API reference#

This document provides the API documentation for the DeepPeak package.

class WaveNet(sequence_length: int, num_filters: int = 64, num_dilation_layers: int = 6, kernel_size: int = 3, optimizer: str | Optimizer = 'adam', loss: str | Loss = 'binary_crossentropy', metrics: Tuple[str | Metric] = 'accuracy', seed: int | None = None)[source]#

Bases: BaseClassifier

WaveNet-style 1D detector for per-timestep peak classification.

Parameters:

  • sequence_length (int) – Length of the input sequences.

  • num_filters (int) – Number of filters in the convolutional layers.

  • num_dilation_layers (int) – Number of dilated convolutional layers.

  • kernel_size (int) – Size of the convolutional kernels.

  • optimizer (Union[str, tf.keras.optimizers.Optimizer]) – Optimizer for model compilation.

  • loss (Union[str, tf.keras.losses.Loss]) – Loss function for model training.

  • metrics (Tuple[Union[str, tf.keras.metrics.Metric]]) – Metrics for model evaluation.

  • seed (Optional[int]) – Random seed for reproducibility.

Notes

Architecture:

  • Input projection (1x1) to num_filters channels

  • Stack of dilated causal Conv1D blocks with exponentially increasing dilation (1, 2, 4, …, 2^(L-1)), residual connections, and skip connections

  • Aggregated skip path -> ReLU -> 1x1 sigmoid for per-step probability

Notes

  • Output shape: (batch, sequence_length, 1), sigmoid probabilities

  • Loss: binary_crossentropy (per time-step)

  • This class encapsulates build/fit/evaluate/predict and plotting utilities

class DenseNet(sequence_length: int, filters: Tuple[int, int, int] = (32, 64, 128), dilation_rates: Tuple[int, int, int] = (1, 2, 4), kernel_size: int = 3, optimizer: str | Optimizer = 'adam', loss: str | Loss = 'binary_crossentropy', metrics: Tuple[str | Metric] = 'accuracy', seed: int | None = None)[source]#

Bases: BaseClassifier

Compact 1D ConvNet for per-timestep peak classification.

Parameters:

  • sequence_length (int) – Length of the input sequences.

  • filters (Tuple[int, int, int]) – Number of filters in each convolutional layer.

  • dilation_rates (Tuple[int, int, int]) – Dilation rates for each convolutional layer.

  • kernel_size (int) – Size of the convolutional kernels.

  • optimizer (Union[str, tf.keras.optimizers.Optimizer]) – Optimizer for model compilation.

  • loss (Union[str, tf.keras.losses.Loss]) – Loss function for model training.

  • metrics (Tuple[Union[str, tf.keras.metrics.Metric]]) – Metrics for model evaluation.

  • seed (Optional[int]) – Random seed for reproducibility.

Notes

Architecture:

  • Three 1D Conv layers with ReLU activations and exponentially increasing dilation (default: 1, 2, 4), padding=’same’

  • Final 1x1 Conv with sigmoid -> per-step probability map named ‘ROI’

Output#

  • ROI: shape (batch, sequence_length, 1) with probabilities in [0, 1]

class Autoencoder(sequence_length: int, dropout_rate: float = 0.3, filters: Tuple[int, int, int] = (32, 64, 128), kernel_size: int = 3, pool_size: int = 2, upsample_size: int = 2, optimizer: str | Optimizer = 'adam', loss: str | Loss = 'binary_crossentropy', metrics: Tuple[str | Metric] = 'accuracy', seed: int | None = None)[source]#

Bases: BaseClassifier

1D convolutional autoencoder for predicting an ROI (Region Of Interest) mask.

Parameters:

  • sequence_length (int) – Length of the input sequences.

  • dropout_rate (float) – Dropout rate for regularization.

  • filters (Tuple[int, int, int]) – Number of filters in each convolutional layer.

  • kernel_size (int) – Size of the convolutional kernels.

  • pool_size (int) – Pooling size for downsampling.

  • upsample_size (int) – Upsampling size for the decoder.

  • optimizer (Union[str, tf.keras.optimizers.Optimizer]) – Optimizer for model compilation.

  • loss (Union[str, tf.keras.losses.Loss]) – Loss function for model training.

  • metrics (Tuple[Union[str, tf.keras.metrics.Metric]]) – Metrics for model evaluation.

  • seed (Optional[int]) – Random seed for reproducibility.

Notes

Architecture:

Encoder:

  • Conv1D(f[0], K, relu, same) -> Dropout(p) -> MaxPool1D(2)

  • Conv1D(f[1], K, relu, same) -> Dropout(p) -> MaxPool1D(2)

Bottleneck:

  • Conv1D(f[2], K, relu, same) -> Dropout(p)

Decoder:

  • UpSampling1D(2) -> Conv1D(f[1], K, relu, same)

  • UpSampling1D(2) -> Conv1D(f[0], K, relu, same)

Output:

  • Conv1D(1, 1, sigmoid, name=”ROI”)

Output shape#

(batch, sequence_length, 1)

Notes

  • Loss: binary_crossentropy on the ‘ROI’ head

  • Metrics: configurable (default: accuracy)

  • The pooling/upsampling ladder assumes sequence_length divisible by 4 to reconstruct the original length exactly (with padding=’same’ this typically works well; validate with a quick model.summary()).

On this page

This Page