TorchLean NN: Convolution and Pooling Layers

def Runtime.Autograd.Model.Layers.conv (batchSize rank inputChannels outputChannels : ℕ) (kernelSize stride padding inputSize : TorchLean.Tensor ℕ [rank]) (weightSeed : ℕ := 0) (weightInit : Torch.Init.Scheme := Torch.Init.Scheme.uniform (-0.1) 0.1) : Layer (((inputSize.to Spec.Shape).prependDim inputChannels).prependDim batchSize) ((((Spec.convOutSpatial inputSize kernelSize stride padding).to Spec.Shape).prependDim outputChannels).prependDim batchSize)

N-D convolution layer for a channels-first tensor (batch, inputChannels, spatial...).

Parameters:

• weight: (outputChannels × inputChannels × kernelSize[0] × ... × kernelSize[rank-1]),
• bias: (outputChannels).

The output spatial shape is computed from (stride, padding, kernelSize).

PyTorch analogy: torch.nn.Conv{d}d / torch.nn.functional.conv{d}d with groups=1 and dilation=1.

def Runtime.Autograd.Model.Layers.convTranspose (batchSize rank inputChannels outputChannels : ℕ) (kernelSize stride padding inputSize : TorchLean.Tensor ℕ [rank]) (weightSeed : ℕ := 0) (weightInit : Torch.Init.Scheme := Torch.Init.Scheme.uniform (-0.1) 0.1) : Layer (((inputSize.to Spec.Shape).prependDim inputChannels).prependDim batchSize) ((((Spec.convTransposeOutSpatial inputSize kernelSize stride padding).to Spec.Shape).prependDim outputChannels).prependDim batchSize)

N-D transpose convolution layer for a channels-first tensor (batch, inputChannels, spatial...).

Parameters:

• weight: (inputChannels × outputChannels × kernelSize[0] × ... × kernelSize[rank-1]),
• bias: (outputChannels).

The output spatial shape uses: output[a] = (input[a] - 1) * stride[a] - 2 * padding[a] + kernelSize[a] (with output_padding = 0).

PyTorch analogy: torch.nn.ConvTranspose{d}d / torch.nn.functional.conv_transpose{d}d with groups=1, dilation=1, and output_padding=0.

def Runtime.Autograd.Model.Layers.maxPool (batchSize rank channels : ℕ) (kernelSize stride padding inputSize : TorchLean.Tensor ℕ [rank]) : Layer (((inputSize.to Spec.Shape).prependDim channels).prependDim batchSize) ((((Spec.poolOutSpatialPad inputSize kernelSize stride padding).to Spec.Shape).prependDim channels).prependDim batchSize)

N-D max pooling layer for a channels-first tensor (batch, channels, spatial...).

Output spatial dimensions follow Spec.poolOutSpatialPad.

PyTorch analogy: torch.nn.functional.max_pool{d}d on an N×C×... tensor.

def Runtime.Autograd.Model.Layers.avgPool (batchSize rank channels : ℕ) (kernelSize stride padding inputSize : TorchLean.Tensor ℕ [rank]) : Layer (((inputSize.to Spec.Shape).prependDim channels).prependDim batchSize) ((((Spec.poolOutSpatialPad inputSize kernelSize stride padding).to Spec.Shape).prependDim channels).prependDim batchSize)

N-D average pooling layer for a channels-first tensor (batch, channels, spatial...).

PyTorch analogy: torch.nn.functional.avg_pool{d}d on an N×C×... tensor.