Residual Classifier Training Example

This command trains the public rank-polymorphic residual classifier on a small CIFAR-10 crop. The model contains a convolutional stem, two residual branches, global average pooling over all spatial axes, and a linear classifier. The compact crop keeps the command useful as a runtime check while still exercising the same residual composition and pooling code used by larger configurations.

def NN.Examples.Models.Vision.ResNet.exeName : String

CLI subcommand name used in terminal banners and parser errors.

def NN.Examples.Models.Vision.ResNet.defaultLogPath : System.FilePath

Default JSON training-log path.

def NN.Examples.Models.Vision.ResNet.batchSize : ℕ

Static minibatch size carried by the checked model shape.

def NN.Examples.Models.Vision.ResNet.inputChannels : ℕ

CIFAR input channels.

def NN.Examples.Models.Vision.ResNet.cropHeight : ℕ

Height of the compact CIFAR crop.

def NN.Examples.Models.Vision.ResNet.cropWidth : ℕ

Width of the compact CIFAR crop.

def NN.Examples.Models.Vision.ResNet.hiddenChannels : ℕ

Channel width of the residual trunk.

@[reducible, inline]

abbrev NN.Examples.Models.Vision.ResNet.modelConfig : TorchLean.nn.models.ResNet.Config 2

Complete residual-classifier architecture.

spatial := [8, 8] fixes the checked input grid. The stem and both residual stages use 3 x 3 same-padding convolutions. Each stage has an identity shortcut, so its branch preserves the grid and channel width. The constructor derives the intermediate and output types from this geometry.

@[reducible, inline]

abbrev NN.Examples.Models.Vision.ResNet.input : Spec.Shape

Batched channel-first input type derived from modelConfig.

@[reducible, inline]

abbrev NN.Examples.Models.Vision.ResNet.output : Spec.Shape

One row of class logits per input sample, also derived from modelConfig.

def NN.Examples.Models.Vision.ResNet.model : TorchLean.nn.Builder (TorchLean.nn.Sequential input output)

Residual classifier from the public model API.

def NN.Examples.Models.Vision.ResNet.train (runtime : TorchLean.Runtime.Config) (flags : Support.Training.Options Support.Npy.Options) : IO TorchLean.Trainer.Report

Train the residual classifier with the public classification trainer.

def NN.Examples.Models.Vision.ResNet.main (args : List String) : IO UInt32

CLI entrypoint for the CIFAR residual-classifier training path.