diff --git a/.DS_Store b/.DS_Store
new file mode 100644
index 0000000..8e57668
Binary files /dev/null and b/.DS_Store differ
diff --git a/evaluate.py b/evaluate.py
new file mode 100644
index 0000000..3ce930a
--- /dev/null
+++ b/evaluate.py
@@ -0,0 +1,60 @@
+import argparse
+import gzip
+import json
+import os
+
+import numpy as np
+
+
+def load_data(data_dir):
+    files = [
+        'train-labels-idx1-ubyte.gz', 'train-images-idx3-ubyte.gz',
+        't10k-labels-idx1-ubyte.gz', 't10k-images-idx3-ubyte.gz'
+    ]
+    paths = []
+    for f in files:
+        paths.append(os.path.join(data_dir, f))
+    with gzip.open(paths[0], 'rb') as f:
+        y_train = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[1], 'rb') as f:
+        x_train = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_train), 28, 28)
+    with gzip.open(paths[2], 'rb') as f:
+        y_test = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[3], 'rb') as f:
+        x_test = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_test), 28, 28)
+    return (x_train, y_train), (x_test, y_test)
+
+
+def setup_config():
+    parser = argparse.ArgumentParser(
+        description="MNIST digit classification Evaluation using prediction results and labels"
+    )
+    parser.add_argument('--label-dir', required=True, help='Directory to labels')
+    parser.add_argument('--predictions', required=True, help='Path to prediction results numpy-array-like file')
+    parser.add_argument('--output-dir', required=True, help='Directory to evaluation result')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = setup_config()
+
+    (_, _), (_, labels) = load_data(args.label_dir)
+    predictions = np.load(args.predictions)
+
+    labels = np.reshape(labels, (-1,))
+    predictions = np.reshape(predictions, (-1,))
+
+    accuracy = (labels == predictions).sum() / labels.shape[0]
+    print(f'Accuracy is {accuracy}')
+
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+        result_file = os.path.join(args.output_dir, "result.json")
+        with open(result_file, 'w') as f:
+            json.dump({'accuracy': accuracy}, f)
+        print(f"Saved evaluation result to {result_file}")
+
+
+if __name__ == '__main__':
+    main()
diff --git a/model/saved_model.pb b/model/saved_model.pb
new file mode 100644
index 0000000..0a4bdf0
Binary files /dev/null and b/model/saved_model.pb differ
diff --git a/model/variables/variables.data-00000-of-00001 b/model/variables/variables.data-00000-of-00001
new file mode 100644
index 0000000..8d36841
Binary files /dev/null and b/model/variables/variables.data-00000-of-00001 differ
diff --git a/model/variables/variables.index b/model/variables/variables.index
new file mode 100644
index 0000000..cec8d0e
Binary files /dev/null and b/model/variables/variables.index differ
diff --git a/predict.py b/predict.py
new file mode 100644
index 0000000..1246a1d
--- /dev/null
+++ b/predict.py
@@ -0,0 +1,126 @@
+import argparse
+import gzip
+import json
+import os
+
+import numpy as np
+import tensorflow as tf
+
+num_workers = 1
+
+
+def load_data(data_dir):
+    files = [
+        'train-labels-idx1-ubyte.gz', 'train-images-idx3-ubyte.gz',
+        't10k-labels-idx1-ubyte.gz', 't10k-images-idx3-ubyte.gz'
+    ]
+    paths = []
+    for f in files:
+        paths.append(os.path.join(data_dir, f))
+    with gzip.open(paths[0], 'rb') as f:
+        y_train = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[1], 'rb') as f:
+        x_train = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_train), 28, 28)
+    with gzip.open(paths[2], 'rb') as f:
+        y_test = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[3], 'rb') as f:
+        x_test = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_test), 28, 28)
+    return (x_train, y_train), (x_test, y_test)
+
+
+def mnist_test_images(batch_size=64, data_dir=None):
+    # load dataset
+    if data_dir:
+        print(f'Loading mnist data from {data_dir}')
+        (x_train, y_train), (x_test, y_test) = load_data(data_dir)
+    else:
+        print('Loading mnist data from tf.keras.datasets.mnist')
+        (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
+    x_train, x_test = x_train / 255.0, x_test / 255.0
+
+    test_images = tf.data.Dataset.from_tensor_slices(x_test).batch(batch_size)
+
+    return test_images
+
+
+def get_strategy(strategy='off'):
+    strategy = strategy.lower()
+    # multiple nodes, every nodes have multiple GPUs
+    if strategy == "multi_worker_mirrored":
+        return tf.distribute.experimental.MultiWorkerMirroredStrategy()
+    # single node with multiple GPUs
+    if strategy == "mirrored":
+        return tf.distribute.MirroredStrategy()
+    # single node with single GPU
+    return tf.distribute.get_strategy()
+
+
+def setup_env(args):
+    tf.config.set_soft_device_placement(True)
+
+    # limit the gpu memory usage as much as it need.
