Convolutional Classifier

The classifier is polymorphic in the number of spatial axes. Its input has shape (batch, channels, spatial...); convolution and pooling use the same vector-valued configuration for signals, images, volumes, and higher-dimensional data.

structure TorchLean.nn.models.CnnConfig (d : ℕ) : Type

Configuration for a compact convolutional classifier.

• 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.

• outDim : ℕ

  Number of classifier outputs per sample.

• conv : Conv d

  Convolution stage.

• pool : Pool d

  Max-pooling stage.

def TorchLean.nn.models.CnnConfig.afterConv {d : ℕ} (cfg : CnnConfig d) : Vector ℕ d

Spatial extent after convolution.

def TorchLean.nn.models.CnnConfig.afterPool {d : ℕ} (cfg : CnnConfig d) : Vector ℕ d

Spatial extent after pooling.

def TorchLean.nn.models.CnnConfig.featureCount {d : ℕ} (cfg : CnnConfig d) : ℕ

Number of features presented to the classifier head.

def TorchLean.nn.models.cnnInShape {d : ℕ} (cfg : CnnConfig d) : Spec.Shape

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

def TorchLean.nn.models.cnnOutShape {d : ℕ} (cfg : CnnConfig d) : Spec.Shape

Classifier output shape (batch, outDim).

def TorchLean.nn.models.cnn {d : ℕ} (cfg : CnnConfig d) (hInChannels : cfg.inChannels ≠ 0 := by decide) : M (Sequential (cnnInShape cfg) (cnnOutShape cfg))

Build convolution -> activation -> max pool -> flatten -> linear.