BinClass.Model (BinClass v0.3.0)

Copy Markdown View Source

Dispatcher for named model architectures.

Adding a new architecture

  1. Create a new module e.g. BinClass.Model.NewArch in lib/bin_class/model/new_arch.ex.
  2. Implement build/2 in that module, using the shared internal model configuration helpers for option validation and seeded dropout.
  3. Add a new clause to BinClass.Model.build/3:
    def build(:new_arch, vocab_size, opts), do: BinClass.Model.NewArch.build(vocab_size, opts)
  4. Update the training configuration's default architecture if it should become the default.

Summary

Functions

architectures()

Returns the supported architecture names.

build(architecture, vocab_size, opts \\ [])

Builds a named binary-classification model.

Functions

architectures()

Returns the supported architecture names.

build(architecture, vocab_size, opts \\ [])

Builds a named binary-classification model.

architecture must be one of the values returned by architectures/0, and vocab_size must be a positive integer greater than :pad_token_id. The returned Axon graph accepts batches of token IDs and produces two-class probability distributions.

Common options

  • :embedding_size - Positive embedding width. Defaults to 64.
  • :dropout_rate - Number from 0.0 up to, but excluding, 1.0. Defaults to 0.2.
  • :pad_token_id - Non-negative padding ID below vocab_size. Defaults to 0.
  • :seed - Optional integer used to seed dropout layers. Defaults to nil.

Architecture-specific options

  • :conv_filters - Positive filter count for :cnn and :cnn_mixed_pooling. Defaults to 128 and 96, respectively.
  • :branch_filters - Positive per-branch filter count for :multi_scale_cnn and :sep_se_cnn. Defaults to 32 and 48, respectively.
  • :ff_dim - Positive feed-forward width for :transformer. Defaults to 128.