Apppublic

jonilaserson/sam-encoder-api

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes

App README

SAM Encoder API

Public microservice that runs the SAM ViT-B vision encoder and returns raw feature embeddings as a compressed numpy file.

API

`POST /embed`

Upload a map image, receive SAM encoder features.

Request: multipart/form-data with field image (PNG/JPEG/WebP, ≤1024px recommended)

Response: application/octet-stream — numpy .npz file with keys:

—neck: float16 (256, 64, 64) — SAM neck output (last_hidden_state)
—early: float16 (768, 64, 64) — ViT block-6 output (BCHW)

`GET /health`

Returns {"status": "ok", "device": "cpu"}.

Usage

python

import requests, io, numpy as np

resp = requests.post(
    "https://your-username-sam-encoder-api.hf.space/embed",
    files={"image": open("map.png", "rb")},
)
data = np.load(io.BytesIO(resp.content))
neck  = data["neck"]   # (256, 64, 64) float16
early = data["early"]  # (768, 64, 64) float16

Model

Uses facebook/sam-vit-base (Meta AI). No fine-tuning — purely the frozen pretrained encoder. Model weights are downloaded from HuggingFace Hub at container build time.