ViTBackbone constructor

ViTBackbone({ required int imageSize, required int patchSize, int numChannels = 3, required int embedSize, int numLayers = 4, int numHeads = 4, })

Implementation

ViTBackbone({
  required this.imageSize,
  required this.patchSize,
  this.numChannels = 3,
  required this.embedSize,
  int numLayers = 4,
  int numHeads = 4,
}) : transformerEncoder = TransformerEncoder(
       vocabSize: 1, // Not used for vision
       embedSize: embedSize,
       blockSize: (imageSize ~/ patchSize) * (imageSize ~/ patchSize) + 1,
       numLayers: numLayers,
       numHeads: numHeads,
     ) {
  int numPatches = (imageSize ~/ patchSize) * (imageSize ~/ patchSize);

  // 1. Projection layer (GPU weights initialized via Layer class)
  patchProjection = Layer(
    patchSize * patchSize * numChannels,
    embedSize,
    useGelu: false,
  );

  // 2. Learnable [CLS] token [1, embedSize]
  clsToken = Tensor.random([1, embedSize], scale: 0.02);

  // 3. Positional Embeddings [N+1, embedSize]
  posEmbeddings = Tensor.random([numPatches + 1, embedSize], scale: 0.02);
}