+    try:
+        gpus = tf.config.experimental.list_physical_devices('GPU')
+        for gpu in gpus:
+            tf.config.experimental.set_memory_growth(gpu, True)
+        logical_gpus = tf.config.list_logical_devices('GPU')
+        print(f"Detected {len(gpus)} Physical GPUs, {len(logical_gpus)} Logical GPUs")
+    except RuntimeError as e:
+        print(e)
+
+    if args.strategy == 'multi_worker_mirrored':
+        index = int(os.environ['VK_TASK_INDEX'])
+        task_name = os.environ["VC_TASK_NAME"].upper()
+        ips = os.environ[f'VC_{task_name}_HOSTS']
+        ips = ips.split(',')
+        global num_workers
+        num_workers = len(ips)
+        ips = [f'{ip}:20000' for ip in ips]
+        os.environ["TF_CONFIG"] = json.dumps({
+            "cluster": {
+                "worker": ips
+            },
+            "task": {"type": "worker", "index": index}
+        })
+
+
+def setup_config():
+    parser = argparse.ArgumentParser(description='MNIST digits classification using trained model')
+    parser.add_argument('--model-dir', help='Directory to model')
+    parser.add_argument('--data-dir', help='Directory to MNIST dataset')
+    parser.add_argument('--output-dir', help='Directory to save models and logs')
+    parser.add_argument(
+        '--strategy',
+        default='off',
+        choices=['off', 'mirrored', 'multi_worker_mirrored'],
+        help='TensorFlow distributed training strategies'
+    )
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = setup_config()
+    # tf2 limitation: Collective ops must be configured at program startup
+    strategy = get_strategy(args.strategy)
+    setup_env(args)
+
+    with strategy.scope():
+        test_images = mnist_test_images(batch_size=64 * num_workers, data_dir=args.data_dir)
+        model = tf.keras.models.load_model(args.model_dir)
+        model.summary()
+
+    logits = model.predict(test_images, verbose=2)
+    probabilities = tf.nn.softmax(logits).numpy()
+    predictions = np.argmax(probabilities, 1)
+
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+        prediction_results_file = os.path.join(args.output_dir, "prediction_results.npy")
+        np.save(prediction_results_file, predictions)
+        print(f"Saved prediction results to {prediction_results_file}")
+
+
+if __name__ == '__main__':
+    print("TensorFlow version:", tf.__version__)
+    main()
diff --git a/serving/.DS_Store b/serving/.DS_Store
new file mode 100644
index 0000000..c9ded55
Binary files /dev/null and b/serving/.DS_Store differ
diff --git a/serving/Dockerfile b/serving/Dockerfile
new file mode 100644
index 0000000..24851fb
--- /dev/null
+++ b/serving/Dockerfile
@@ -0,0 +1,26 @@
+FROM tensorflow/tensorflow:2.4.3-gpu
+
+USER root
+ENV LANG=C.UTF-8 LC_ALL=C.UTF-8 SHELL=/bin/bash
+
+RUN sed -i "s@http://.*archive.ubuntu.com@http://mirrors.tuna.tsinghua.edu.cn@g" /etc/apt/sources.list && \
+    sed -i "s@http://.*security.ubuntu.com@http://mirrors.tuna.tsinghua.edu.cn@g" /etc/apt/sources.list
+
+RUN rm -rf /var/lib/apt/lists/* \
+        /etc/apt/sources.list.d/cuda.list \
+        /etc/apt/sources.list.d/nvidia-ml.list && \
+    apt-get update && \
+    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        apt-utils ca-certificates wget curl vim git openssh-server && \
+    ldconfig && \
+    apt-get clean && \
+    apt-get autoremove && \
+    rm -rf /var/lib/apt/lists/* /tmp/* ~/*
+
+RUN python -m pip --no-cache-dir install -i https://mirrors.ustc.edu.cn/pypi/web/simple \
+    Flask \
+    Pillow
+
+COPY . /mnist-serving/
+
+CMD ['cd /mnist-serving; python serving.py']
diff --git a/serving/serving.py b/serving/serving.py
new file mode 100644
index 0000000..c809fa9
--- /dev/null
+++ b/serving/serving.py
@@ -0,0 +1,100 @@
+import os
+import sys
+import traceback
+
+import numpy as np
+import tensorflow as tf
+from flask import Flask
+from flask import jsonify
+from flask import request
+
+import utils
+
+app = Flask(__name__)
+model = None
+
+
+def load_model():
+    tf.config.set_soft_device_placement(True)
+
+    # limit the gpu memory usage as much as it need.
