tl.zoo.utils

Model blocks, utility functions and custom layers for model building.

Classes

AttentionPool1D	AttentionPool from the FastISM repository.
MultiheadAttention	Creates a MultiheadAttention module.

Functions

activate(current, activation[, verbose, name])	Apply activation function to a tensor.
conv_block(inputs, filters, kernel_size, ...)	Convolution building block.
conv_block_bs(inputs[, filters, ...])	Construct a convolution block (for Basenji/Enformer).
dense_block(inputs, units, activation[, ...])	Dense building block.
dilated_residual(inputs, filters[, ...])	Construct a residual dilated convolution block.
ffn_block_enf(inputs, filters[, ...])	Construct a feedforward block (for Enformer), consisting of Residual(LayerNorm+PointwiseConv+Dropout+ReLU+PointwiseConv+Dropout).
gelu_approx(x)	Wrap around keras.activations.gelu with approximate = True.
gelu_enf(x)	Very simple gelu approximation, used in Enformer, so needed to get equivalent results.
get_output(input_layer, hidden_layers)	Pass input layer through hidden layers.
mha_block_enf(inputs, num_heads, ...[, ...])	Construct a MHA block (for Enformer/Borzoi), consisting of Residual(LayerNorm+MHSelfAttention+Dropout).
pool(current, pool_type[, pool_size, ...])	Apply activation function to a tensor.