Supervised Samples

Typed input-target records used by datasets and trainers.

Application code works with the named input and target fields. The runtime conversion to its heterogeneous graph-argument representation is confined to Sample.Internal.

structure TorchLean.Sample.Supervised (α : Type) [Storage α] (σ τ : Shape) : Type

A supervised input-target sample with both shapes tracked statically.

• input : Tensor α σ

  Model input.

• target : Tensor α τ

  Expected model output.

def TorchLean.Sample.instReprSupervised.repr {α✝ : Type} {inst✝ : Storage α✝} {σ✝ τ✝ : Shape} [Repr α✝] : Supervised α✝ σ✝ τ✝ → ℕ → Std.Format

@[instance_reducible]

instance TorchLean.Sample.instReprSupervised {α✝ : Type} {inst✝ : Storage α✝} {σ✝ τ✝ : Shape} [Repr α✝] : Repr (Supervised α✝ σ✝ τ✝)

@[reducible, inline]

abbrev TorchLean.Sample.Batch (α : Type) [Storage α] (n : ℕ) (σ τ : Shape) : Type

A fixed-size minibatch whose tensors share the leading dimension n.

def TorchLean.Sample.mapInput {α : Type} [Storage α] {σ σ' τ : Shape} (f : Tensor α σ → Tensor α σ') (sample : Supervised α σ τ) : Supervised α σ' τ

Map the input tensor, optionally changing its shape.

def TorchLean.Sample.mapTarget {α : Type} [Storage α] {σ τ τ' : Shape} (f : Tensor α τ → Tensor α τ') (sample : Supervised α σ τ) : Supervised α σ τ'

Map the target tensor, optionally changing its shape.

def TorchLean.Sample.map {α β : Type} [Storage α] [Storage β] {σ τ σ' τ' : Shape} (mapInput : Tensor α σ → Tensor β σ') (mapTarget : Tensor α τ → Tensor β τ') (sample : Supervised α σ τ) : Supervised β σ' τ'

Map both tensors, optionally changing their element type and shapes.

def TorchLean.Sample.Internal.arguments {α : Type} [Storage α] {σ τ : Shape} (sample : Supervised α σ τ) : Arguments α [σ, τ]

Convert a supervised record to the generic argument representation used by graph runtimes.