papi_cc / app /main.py
CJR61's picture
Claude Sonnet 4.6
Replace B3 ONNX with EfficientNet-B0 (3-fold CV, orig fold_3)
a5ec00a
Raw History Blame Contribute Delete
5.98 kB
# 20251015 update for open data classification
from fastapi import FastAPI, UploadFile, File
from app.model import load_model, predict_image
from app.utils import save_uploaded_file, append_to_csv
from app.log import logger, configure_logger
from app.dashboard.router import router as dashboard_router
from app.auth import check_predict_key
from fastapi import Depends
import os
from app.utils import preprocess_image, is_valid_image
from contextlib import asynccontextmanager
import numpy as np
SIM_THRES = 0.35 # Similarity threshold (lowered from 0.55; cluster_centers.npy covers broader distribution)
@asynccontextmanager
async def lifespan(app: FastAPI):
# Set model loading path
current_directory = os.path.dirname(os.path.abspath(__file__))
parent_folder = os.path.dirname(current_directory)
resource_dir = os.path.join(parent_folder, "resource")
app.state.database_dir = os.path.join(parent_folder, "database")
app.state.uploads_dir = os.path.join(parent_folder, "uploads")
app.state.logs_dir = os.path.join(parent_folder, "logs")
configure_logger(app.state.logs_dir)
# MODEL_PATH = os.path.join(parent_folder, "resource", "model_2c_v1.onnx")
# 改:使用 emb 版 ONNX,才能同時拿到 embedding + prob
MODEL_PATH = os.path.join(resource_dir, "model_b0_balanced_v2_emb.onnx")
REF_EMB_PATH = os.path.join(resource_dir, "avg12_embedding.npy")
# Load model
logger.info("Starting application... Loading model.")
model = load_model(MODEL_PATH)
app.state.model = model # Store the model in `app.state` for later access
logger.info("Model loaded successfully!")
# 優先載入 cluster_centers.npy(k-means 多中心點)
# fallback: avg12_embedding.npy(舊版單中心點)
cluster_path = os.path.join(resource_dir, "cluster_centers.npy")
ref_emb = np.load(cluster_path if os.path.exists(cluster_path) else REF_EMB_PATH).astype(np.float32)
# 舊版 avg12 shape=(1,1536) → squeeze 成 1-D;cluster_centers shape=(k,1536) 保持 2-D
if ref_emb.ndim == 2 and ref_emb.shape[0] == 1:
ref_emb = ref_emb.squeeze(0)
mode = "cluster_centers" if (ref_emb.ndim == 2) else "avg_single"
logger.info(f"Reference embedding loaded: {mode}, shape={ref_emb.shape}")
app.state.ref_emb = ref_emb
app.state.sim_thres = SIM_THRES
yield # Enter running state, FastAPI will start accepting requests at this point
# Release resources
logger.info("Shutting down application... Cleaning up resources.")
del app.state.model
logger.info("Model resources cleaned up.")
# Initialize app and bind lifespan
app = FastAPI(lifespan=lifespan)
app.include_router(dashboard_router)
@app.get("/")
async def read_root():
return {"message": "Connection successful!"}
# Inference API
@app.post("/predict/", dependencies=[Depends(check_predict_key)])
async def predict(file: UploadFile = File(...)):
# Save uploaded image and generate ID
try:
saved_path, image_id = save_uploaded_file(app.state.uploads_dir, file)
except ValueError as e:
return {"error": str(e)}
logger.info(f"Image saved: {saved_path}, ID: {image_id}")
# Check if the image is valid
if not is_valid_image(saved_path):
logger.error(f"Invalid image file - ID: {image_id}, Path: {saved_path}")
os.remove(saved_path) # Delete invalid file
return {"error": "Invalid image file, please upload a valid image format"}
# Preprocess image
try:
preprocessed_image = preprocess_image(saved_path)
except Exception as e:
logger.error(f"Image preprocessing failed - ID: {image_id}, Error: {str(e)}")
os.remove(saved_path)
return {"error": f"Image preprocessing failed: {str(e)}"}
# Inference
try:
model = app.state.model
ref_emb = app.state.ref_emb
thres = app.state.sim_thres
# 新版 predict_image:回傳 (is_in_criteria: bool, prob: float, sim: float)
is_in_criteria, probability, sim = predict_image(model, preprocessed_image, ref_emb, thres=thres)
# 先把 positive 類別的原始機率四捨五入(保留原本行為)
probability = round(float(probability), 2)
# 所屬類別機率(bacteria prob 直接用;virus prob = 1 - bacteria prob)
class_prob = probability if probability > 0.5 else round(1 - probability, 2)
# 高信心 bypass:class_prob >= HIGH_CONF 且 sim >= MIN_SIM 時直接分類
# 解決非典型拍攝角度(如臟器黑底標本)被 sim gate 誤擋的問題
# MIN_SIM 防止完全無關圖片通過 bypass(B0 模型對非豬圖片也極端自信)
HIGH_CONF = 0.85
MIN_SIM_BYPASS = 0.24
if class_prob >= HIGH_CONF and sim >= MIN_SIM_BYPASS:
is_in_criteria = True # bypass sim gate
if not is_in_criteria:
prediction = "None"
else:
if probability > 0.5:
prediction = "bacteria"
else:
prediction = "virus"
probability = round(1 - probability, 2)
logger.info(
f"Inference result - ID: {image_id}, Class: {prediction}, "
f"Probability: {probability:.2f}, ClassProb: {class_prob:.2f}, Sim: {sim:.3f}"
)
except Exception as e:
logger.error(f"Model inference failed - ID: {image_id}, Error: {str(e)}")
return {"error": f"Model inference failed: {str(e)}"}
# Save result to CSV
try:
csv_path = os.path.join(app.state.database_dir, "inference_results.csv")
append_to_csv(csv_path, image_id, prediction, probability, sim)
except Exception as e:
logger.error(f"Failed to save result to CSV: {str(e)}")
return {"error": f"Failed to save result to CSV: {str(e)}"}
# Return inference result
return {"id": image_id, "class": prediction, "probability": probability}