CNN PyTorch Reference Import

CNN reference weight import from a PyTorch-style state_dict.

We mirror the common PyTorch naming convention for modules:

• conv1.weight, conv1.bias
• conv2.weight, conv2.bias
• fc.weight, fc.bias

Each tensor is expected to be encoded as nested JSON arrays whose shape matches the expected TorchLean Shape.

structure Import.PyTorch.CNN.Parameters (inputChannels outputChannels kernelHeight kernelWidth flattenedWidth : ℕ) : Type

Parameters for the example two-block CNN imported from a PyTorch state_dict.

This matches the keys used by the exporter (conv1.*, conv2.*, fc.*) and pins down the exact shapes expected by TorchLean.

• firstConvolutionWeight : TorchLean.Tensor Float [outputChannels, inputChannels, kernelHeight, kernelWidth]

  First convolution kernel, in PyTorch (output, input, height, width) layout.

• firstConvolutionBias : TorchLean.Tensor Float [outputChannels]

  First convolution bias.

• secondConvolutionWeight : TorchLean.Tensor Float [outputChannels, outputChannels, kernelHeight, kernelWidth]

  Second convolution kernel.

• secondConvolutionBias : TorchLean.Tensor Float [outputChannels]

  Second convolution bias.

• classifierWeight : TorchLean.Tensor Float [outputChannels, flattenedWidth]

  Classifier weight, in PyTorch (output, input) layout.

• classifierBias : TorchLean.Tensor Float [outputChannels]

  Classifier bias.

def Import.PyTorch.CNN.load (inputChannels outputChannels kernelHeight kernelWidth flattenedWidth : ℕ) (json : Lean.Json) : Option (Parameters inputChannels outputChannels kernelHeight kernelWidth flattenedWidth)

Load CNN parameters from JSON using PyTorch state_dict keys.