efficientnet_custom#
This module contains custom layers for predicting Dirichlet concentrations.
The module name is unhelpful - it used to be about EfficientNet, but not any more. Sorry.
- zoobot.pytorch.estimators.efficientnet_custom.custom_top_dirichlet(input_dim, output_dim)#
Final dense layer used in GZ DECaLS (after global pooling).
output_dimneurons with an activation oftf.nn.sigmoid(x) * 100. + 1., chosen to ensure 1-100 output range This range is suitable for parameters of Dirichlet distribution.- Parameters
output_dim (int) – Dimension of dense layer e.g. 34 for decision tree with 34 answers
- Returns
nn.Linear followed by 1-101 sigmoid activation
- Return type
nn.Sequential