfold_cv(X, y, models, genus_mapping, emb_matrix=None, n_splits=4, epochs=30, batch_size=64, sm=None, model_callbacks=None, sampling_rate=16000, window=None, nperseg=None, noverlap=None, cutoff=None)[source]

Execute KFold Cross-Validation on dataset (X, y) which is to be split into n_splits stratified folds. User should provide n_splits models to be trained on each fold. Each model name should follow the pattern architecture_inputFormat (e.g. HieraCls_spectro). The n_splits-1 folds held for training are also augmented using SMOTE, if sm is not None.

Parameters

  • X (list) – Matrix of signals.

  • y (list) – Label vector.

  • genus_mapping (list) – List containing genus indexes ef every species.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Defaults to None. Only needed for Embedder models.

  • n_splits (int) – Number of folds. Defaults to 4.

  • epochs (int) – Number of epochs to train models each fold. Defaults to 30.

  • batch_size (int) – Size of one signal batch in tf.Dataset. Defaults to 64.

  • sm (imblearn.smote, optional) – SMOTE object to augment data. Defaults to None.

  • model_callbacks (list) – Callbacks for model training (e.g. Early Stopping, Model Checkpoint, Learning Rate Schedules). Defaults to None.

  • sampling_rate (int) – Sampling frequency. Defaults to 16000.

  • window (str (for psd) or function pointer (for spectrograms)) – Window-function to multiply each segment with i.e. ‘hann’ (for psd) or tf.signal.hann_window (for spectrograms). Defaults to None.

  • nperseg (int) – Length of a segment for applying the Welch-Transform or STFT. Defaults to None.

  • noverlap (int) – Lenth of overlapping region between segments. Defaults to None.

  • cutoff (int, optional) – How many PSD frequencies should be kept. Defaults to None.

Returns

Mean and std. dev. confusion matrices over all folds for genus and species and list of training histories for every fold.

wingbeats.modelling.layers module

Library for custom layers

class wingbeats.modelling.layers.CNN1D(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a CNN made up of 5 ConvBlock1d’s of increasing filter sizes.

Parameters

  • drop_rate (float) – Dropout rate (between 0 and 1).

  • mcdrop (bool, optional) – Whether to apply Monte-Carlo Dropout. Defaults to False.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, len_signal, 1).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.CNN2D(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a CNN made up of 5 ConvBlock2d’s of increasing filter sizes.

Parameters

  • drop_rate (float) – Dropout rate (between 0 and 1).

  • mcdrop (bool, optional) – Whether to apply Monte-Carlo Dropout. Defaults to False.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, height, width, 3).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.CNN_Efficient(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a CNN made up of the feature extractor of an EfficientNet.

Parameters

  • in_shape (tuple) – Input shape.

  • drop_rate (float) – Dropout rate (between 0 and 1).

  • mcdrop (bool, optional) – Whether to apply Monte-Carlo Dropout. Defaults to False.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, height, width, 3).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.CNN_Mobile(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a CNN made up of the feature extractor of a MobileNet.

Parameters

  • in_shape (tuple) – Input shape.

  • alpha (float) – Network width parameter. If alpha < 1.0, the number of filters proportionally decreases in each layer. For alpha > 1.0, it increases.

  • drop_rate (float) – Dropout rate (between 0 and 1).

  • mcdrop (bool, optional) – Whether to apply Monte-Carlo Dropout. Defaults to False.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, height, width, 3).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.ConvBlock1d(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a simple one-dimensional convolutional block. The block consists of Conv1D(num_filters, kernel_size = 3, strides = 1) + BatchNorm + Relu + MaxPool1D(pool_size = ", strides = 2).

Parameters

num_filters (int) – Number of convolutional filters.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, len_signal, 1).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.ConvBlock2d(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a simple two-dimensional convolutional block. The block consists of Conv2D(num_filters, kernel_size = 3, strides = 1) + BatchNorm + Relu + MaxPool1D(pool_size = ", strides = 2).

Parameters

num_filters (int) – Number of convolutional filters.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch of shape (batch_size, height, width, 3).

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.DenseBlock(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Class for a simple Dense block. The block consists of BatchNorm + Relu + Dense(units).

Parameters

  • units (int) – Number of nodes in the Dense layer.

  • reg_param (float) – L2-Regularization factor.

call(inputs, training=None)[source]

Apply forward propagation to inputs.

Parameters

  • inputs (Tensor) – Tensor batch.

  • training (bool, optional) – Whether to run in training mode (relevant for Batch Normalization). Defaults to None.

Returns

Transformed inputs.

get_config()[source]

Get configuration (needed for serialization when extra arguments are provided in __init__()).

Returns

Configuration dictionary.

Return type

dict

class wingbeats.modelling.layers.Identity(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

Identity layer (just returns input unmodified)

call(inputs)[source]

Apply identity function on inputs.

Parameters

inputs (Tensor) – Tensor batch.

Returns

The same inputs.

class wingbeats.modelling.layers.L2_Norm(*args, **kwargs)[source]

Bases: tensorflow.python.keras.engine.base_layer.Layer

L2-Normalization layer (meant for hierarchical embedding output).

call(inputs)[source]

Apply L2-Normalization on inputs.

Parameters

inputs (Tensor) – Tensor batch.

Returns

Transformed inputs.

wingbeats.modelling.metrics module

Library for custom metrics and losses

wingbeats.modelling.metrics.embedding_loss(emb_matrix)[source]

Compute embedding loss of current batch as 1.0 - embedding_similarity.

Parameters

emb_matrix (array) – Matrix of hierarchical embeddings.

Returns

Function that computes the embedding loss w.r.t. true and predicted embeddings.

wingbeats.modelling.metrics.embedding_similarity(emb_matrix)[source]

Compute embedding similarity of current batch.

Parameters

emb_matrix (array) – Matrix of hierarchical embeddings.

Returns

Function that computes similarity between true and predicted embeddings.

wingbeats.modelling.metrics.focal_loss(gamma=2.0)[source]

Compute focal loss as modified cross entropy loss. The goal is to penalize hard examples harsher.

Parameters

gamma (float) – Penalty exponent. If gamma is 0.0, the focal loss becomes the standard cross entropy loss. Defaults to 2.0.

Returns

Focal loss function w.r.t. true and predicted probabilities.

wingbeats.modelling.metrics.get_species_from_embeddings(pred_embs, genus_mapping, emb_matrix)[source]

Infer predicted species from predicted embeddings.

Parameters

  • pred_embs (array) – Predicted embeddings to be compared via nearest neighbor to the true embeddings.

  • genus_mapping (list) – List that maps the index of the species to the index of the genus.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Only needed for Emb.

Returns

Predicted species.

Return type

list

wingbeats.modelling.metrics.predict_gen_spec(model, X, model_name, genus_mapping, emb_matrix)[source]

Make genus and species predictions on dataset X according to model architecture.

Parameters

  • model (tf.Model) – Pretrained classifier.

  • X (tf.Dataset) – Matrix of signals.

  • model_name (str) – Name of the architecture. Currently only allowed: SimpleCls, Emb, SimpleEmbCls, HieraCls, HieraEmbCls.

  • genus_mapping (list) – List that maps the index of the species to the index of the genus.

  • emb_matrix (array) – Matrix of hierarchical embeddings. Only needed for Emb.

Returns

Predicted genus and species

Return type

tuple

Module contents