File size: 1,071 Bytes
ebeed67 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | """CLIP image embedding postprocessor.
Input: raw embedding tensor from a CLIP vision encoder (e.g. MobileCLIP S0/S1/S2)
Output: {"kind": "embedding", "embedding": [...], "embeddingNorm": [...]}
MobileCLIP uses resize 256×256 + rescale /255 only (no ImageNet mean/std).
The existing inference_pool resize_image path already produces this layout —
preprocessMode:'resize' + inputNormalization:'none' is the correct config.
"""
import numpy as np
def postprocess_clip(
predictions: dict,
config: dict,
orig_w: int,
orig_h: int,
scale: float,
pad: tuple[int, int],
) -> dict:
"""L2-normalize the CLIP embedding vector."""
# Get the first output tensor (vision encoder emits a single embedding)
raw = np.array(list(predictions.values())[0]).flatten().astype(np.float32)
norm = np.linalg.norm(raw)
if norm == 0:
normalized = np.zeros_like(raw)
else:
normalized = raw / norm
return {
"kind": "embedding",
"embedding": raw.tolist(),
"embeddingNorm": normalized.tolist(),
}
|