papi_cc / app /model.py
CJR61's picture
Claude Sonnet 4.6
Add k-means multi-center reference for similarity gate
f8bc55b
Raw History Blame Contribute Delete
2.99 kB
import onnxruntime as ort
import numpy as np
from numpy.linalg import norm
def load_model(path):
return ort.InferenceSession(path, providers=["CUDAExecutionProvider","CPUExecutionProvider"])
def cosine_similarity(a, b):
a = a / (norm(a) + 1e-8)
b = b / (norm(b) + 1e-8)
return float(np.dot(a, b))
def predict_image(model, image, ref_emb, thres=0.55):
"""
ref_emb: 1-D (single reference) or 2-D (k cluster centers, shape [k, dim]).
When 2-D, sim = max cosine similarity to any cluster center.
Returns (is_in_criteria: bool, prob: float, sim: float).
"""
inp = model.get_inputs()[0].name
out_map = {o.name: i for i, o in enumerate(model.get_outputs())}
outs = model.run(None, {inp: image})
missing = [k for k in ("embedding", "probability") if k not in out_map]
if missing:
raise KeyError(f"Model output missing expected keys: {missing}. Found: {list(out_map.keys())}")
emb = outs[out_map["embedding"]].squeeze().astype(np.float32)
prob = float(outs[out_map["probability"]].squeeze())
# Multi-center: take max similarity to any cluster center
if ref_emb.ndim == 2:
sim = float(max(cosine_similarity(emb, c) for c in ref_emb))
else:
sim = cosine_similarity(emb, ref_emb)
return sim >= thres, prob, sim
if __name__ == "__main__":
from utils import preprocess_image
MODEL_PATH = "/mnt/d/pig/api/resource/model_2c_v1_emb.onnx"
REF_EMB_PATH = "/mnt/d/pig/api/resource/avg12_embedding.npy"
IMAGE_PATH = "/mnt/d/pig/api/test/data/ec.jpg"
THRES = 0.75
sess = load_model(MODEL_PATH)
ref = np.load(REF_EMB_PATH).astype(np.float32)
if ref.ndim > 1:
ref = ref.mean(axis=0)
x = preprocess_image(IMAGE_PATH) # [1,3,H,W] float32, ImageNet norm
flag, prob, sim = predict_image(sess, x, ref, thres=THRES)
print(f"🔹 Probability = {prob:.4f}")
print(f"🔹 Similarity = {sim:.4f} (≥{THRES}? {flag})")
#################################################################################
# import onnxruntime as ort
# def load_model(MODEL_PATH):
# return ort.InferenceSession(MODEL_PATH)
# def predict_image(model, image):
# input_name = model.get_inputs()[0].name
# output_name = model.get_outputs()[0].name
# outputs = model.run([output_name], {input_name: image})
# probability = outputs[0][0][0]
# # prediction = "bacteria" if probability > 0.5 else "virus"
# return probability
# if __name__ == "__main__":
# import os
# from PIL import Image
# import numpy as np
# from utils import preprocess_image
# MODEL_PATH = "/mnt/d/pig/api/resource/model_2c_v1_tp2.onnx"
# IMAGE_DIR = "/mnt/d/pig/api/test/data/ec.jpg"
# # 載入模型
# model = load_model(MODEL_PATH)
# print("✅ Model loaded:", MODEL_PATH)
# x = preprocess_image(IMAGE_DIR)
# # 預測
# prob = predict_image(model, x)
# print(f"🔹 Prediction probability = {prob:.4f}")