Add k-means multi-center reference for similarity gate
Browse files- model.py: predict_image accepts 2-D ref_emb (cluster centers);
sim = max cosine similarity to any center; thres default 0.55
- main.py: load cluster_centers.npy if present, else avg12_embedding.npy
- scripts/compute_cluster_centers.py: extract embeddings from all
training images, k-means k=5, save cluster_centers.npy
Run script once to generate cluster_centers.npy, then redeploy.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- app/main.py +9 -4
- app/model.py +14 -9
- scripts/compute_cluster_centers.py +118 -0
app/main.py
CHANGED
|
@@ -36,10 +36,15 @@ async def lifespan(app: FastAPI):
|
|
| 36 |
app.state.model = model # Store the model in `app.state` for later access
|
| 37 |
logger.info("Model loaded successfully!")
|
| 38 |
|
| 39 |
-
#
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
app.state.ref_emb = ref_emb
|
| 44 |
app.state.sim_thres = SIM_THRES
|
| 45 |
|
|
|
|
| 36 |
app.state.model = model # Store the model in `app.state` for later access
|
| 37 |
logger.info("Model loaded successfully!")
|
| 38 |
|
| 39 |
+
# 優先載入 cluster_centers.npy(k-means 多中心點)
|
| 40 |
+
# fallback: avg12_embedding.npy(舊版單中心點)
|
| 41 |
+
cluster_path = os.path.join(resource_dir, "cluster_centers.npy")
|
| 42 |
+
ref_emb = np.load(cluster_path if os.path.exists(cluster_path) else REF_EMB_PATH).astype(np.float32)
|
| 43 |
+
# 舊版 avg12 shape=(1,1536) → squeeze 成 1-D;cluster_centers shape=(k,1536) 保持 2-D
|
| 44 |
+
if ref_emb.ndim == 2 and ref_emb.shape[0] == 1:
|
| 45 |
+
ref_emb = ref_emb.squeeze(0)
|
| 46 |
+
mode = "cluster_centers" if (ref_emb.ndim == 2) else "avg_single"
|
| 47 |
+
logger.info(f"Reference embedding loaded: {mode}, shape={ref_emb.shape}")
|
| 48 |
app.state.ref_emb = ref_emb
|
| 49 |
app.state.sim_thres = SIM_THRES
|
| 50 |
|
app/model.py
CHANGED
|
@@ -10,10 +10,14 @@ def cosine_similarity(a, b):
|
|
| 10 |
b = b / (norm(b) + 1e-8)
|
| 11 |
return float(np.dot(a, b))
|
| 12 |
|
| 13 |
-
def predict_image(model, image, ref_emb, thres=0.
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
inp = model.get_inputs()[0].name
|
| 16 |
-
out_map = {o.name: i for i, o in enumerate(model.get_outputs())}
|
| 17 |
|
| 18 |
outs = model.run(None, {inp: image})
|
| 19 |
|
|
@@ -21,15 +25,16 @@ def predict_image(model, image, ref_emb, thres=0.75):
|
|
| 21 |
if missing:
|
| 22 |
raise KeyError(f"Model output missing expected keys: {missing}. Found: {list(out_map.keys())}")
|
| 23 |
|
| 24 |
-
emb
|
| 25 |
prob = float(outs[out_map["probability"]].squeeze())
|
| 26 |
|
| 27 |
-
#
|
| 28 |
-
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
|
| 31 |
-
|
| 32 |
-
return is_true, prob, sim
|
| 33 |
|
| 34 |
if __name__ == "__main__":
|
| 35 |
from utils import preprocess_image
|
|
|
|
| 10 |
b = b / (norm(b) + 1e-8)
|
| 11 |
return float(np.dot(a, b))
|
| 12 |
|
| 13 |
+
def predict_image(model, image, ref_emb, thres=0.55):
|
| 14 |
+
"""
|
| 15 |
+
ref_emb: 1-D (single reference) or 2-D (k cluster centers, shape [k, dim]).
