Position Weight Generator

Sample sequences from position-specific logit distributions

This generator is open source. Any third-party models, product names, or trademarks referenced are the property of their respective owners, and Proto is not affiliated with them.

---

Source

proto-bio/proto-language/proto_language/generator/position_weight_generator.py

View source

Convert logit distributions into discrete proposal sequences.

Reads seq.logits from each proposal sequence, applies softmax, and writes the decoded string back to seq.sequence. Supports deterministic argmax decoding or stochastic categorical sampling.

Designed for gradient-based optimizers that update seq.logits directly and need discrete sequences for handoff or tracking.

API Reference

ConfigPositionWeightGeneratorConfig Source

Configuration for position-specific sequence sampling.

This generator is intended for optimizers that own position-specific sequence distributions and need to materialize discrete proposal sequences for the rest of the framework.

sampling_mode · enum · default: "argmax"

How to convert position-specific distributions into discrete proposals.

Options: argmax, categorical

temperature · number · default: "1.0"

Softmax temperature on logits before decoding. Below 1 sharpens; above 1 flattens. Must be > 0.

sequence_bias · SequenceLogitBiasConfig

Optional declarative sequence-symbol bias applied before decoding.

logit_scale · number · default: "1.0"

Optional scale factor applied to logits before adding sequence_bias and decoding.

peak_probability_positions · array

Position indices used in the mean per-position peak-probability metric; None averages over all.

Usage

python

>>> segment = Segment(sequence="ACGT", sequence_type="dna")
>>> gen = PositionWeightGenerator(PositionWeightGeneratorConfig(sampling_mode="argmax"))
>>> gen.assign(segment)
>>> segment.proposal_sequences[0].logits = np.array([[5, 0, 0, 0], [0, 4, 0, 0], [0, 0, 3, 0], [0, 0, 0, 2]])
>>> gen.sample()
>>> segment.proposal_sequences[0].sequence
'ACGT'

Metadata

Property	Value
Key	position-weight
Class	PositionWeightGenerator
Category	gradient
Input Type	logits
Uses GPU	False
Supported Sequence Types	dna, rna, protein
Allows Empty Start	False