+    try:
+        gpus = tf.config.experimental.list_physical_devices('GPU')
+        for gpu in gpus:
+            tf.config.experimental.set_memory_growth(gpu, True)
+        logical_gpus = tf.config.list_logical_devices('GPU')
+        print(f"Detected {len(gpus)} Physical GPUs, {len(logical_gpus)} Logical GPUs")
+    except RuntimeError as e:
+        print(e)
+
+    global model
+    model = tf.keras.models.load_model(os.environ['MODEL_PATH'])
+    model.summary()
+    return model
+
+
+@app.route('/')
+def hello_world():
+    return jsonify({'status': {'code': 0, 'msg': 'success'}, 'data': 'hello world'})
+
+
+@app.route('/image', methods=['POST'])
+def digits_classification():
+    global model
+    if model is None:
+        model = load_model()
+
+    # accept files as URL, base64 encoded or file stream
+    if request.content_type and request.content_type.startswith('application/json'):
+        data = request.get_json()
+        image_files = None
+        image_base64s = data.get('base64s')
+        image_urls = data.get('urls')
+    else:
+        # content-type: "application/x-www-form-unlencoded" or "multipart/form-data"
+        data = request.form
+        image_files = request.files.getlist('files')
+        image_base64s = data.getlist('base64s')
+        image_urls = data.getlist('urls')
+
+    # image bytes
+    images, request_image_type = utils.get_request_images_as_file(
+        image_files=image_files,
+        image_base64s=image_base64s,
+        image_urls=image_urls
+    )
+
+    # preprocessing: resize, convert to gray, normalize
+    images = utils.mnist_preprocess(images)
+
+    logits = model.predict(images, verbose=0)
+    probability = tf.nn.softmax(logits).numpy()
+    prediction = np.argmax(probability, 1)
+
+    return jsonify(
+        {
+            'status': {'code': 0, 'msg': 'success'},
+            'data': {
+                'probabilities': probability.tolist(),
+                'predictions': prediction.tolist()
+            }
+        }
+    )
+
+
+@app.errorhandler(Exception)
+def handle_unknown_error(e):
+    return jsonify({
+        'status': {'code': 500, 'msg': repr(traceback.format_exception(*sys.exc_info()))},
+        'data': None
+    })
+
+
+if __name__ == '__main__':
+    # 如果服务运行于 Nginx 之类的代理下，由于代理设置的某些 HTTP 报文头，可能存在请求不能转发的情况
+    # 使用下面的方式能保证正常转发，但是仅在信任代理、信任请求方的情况下使用
+    # https://flask.palletsprojects.com/en/1.1.x/deploying/wsgi-standalone/#proxy-setups
+    # from werkzeug.contrib.fixers import ProxyFix
+    # app.wsgi_app = ProxyFix(app.wsgi_app)
+    model = load_model()
+    app.run(host='0.0.0.0', port=5000)
diff --git a/serving/utils.py b/serving/utils.py
new file mode 100644
index 0000000..f64ea30
--- /dev/null
+++ b/serving/utils.py
@@ -0,0 +1,102 @@
+import base64
+from io import BytesIO
+
+import numpy as np
+import requests
+from PIL import Image
+
+
+def download_image_as_bytes(url, max_file_size=0):
+    """根据 URL 下载图片。
+
+    Args:
+        url: string, 图片URL.
