TorchLean API

Docs HomeInstallationGuideExamplesGraphsTools

NN.API.Models.ResNet

Residual Convolutional Classifier #

The model is polymorphic in spatial rank. Residual branches operate on a common typed shape, and global average pooling reduces every spatial axis before the classifier head.

source
structure TorchLean.nn.models.ResNetConfig (d : ℕ) :
Type

Configuration for a residual classifier over d spatial axes.

  • batch : ℕ

    Number of independent samples processed together.

  • inChannels : ℕ

    Number of channels in each input sample.

  • spatial : Vector ℕ d

    Extent of each spatial axis.

  • spatialNonzero (i : Fin d) : self.spatial.get i ≠ 0

    Spatial axes are nonempty, as required by global average pooling.

  • hiddenChannels : ℕ

    Channel width used by the residual trunk.

  • numClasses : ℕ

    Number of classifier logits per sample.

Instances For
    source
    def TorchLean.nn.models.resnetInShape {d : ℕ} (cfg : ResNetConfig d) :
    Spec.Shape

    Input tensor shape (batch, inChannels, spatial...).

    Instances For
      source
      def TorchLean.nn.models.resnetHiddenShape {d : ℕ} (cfg : ResNetConfig d) :
      Spec.Shape

      Activation shape shared by the residual branches.

      Instances For
        source
        def TorchLean.nn.models.resnetOutShape {d : ℕ} (cfg : ResNetConfig d) :
        Spec.Shape

        Classifier output shape (batch, numClasses).

        Instances For
          source
          def TorchLean.nn.models.Internal.sameSpatialConv {d batch : ℕ} (spatial : Vector ℕ d) (spatialNonzero : ∀ (i : Fin d), spatial.get i ≠ 0) (inChannels outChannels : ℕ) [NeZero inChannels] :
          M (Sequential (Spec.Shape.dim batch (Spec.Shape.ofList (inChannels :: spatial.toList))) (Spec.Shape.dim batch (Spec.Shape.ofList (outChannels :: spatial.toList))))

          Shape-preserving convolution used by the residual trunk.

          Instances For
            source
            def TorchLean.nn.models.resnet {d : ℕ} (cfg : ResNetConfig d) (hInChannels : cfg.inChannels ≠ 0 := by decide) (hHiddenChannels : cfg.hiddenChannels ≠ 0 := by decide) :
            M (Sequential (resnetInShape cfg) (resnetOutShape cfg))

            Build a convolutional stem, two residual blocks, global pooling, and a linear classifier.

            Instances For