AI/ML Type Utilities

Since 1.6.0

Types for machine learning and AI applications.

Overview

AI/ML Type Utilities provides types for building machine learning pipelines, neural networks, and AI inference systems. It supports tensor operations, model definitions, training configurations, and inference results.

This module enables building type-safe ML applications with proper type constraints for tensors, layers, optimizers, and training workflows.

Basic Usage

import type { Tensor, Model, Layer, Optimizer, Dataset, Prediction } from 'uni-types'

// Define tensor types
type ImageTensor = Tensor<[224, 224, 3], 'float32'>

// Define model configuration
type CNNModel = Model<ImageTensor, Prediction<string>>

// Training setup
type TrainingSetup = {
  optimizer: Optimizer
  loss: LossFunction
  metrics: MLMetric[]
}

Key Types

Tensor

Tensor type with shape and data type constraints.

type Tensor<Shape extends number[] = number[], DType extends TensorDType = 'float32'> = {
  shape: Shape
  dtype: DType
  data: number[]
}

TensorDType

Supported tensor data types.

type TensorDType = 'float32' | 'float64' | 'int32' | 'int64' | 'uint8' | 'bool' | 'string'

Model

Model type with input/output specifications.

type Model<Input = unknown, Output = unknown> = {
  name: string
  version: string
  input: Input
  output: Output
  layers: Layer[]
}

LayerType

Supported layer types.

type LayerType = 'dense' | 'conv2d' | 'conv3d' | 'lstm' | 'gru' | 'attention' | 'pooling' | 'normalization' | 'activation' | 'dropout' | 'embedding' | 'transformer'

Optimizer

Optimizer types for training.

type Optimizer = 'sgd' | 'adam' | 'adamw' | 'rmsprop' | 'adagrad' | 'adadelta' | 'nadam' | 'ftrl' | 'lbfgs'

LossFunction

Loss function types.

type LossFunction = 'mse' | 'mae' | 'crossentropy' | 'binary_crossentropy' | 'categorical_crossentropy' | 'hinge' | 'huber' | 'log_cosh'

Dataset

Dataset type for training data.

type Dataset<T = unknown> = {
  size: number
  shape: number[]
  batch(index: number): Batch<T>
  shuffle(): Dataset<T>
  split(ratio: number): [Dataset<T>, Dataset<T>]
}

Prediction

Prediction result type.

type Prediction<T = unknown> = {
  label: T
  probability: number
  confidence: Confidence
}

Related

• Functional Programming - Monadic types
• Type Inference - Type inference utilities