# 自定义脚本代码示例
[从OBS中导入模型文件创建模型](https://support.huaweicloud.com/usermanual-standard-modelarts/inference-modelarts-0008.html)时，模型文件包需符合ModelArts的模型包规范，推理代码和配置文件也需遵循ModelArts的要求。
本章节提供针对常用AI引擎的自定义脚本代码示例（包含推理代码示例）。模型推理代码编写的通用方法及说明请见[模型推理代码编写说明](https://support.huaweicloud.com/usermanual-standard-modelarts/inference-modelarts-0057.html)。
 #### TensorFlow
TensorFlow存在两种接口类型，keras接口和tf接口，其训练和保存模型的代码存在差异，但是推理代码编写方式一致。
**训练模型（keras接口）**
```
from keras.models import Sequential
model = Sequential()
from keras.layers import Dense
import tensorflow as tf
# 导入训练数据集
mnist = tf.keras.datasets.mnist
(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
print(x_train.shape)
from keras.layers import Dense
from keras.models import Sequential
import keras
from keras.layers import Dense, Activation, Flatten, Dropout
# 定义模型网络
model = Sequential()
model.add(Flatten(input_shape=(28,28)))
model.add(Dense(units=5120,activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(units=10, activation='softmax'))
# 定义优化器，损失函数等
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
model.summary()
# 训练
model.fit(x_train, y_train, epochs=2)
# 评估
model.evaluate(x_test, y_test)
```
**保存模型（keras接口）**
```
from keras import backend as K  
# K.get_session().run(tf.global_variables_initializer())
# 定义预测接口的inputs和outputs
# inputs和outputs字典的key值会作为模型输入输出tensor的索引键
# 模型输入输出定义需要和推理自定义脚本相匹配
predict_signature = tf.saved_model.signature_def_utils.predict_signature_def(
    inputs={"images" : model.input},
    outputs={"scores" : model.output}
)
# 定义保存路径
builder = tf.saved_model.builder.SavedModelBuilder('./mnist_keras/')
builder.add_meta_graph_and_variables(
    sess = K.get_session(),
    # 推理部署需要定义tf.saved_model.tag_constants.SERVING标签
    tags=[tf.saved_model.tag_constants.SERVING],
    """
    signature_def_map：items只能有一个，或者需要定义相应的key为
    tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
    """
    signature_def_map={
        tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
            predict_signature
    }
)
builder.save()
```
**训练模型（tf接口）**
```
from __future__ import print_function
import gzip
import os
import urllib
import numpy
import tensorflow as tf
from six.moves import urllib
# 训练数据来源于yann lecun官方网站http://yann.lecun.com/exdb/mnist/
SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
TRAIN_IMAGES = 'train-images-idx3-ubyte.gz'
TRAIN_LABELS = 'train-labels-idx1-ubyte.gz'
TEST_IMAGES = 't10k-images-idx3-ubyte.gz'
TEST_LABELS = 't10k-labels-idx1-ubyte.gz'
VALIDATION_SIZE = 5000
def maybe_download(filename, work_directory):
    """Download the data from Yann's website, unless it's already here."""
    if not os.path.exists(work_directory):
        os.mkdir(work_directory)
    filepath = os.path.join(work_directory, filename)
    if not os.path.exists(filepath):
        filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
        statinfo = os.stat(filepath)
        print('Successfully downloaded %s %d bytes.' % (filename, statinfo.st_size))
    return filepath
def _read32(bytestream):
    dt = numpy.dtype(numpy.uint32).newbyteorder('>')
    return numpy.frombuffer(bytestream.read(4), dtype=dt)[0]
def extract_images(filename):
    """Extract the images into a 4D uint8 numpy array [index, y, x, depth]."""
    print('Extracting %s' % filename)
    with gzip.open(filename) as bytestream:
        magic = _read32(bytestream)
        if magic != 2051:
            raise ValueError(
                'Invalid magic number %d in MNIST image file: %s' %
                (magic, filename))
        num_images = _read32(bytestream)
        rows = _read32(bytestream)
        cols = _read32(bytestream)
        buf = bytestream.read(rows * cols * num_images)
        data = numpy.frombuffer(buf, dtype=numpy.uint8)
        data = data.reshape(num_images, rows, cols, 1)
        return data
def dense_to_one_hot(labels_dense, num_classes=10):
    """Convert class labels from scalars to one-hot vectors."""
    num_labels = labels_dense.shape[0]
    index_offset = numpy.arange(num_labels) * num_classes
    labels_one_hot = numpy.zeros((num_labels, num_classes))
    labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
    return labels_one_hot
def extract_labels(filename, one_hot=False):
    """Extract the labels into a 1D uint8 numpy array [index]."""
    print('Extracting %s' % filename)
    with gzip.open(filename) as bytestream:
        magic = _read32(bytestream)
        if magic != 2049:
            raise ValueError(
                'Invalid magic number %d in MNIST label file: %s' %
                (magic, filename))
        num_items = _read32(bytestream)
        buf = bytestream.read(num_items)
        labels = numpy.frombuffer(buf, dtype=numpy.uint8)
        if one_hot:
            return dense_to_one_hot(labels)
        return labels
class DataSet(object):
    """Class encompassing test, validation and training MNIST data set."""
    def __init__(self, images, labels, fake_data=False, one_hot=False):
        """Construct a DataSet. one_hot arg is used only if fake_data is true."""