|
| 16 |
+
When 2-D, sim = max cosine similarity to any cluster center.
|
| 17 |
+
Returns (is_in_criteria: bool, prob: float, sim: float).
|
| 18 |
+
"""
|
| 19 |
inp = model.get_inputs()[0].name
|
| 20 |
+
out_map = {o.name: i for i, o in enumerate(model.get_outputs())}
|
| 21 |
|
| 22 |
outs = model.run(None, {inp: image})
|
| 23 |
|
|
|
|
| 25 |
if missing:
|
| 26 |
raise KeyError(f"Model output missing expected keys: {missing}. Found: {list(out_map.keys())}")
|
| 27 |
|
| 28 |
+
emb = outs[out_map["embedding"]].squeeze().astype(np.float32)
|
| 29 |
prob = float(outs[out_map["probability"]].squeeze())
|
| 30 |
|
| 31 |
+
# Multi-center: take max similarity to any cluster center
|
| 32 |
+
if ref_emb.ndim == 2:
|
| 33 |
+
sim = float(max(cosine_similarity(emb, c) for c in ref_emb))
|
| 34 |
+
else:
|
| 35 |
+
sim = cosine_similarity(emb, ref_emb)
|
| 36 |
|
| 37 |
+
return sim >= thres, prob, sim
|
|
|
|
| 38 |
|
| 39 |
if __name__ == "__main__":
|
| 40 |
from utils import preprocess_image
|
scripts/compute_cluster_centers.py
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
compute_cluster_centers.py
|
| 3 |
+
--------------------------
|
| 4 |
+
Extract embeddings from all training images, run k-means (k=5),
|
| 5 |
+
save cluster centers to resource/cluster_centers.npy.
|
| 6 |
+
|
| 7 |
+
Usage (from repo root):
|
| 8 |
+
python scripts/compute_cluster_centers.py
|
| 9 |
+
|
| 10 |
+
Requirements: onnxruntime, opencv-python, numpy, scikit-learn
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
import os
|
| 14 |
+
import sys
|
| 15 |
+
import numpy as np
|
| 16 |
+
import cv2
|
| 17 |
+
from pathlib import Path
|
| 18 |
+
|
| 19 |
+
# ── Config ────────────────────────────────────────────────────────────────────
|
| 20 |
+
REPO_ROOT = Path(__file__).parent.parent
|
| 21 |
+
DATA_DIR = Path("/mnt/d/pig/data") # training images root
|
| 22 |
+
MODEL_PATH = REPO_ROOT / "resource" / "model_2c_v1_emb.onnx"
|
| 23 |
+
OUT_PATH = REPO_ROOT / "resource" / "cluster_centers.npy"
|
| 24 |
+
K = 5
|
| 25 |
+
IMG_EXTS = {".jpg", ".jpeg", ".png", ".bmp", ".tiff", ".tif"}
|
| 26 |
+
# ──────────────────────────────────────────────────────────────────────────────
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def preprocess(image_path: str) -> np.ndarray:
|
| 30 |
+
img = cv2.imread(image_path)
|
| 31 |
+
if img is None:
|
| 32 |
+
return None
|
| 33 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
| 34 |
+
img = cv2.resize(img, (600, 600))
|
| 35 |
+
img = img / 255.0
|
| 36 |
+
mean = np.array([0.485, 0.456, 0.406])
|
| 37 |
+
std = np.array([0.229, 0.224, 0.225])
|
| 38 |
+
img = (img - mean) / std
|
| 39 |
+
img = np.transpose(img, (2, 0, 1))
|
| 40 |
+
return np.expand_dims(img, 0).astype(np.float32)
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def extract_embedding(sess, image_path: str) -> np.ndarray | None:
|
| 44 |
+
x = preprocess(image_path)
|
| 45 |
+
if x is None:
|
| 46 |
+
return None
|
| 47 |
+
inp_name = sess.get_inputs()[0].name
