ConvSpec

The geometry of a 2-D convolution over batched, channel-major images.

Everything stays rank-2, which is the shape the rest of the library already speaks: one image per row, channels laid out one after another.

  input    (batch, inChannels * inHeight * inWidth)
  weights  (outChannels, inChannels * kernel * kernel)
  output   (batch, outChannels * outHeight * outWidth)

Square kernels only, and no bias: a bias is one broadcast add away and does not need to be inside the convolution.

View source ↗Full module reference on javadoc.io ↗

Record components

inChannels: int
inHeight: int
inWidth: int
outChannels: int
kernel: int
stride: int
pad: int

Methods

int outHeight()

int outWidth()

int inputSize()

Elements per input image: inChannels * inHeight * inWidth.

int outputSize()

Elements per output image: outChannels * outHeight * outWidth.

int weightSize()

Elements per filter: inChannels * kernel * kernel.

Shape weightShape()

Shape inputShape(int batch)

Shape outputShape(int batch)

ConvSpec next(int nextOutChannels, int nextKernel, int nextStride, int nextPad)

The spec of a layer that consumes this layer's output.