+        max_file_size: integer, 图片文件最大值，0则不限制图片大小。
+
+    Returns:
+        bytes, 图片
+
+    Raises:
+        超过最大图片文件大小限制
+        ImageUrlDownloadFailError 图片通过 URL 下载失败
+    """
+    try:
+        # requests, streaming request and streaming uploads
+        # https://requests.readthedocs.io/en/stable/user/advanced/
+        res = requests.get(url, allow_redirects=True, stream=True)
+        if res.status_code != 200:
+            raise Exception('ImageUrlDownloadFailError')
+    except:
+        raise Exception('ImageUrlDownloadFailError')
+
+    file_size = int(res.headers['content-length'])
+    if 0 < max_file_size < file_size:
+        raise Exception(
+            f'image size must less than {max_file_size / 1024 / 1024}MB, '
+            f'but get {file_size / 1024 / 1024}MB, url is {url}'
+        )
+
+    image = res.content
+    res.close()
+    return image
+
+
+def get_request_images_as_file(
+        image_files=None, image_base64s=None, image_urls=None, max_file_size=0
+):
+    """由于允许客户端以 file, base64, url 三种方式发送图片，但同一请求只能以其中一种方式传输图片，
+    判断上传图片方式，并将数据转换成 bytes 类型的图片文件。
+
+    Args:
+        image_files: list of object which implements `.read()`, `.tell()`, `.seek()`, such as
+            werkzeug.FileStorage, 图片文件
+        image_base64s: list of string, base64 编码后的图片文件
+        image_urls: list of string, 图片URL地址
+        max_file_size: integer, 单张图片文件大小最大值，如果超过了会抛 InvalidContentLengthError 异常
+
+    Returns:
+        Tuple[List[bytes], string], 图片文件对象，传入的图片数据类型，包括 "file", "base64", "url".
+
+    Raises:
+        InvalidContentLengthError, 超过最大图片文件大小限制
+        InvalidArgumentError, 图片参数或值为空
+        ImageUrlDownloadFailError, 图片通过 URL 下载失败
+    """
+    if image_files and len(image_files) > 0 and None not in image_files:
+        image_bytes = [f.read() for f in image_files]
+        type_ = 'file'
+
+    elif image_base64s and len(image_base64s) > 0 and None not in image_base64s:
+        image_bytes = [base64.b64decode(s) for s in image_base64s]
+        type_ = 'base64'
+
+    elif image_urls and len(image_urls) > 0 and None not in image_urls:
+        image_bytes = [download_image_as_bytes(url, max_file_size=max_file_size)
+                       for url in image_urls]
+        type_ = 'url'
+
+    else:
+        raise Exception('no image')
+
+    if max_file_size > 0:
+        for idx, f in enumerate(image_bytes):
+            if len(f) > max_file_size:
+                raise Exception(f'image size must less than {max_file_size / 1024 / 1024}MB, '
+                                f'but image-{idx + 1} is {len(f) / 1024 / 1024}MB.')
+    return image_bytes, type_
+
+
+def mnist_preprocess(images):
+    results = []
+    for img in images:
+        if isinstance(img, bytes):
+            image_file = BytesIO(img)
+        else:
+            image_file = img
+
+        image = Image.open(image_file)
+        image = image.resize((28, 28))
+        image = image.convert('L')
+        image = np.asarray(image)[:, :] / 255.
