PatchEmbedHelper

Helpers & Utilities · AiDotNet.Helpers

ViT-style patch embedding (Dosovitskiy et al. 2020 §3.1): turns raw NCHW image input into a BSC token sequence for transformer-backed vision encoders (PaLM-E, BiomedCLIP, PaLI-3, OpenCLIP, ViT). Already-tokenized inputs pass through unchanged.

Methods

Method Summary
Tensor<T> TokenizeImageNCHWToBSC(Tensor<T> input, int visionDim, int imageSize, ConvolutionalLayer<T> patchEmbed, IEngine engine)