|
| 48 |
+
out_map = {o.name: i for i, o in enumerate(sess.get_outputs())}
|
| 49 |
+
if "embedding" not in out_map:
|
| 50 |
+
raise KeyError(f"Model has no 'embedding' output. Outputs: {list(out_map)}")
|
| 51 |
+
outs = sess.run(None, {inp_name: x})
|
| 52 |
+
return outs[out_map["embedding"]].squeeze().astype(np.float32)
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
def collect_images(root: Path) -> list[str]:
|
| 56 |
+
paths = []
|
| 57 |
+
for p in root.rglob("*"):
|
| 58 |
+
if p.suffix.lower() in IMG_EXTS:
|
| 59 |
+
paths.append(str(p))
|
| 60 |
+
return sorted(paths)
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
def main():
|
| 64 |
+
import onnxruntime as ort
|
| 65 |
+
from sklearn.cluster import KMeans
|
| 66 |
+
|
| 67 |
+
print(f"Model : {MODEL_PATH}")
|
| 68 |
+
print(f"Data : {DATA_DIR}")
|
| 69 |
+
print(f"Output : {OUT_PATH}")
|
| 70 |
+
print(f"k : {K}")
|
| 71 |
+
print()
|
| 72 |
+
|
| 73 |
+
sess = ort.InferenceSession(str(MODEL_PATH),
|
| 74 |
+
providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
|
| 75 |
+
|
| 76 |
+
images = collect_images(DATA_DIR)
|
| 77 |
+
print(f"Found {len(images)} images — extracting embeddings...")
|
| 78 |
+
|
| 79 |
+
embeddings, failed = [], []
|
| 80 |
+
for i, path in enumerate(images, 1):
|
| 81 |
+
emb = extract_embedding(sess, path)
|
| 82 |
+
if emb is not None:
|
| 83 |
+
embeddings.append(emb)
|
| 84 |
+
else:
|
| 85 |
+
failed.append(path)
|
| 86 |
+
if i % 100 == 0 or i == len(images):
|
| 87 |
+
print(f" {i}/{len(images)} ok={len(embeddings)} fail={len(failed)}", end="\r")
|
| 88 |
+
|
| 89 |
+
print(f"\nExtracted {len(embeddings)} embeddings ({len(failed)} failed)")
|
| 90 |
+
if failed:
|
| 91 |
+
print("Failed files:")
|
| 92 |
+
for f in failed[:5]:
|
| 93 |
+
print(f" {f}")
|
| 94 |
+
|
| 95 |
+
X = np.stack(embeddings) # (N, 1536)
|
| 96 |
+
print(f"\nRunning k-means (k={K}) on {X.shape} ...")
|
| 97 |
+
km = KMeans(n_clusters=K, random_state=42, n_init=10)
|
| 98 |
+
km.fit(X)
|
| 99 |
+
|
| 100 |
+
centers = km.cluster_centers_.astype(np.float32) # (K, 1536)
|
| 101 |
+
# L2-normalise so cosine similarity == dot product
|
| 102 |
+
norms = np.linalg.norm(centers, axis=1, keepdims=True)
|
| 103 |
+
centers = centers / (norms + 1e-8)
|
| 104 |
+
|
| 105 |
+
np.save(str(OUT_PATH), centers)
|
| 106 |
+
print(f"Saved {centers.shape} cluster centers → {OUT_PATH}")
|
| 107 |
+
|
| 108 |
+
# Quick sanity: show avg max-sim of training images to centers
|
| 109 |
+
sims = X / (np.linalg.norm(X, axis=1, keepdims=True) + 1e-8)
|
| 110 |
+
max_sims = (sims @ centers.T).max(axis=1)
|
| 111 |
+
print(f"\nSanity check (training set max-sim to nearest center):")
|
| 112 |
+
print(f" mean={max_sims.mean():.3f} min={max_sims.min():.3f} "
|
| 113 |
+
f"p5={np.percentile(max_sims, 5):.3f} p10={np.percentile(max_sims, 10):.3f}")
|
| 114 |
+
print("Done.")
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
if __name__ == "__main__":
|
| 118 |
+
main()
|