+        results.append(image)
+    return np.asarray(results)
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..745b2cb
--- /dev/null
+++ b/train.py
@@ -0,0 +1,148 @@
+import argparse
+import gzip
+import json
+import os
+
+import numpy as np
+import tensorflow as tf
+
+num_workers = 1
+
+
+def load_data(data_dir):
+    files = [
+        'train-labels-idx1-ubyte.gz', 'train-images-idx3-ubyte.gz',
+        't10k-labels-idx1-ubyte.gz', 't10k-images-idx3-ubyte.gz'
+    ]
+    paths = []
+    for f in files:
+        paths.append(os.path.join(data_dir, f))
+    with gzip.open(paths[0], 'rb') as f:
+        y_train = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[1], 'rb') as f:
+        x_train = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_train), 28, 28)
+    with gzip.open(paths[2], 'rb') as f:
+        y_test = np.frombuffer(f.read(), np.uint8, offset=8)
+    with gzip.open(paths[3], 'rb') as f:
+        x_test = np.frombuffer(f.read(), np.uint8, offset=16).reshape(len(y_test), 28, 28)
+    return (x_train, y_train), (x_test, y_test)
+
+
+def mnist_dataset(batch_size=64, data_dir=None):
+    # load dataset
+    if data_dir:
+        print(f'Loading mnist data from {data_dir}')
+        (x_train, y_train), (x_test, y_test) = load_data(data_dir)
+    else:
+        print('Loading mnist data from tf.keras.datasets.mnist')
+        (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
+    x_train, x_test = x_train / 255.0, x_test / 255.0
+
+    train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(60000).repeat().batch(batch_size)
+    test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(batch_size)
+
+    return train_dataset, test_dataset
+
+
+def get_strategy(strategy='off'):
+    strategy = strategy.lower()
+    # multiple nodes, every nodes have multiple GPUs
+    if strategy == "multi_worker_mirrored":
+        return tf.distribute.experimental.MultiWorkerMirroredStrategy()
+    # single node with multiple GPUs
+    if strategy == "mirrored":
+        return tf.distribute.MirroredStrategy()
+    # single node with single GPU
+    return tf.distribute.get_strategy()
+
+
+def setup_env(args):
+    tf.config.set_soft_device_placement(True)
+
+    # limit the gpu memory usage as much as it need.
+    try:
+        gpus = tf.config.experimental.list_physical_devices('GPU')
+        for gpu in gpus:
+            tf.config.experimental.set_memory_growth(gpu, True)
+        logical_gpus = tf.config.list_logical_devices('GPU')
+        print(f"Detected {len(gpus)} Physical GPUs, {len(logical_gpus)} Logical GPUs")
+    except RuntimeError as e:
+        print(e)
+
+    if args.strategy == 'multi_worker_mirrored':
+        index = int(os.environ['VK_TASK_INDEX'])
+        task_name = os.environ["VC_TASK_NAME"].upper()
+        ips = os.environ[f'VC_{task_name}_HOSTS']
+        ips = ips.split(',')
+        global num_workers
+        num_workers = len(ips)
+        ips = [f'{ip}:20000' for ip in ips]
+        os.environ["TF_CONFIG"] = json.dumps({
+            "cluster": {
+                "worker": ips
+            },
+            "task": {"type": "worker", "index": index}
+        })
+
+
+def setup_config():
+    parser = argparse.ArgumentParser(description='Train MNIST digits classification')
+    parser.add_argument('--data-dir', help='Directory to MNIST dataset')
+    parser.add_argument('--output-dir', help='Directory to save models and logs')
+    parser.add_argument('--epochs', default=2, help='Number of epochs')
+    parser.add_argument('--eval', action='store_true', help='whether do evaluation after training finished')
+    parser.add_argument(
+        '--strategy',
+        default='off',
+        choices=['off', 'mirrored', 'multi_worker_mirrored'],
+        help='TensorFlow distributed training strategies'
+    )
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = setup_config()
+    # tf2 limitation: Collective ops must be configured at program startup
+    strategy = get_strategy(args.strategy)
+    setup_env(args)
+
+    with strategy.scope():
+        # build dataset
+        train_dataset, test_dataset = mnist_dataset(
+            batch_size=64 * num_workers,
+            data_dir=args.data_dir
+        )
+
+        # build model
+        model = tf.keras.models.Sequential([
+            tf.keras.layers.Flatten(input_shape=(28, 28)),
+            tf.keras.layers.Dense(128, activation='relu'),
+            tf.keras.layers.Dropout(0.2),
+            tf.keras.layers.Dense(10)
+        ])
+        model.compile(
+            optimizer='adam',
+            loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+            metrics=['accuracy']
+        )
+        model.summary()
+
+    # training
+    model.fit(train_dataset, epochs=args.epochs, steps_per_epoch=70)
+
+    # evaluation
+    if args.eval:
+        model.evaluate(test_dataset, verbose=2)
+
+    # save model
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+        model.save(args.output_dir)
+        print(f'Saved model to {args.output_dir}')
+
+
+if __name__ == '__main__':
+    # python train.py --data-dir=/path/to/MNIST/dataset --output-dir=/path/to/output
+    print("TensorFlow version:", tf.__version__)
+    main()