api_server.py

import os
import time
import numpy as np
from PIL import Image
import torchvision.transforms as transforms
from pathlib import Path

# Disable tensorflow warnings
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

from tensorflow import keras
from flask import Flask, jsonify, request, render_template
import torch

load_type = 'local'

MODEL_NAME = "yolo11_detect_best_241018_1.pt"
MODEL_DIR = "./artifacts/models"
#REPO_ID = "1vash/mnist_demo_model"

# Load the saved YOLO model into memory
if load_type == 'local':
    # 本地模型路徑
    model_path = f'{MODEL_DIR}/{MODEL_NAME}'
    if not os.path.exists(model_path):
        raise FileNotFoundError(f"Model file not found at {model_path}")
    # 使用 torch 來載入 YOLO 模型
    model = torch.load(model_path)
    model.eval()  # 設定模型為推理模式
elif load_type == 'remote_hub_download':
    from huggingface_hub import hf_hub_download

    # 從 Hugging Face Hub 下載模型
    model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_NAME)
    model = torch.load(model_path)
    model.eval()
elif load_type == 'remote_hub_from_pretrained':
    # 使用 Hugging Face Hub 預訓練的模型方式下載
    os.environ['TRANSFORMERS_CACHE'] = str(Path(MODEL_DIR).absolute())
    from huggingface_hub import from_pretrained

    model = from_pretrained(REPO_ID, filename=MODEL_NAME, cache_dir=MODEL_DIR)
    model.eval()
else:
    raise AssertionError('No load type is specified!')

# Initialize the Flask application
app = Flask(__name__)


# API route for prediction(YOLO)
@app.route('/predict', methods=['POST'])
def predict():
    """
    Predicts the class label of an input image.

    Request format:
    {
        "image": [[pixel_values_gray]]
    }

    Response format:
    {
        "label": predicted_label,
        "pred_proba" prediction class probability
        "ml-latency-ms": latency_in_milliseconds
            (Measures time only for ML operations preprocessing with predict)
    }
    """
    if 'image' not in request.files:
        # Handle if no file is selected
        return 'No file selected'

    start_time = time.time()

    file = request.files['image']

    # Get pixels out of file
    image_data = Image.open(file)

    # # Check image shape
    # if image_data.size != (28, 28):
    #     return "Invalid image shape. Expected (28, 28), take from 'demo images' folder."

    # Preprocess the image
    processed_image = preprocess_image(image_data)

    # Make a prediction using YOLO
    results = model(processed_image)

    # Process the YOLO output
    detections = []
    for det in results.xyxy[0]:  # Assuming results are in xyxy format (xmin, ymin, xmax, ymax, confidence, class)
        x_min, y_min, x_max, y_max, confidence, class_idx = det
        width = x_max - x_min
        height = y_max - y_min
        detection = {
            "label": int(class_idx),
            "confidence": float(confidence),
            "bbox": [float(x_min), float(y_min), float(width), float(height)]
        }
        detections.append(detection)

    # Calculate latency in milliseconds
    latency_ms = (time.time() - start_time) * 1000

    # Return the detection results and latency as JSON response
    response = {
        'detections': detections,
        'ml-latency-ms': round(latency_ms, 4)
    }

    # dictionary is not a JSON: https://www.quora.com/What-is-the-difference-between-JSON-and-a-dictionary
    # flask.jsonify vs json.dumps https://sentry.io/answers/difference-between-json-dumps-and-flask-jsonify/
    # The flask.jsonify() function returns a Response object with Serializable JSON and content_type=application/json.
    return jsonify(response)


# Helper function to preprocess the image
def preprocess_image(image_data):
    """Preprocess image for YOLO Model Inference

    :param image_data: Raw image (PIL.Image)
    :return: image: Preprocessed Image (Tensor)
    """
    # Define the YOLO input size (example 640x640, you can modify this based on your model)
    input_size = (640, 640)

    # Define transformation: Resize the image, convert to Tensor, and normalize pixel values
    transform = transforms.Compose([
        transforms.Resize(input_size),       # Resize to YOLO input size
        transforms.ToTensor(),               # Convert image to PyTorch Tensor (通道數、影像高度和寬度)
        transforms.Normalize([0.0, 0.0, 0.0], [1.0, 1.0, 1.0])  # Normalization (if needed)
    ])

    # Apply transformations to the image
    image = transform(image_data)

    # Add batch dimension (1, C, H, W) since YOLO expects a batch
    image = image.unsqueeze(0)

    return image


# API route for health check
@app.route('/health', methods=['GET'])
def health():
    """
    Health check API to ensure the application is running.
    Returns "OK" if the application is healthy.
    Demo Usage: "curl http://localhost:5000/health" or using alias "curl http://127.0.0.1:5000/health"
    """
    return 'OK'


# API route for version
@app.route('/version', methods=['GET'])
def version():
    """
    Returns the version of the application.
    Demo Usage: "curl http://127.0.0.1:5000/version" or using alias "curl http://127.0.0.1:5000/version"
    """
    return '1.0'


@app.route("/")
def hello_world():
    return render_template("index.html")
    # return "<p>Hello, Team!</p>"


# Start the Flask application
if __name__ == '__main__':
    app.run(debug=True)