KosmosOptions

Options & Configuration · AiDotNet.VisionLanguage.Generative

Hyperparameters shared by KOSMOS-1 and KOSMOS-2:

  • A CLIP ViT-L/14 image encoder: 224 px, 1024 wide, 24 layers, 16 heads, patch 14.
  • A 64-token image resampler.
  • A MAGNETO causal decoder: 2048 wide, 24 layers, 32 heads, FFN 8192, 2048 positions.

Inheritance: ModelOptions → GenerativeVLMOptions → KosmosOptions

Known derived types: KOSMOS1Options, KOSMOS2Options

For Beginners

The defaults reproduce the published models. The token ids of the image markers must match the tokenizer used to train the model.

Properties

Property Summary
int BosTokenId { get; set; } Gets or sets the begin-of-sequence token id.
int DecoderFeedForwardDim { get; set; } Gets or sets the decoder feed-forward width.
int EosTokenId { get; set; } Gets or sets the end-of-sequence token id.
int ImageEndTokenId { get; set; } Gets or sets the </image> token id.
int ImageStartTokenId { get; set; } Gets or sets the <image> token id.
int NumImageTokens { get; set; } Gets or sets the number of image embeddings placed in the sequence (latent queries).
int PatchSize { get; set; } Gets or sets the vision encoder’s patch size.
int VisionHeads { get; set; } Gets or sets the vision encoder’s attention heads.