preprocessing.json

{
  "model": "relu-clip efflite4 distilled from CLIP ViT-L/14",
  "input": {
    "size": 224,
    "layout": "NHWC",
    "channels": "RGB",
    "resize": "shortest side to 224, bicubic",
    "crop": "center 224x224",
    "scale": "divide by 255 to [0,1], then (x - mean) / std",
    "mean": [
      0.48145466,
      0.4578275,
      0.40821073
    ],
    "std": [
      0.26862954,
      0.26130258,
      0.27577711
    ],
    "quantization": {
      "dtype": "int8",
      "scale": 0.015443762764334679,
      "zero_point": -12,
      "formula": "q = clip(round(x / scale + zero_point), -128, 127)"
    }
  },
  "output": {
    "embed_dim": 768,
    "normalised": false,
    "note": "L2 normalisation is deliberately left out of the graph so the accelerator subgraph stays free of reductions; do it on the host after dequantising.",
    "quantization": {
      "dtype": "int8",
      "scale": 0.3109022378921509,
      "zero_point": -5,
      "formula": "x = (q - zero_point) * scale"
    }
  },
  "text_embeddings": {
    "imagenet1k_text_emb_vitl14.npz": {
      "keys": [
        "embs",
        "labels"
      ],
      "shape": [
        1000,
        768
      ],
      "normalised": false,
      "note": "average over the 80 OpenAI ImageNet templates; row norms 0.81-0.93, normalise before use"
    },
    "demo_prompt_emb_vitl14.npz": {
      "keys": [
        "text_emb",
        "labels"
      ],
      "shape": [
        11,
        768
      ],
      "normalised": true,
      "note": "prompts encoded verbatim, no templating"
    }
  },
  "classification": "cosine(L2norm(image_emb), L2norm(text_emb)) -> argmax"
}