        if fake_data:
            self._num_examples = 10000
            self.one_hot = one_hot
        else:
            assert images.shape[0] == labels.shape[0], (
                    'images.shape: %s labels.shape: %s' % (images.shape,
                                                           labels.shape))
            self._num_examples = images.shape[0]
            # Convert shape from [num examples, rows, columns, depth]
            # to [num examples, rows*columns] (assuming depth == 1)
            assert images.shape[3] == 1
            images = images.reshape(images.shape[0],
                                    images.shape[1] * images.shape[2])
            # Convert from [0, 255] -> [0.0, 1.0].
            images = images.astype(numpy.float32)
            images = numpy.multiply(images, 1.0 / 255.0)
        self._images = images
        self._labels = labels
        self._epochs_completed = 0
        self._index_in_epoch = 0
    @property
    def images(self):
        return self._images
    @property
    def labels(self):
        return self._labels
    @property
    def num_examples(self):
        return self._num_examples
    @property
    def epochs_completed(self):
        return self._epochs_completed
    def next_batch(self, batch_size, fake_data=False):
        """Return the next `batch_size` examples from this data set."""
        if fake_data:
            fake_image = [1] * 784
            if self.one_hot:
                fake_label = [1] + [0] * 9
            else:
                fake_label = 0
            return [fake_image for _ in range(batch_size)], [
                fake_label for _ in range(batch_size)
            ]
        start = self._index_in_epoch
        self._index_in_epoch += batch_size
        if self._index_in_epoch > self._num_examples:
            # Finished epoch
            self._epochs_completed += 1
            # Shuffle the data
            perm = numpy.arange(self._num_examples)
            numpy.random.shuffle(perm)
            self._images = self._images[perm]
            self._labels = self._labels[perm]
            # Start next epoch
            start = 0
            self._index_in_epoch = batch_size
            assert batch_size <= self._num_examples
        end = self._index_in_epoch
        return self._images[start:end], self._labels[start:end]
def read_data_sets(train_dir, fake_data=False, one_hot=False):
    """Return training, validation and testing data sets."""
    class DataSets(object):
        pass
    data_sets = DataSets()
    if fake_data:
        data_sets.train = DataSet([], [], fake_data=True, one_hot=one_hot)
        data_sets.validation = DataSet([], [], fake_data=True, one_hot=one_hot)
        data_sets.test = DataSet([], [], fake_data=True, one_hot=one_hot)
        return data_sets
    local_file = maybe_download(TRAIN_IMAGES, train_dir)
    train_images = extract_images(local_file)
    local_file = maybe_download(TRAIN_LABELS, train_dir)
    train_labels = extract_labels(local_file, one_hot=one_hot)
    local_file = maybe_download(TEST_IMAGES, train_dir)
    test_images = extract_images(local_file)
    local_file = maybe_download(TEST_LABELS, train_dir)
    test_labels = extract_labels(local_file, one_hot=one_hot)
    validation_images = train_images[:VALIDATION_SIZE]
    validation_labels = train_labels[:VALIDATION_SIZE]
    train_images = train_images[VALIDATION_SIZE:]
    train_labels = train_labels[VALIDATION_SIZE:]
    data_sets.train = DataSet(train_images, train_labels)
    data_sets.validation = DataSet(validation_images, validation_labels)
    data_sets.test = DataSet(test_images, test_labels)
    return data_sets
training_iteration = 1000
modelarts_example_path =  './modelarts-mnist-train-save-deploy-example'
export_path = modelarts_example_path + '/model/'
data_path = './'
print('Training model...')
mnist = read_data_sets(data_path, one_hot=True)
sess = tf.InteractiveSession()
serialized_tf_example = tf.placeholder(tf.string, name='tf_example')
feature_configs = {'x': tf.FixedLenFeature(shape=[784], dtype=tf.float32), }
tf_example = tf.parse_example(serialized_tf_example, feature_configs)
x = tf.identity(tf_example['x'], name='x')  # use tf.identity() to assign name
y_ = tf.placeholder('float', shape=[None, 10])
w = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
sess.run(tf.global_variables_initializer())
y = tf.nn.softmax(tf.matmul(x, w) + b, name='y')
cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)
values, indices = tf.nn.top_k(y, 10)
table = tf.contrib.lookup.index_to_string_table_from_tensor(
    tf.constant([str(i) for i in range(10)]))
prediction_classes = table.lookup(tf.to_int64(indices))
for _ in range(training_iteration):
    batch = mnist.train.next_batch(50)
    train_step.run(feed_dict={x: batch[0], y_: batch[1]})
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
print('training accuracy %g' % sess.run(
    accuracy, feed_dict={
        x: mnist.test.images,
        y_: mnist.test.labels
    }))
print('Done training!')
```
**保存模型（tf接口）**
```
# 导出模型
# 模型需要采用saved_model接口保存
print('Exporting trained model to', export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
tensor_info_y = tf.saved_model.utils.build_tensor_info(y)
# 定义预测接口的inputs和outputs
# inputs和outputs字典的key值会作为模型输入输出tensor的索引键
# 模型输入输出定义需要和推理自定义脚本相匹配
prediction_signature = (
    tf.saved_model.signature_def_utils.build_signature_def(
        inputs={'images': tensor_info_x},
        outputs={'scores': tensor_info_y},
        method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))
legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
builder.add_meta_graph_and_variables(
    # tag设为serve/tf.saved_model.tag_constants.SERVING
    sess, [tf.saved_model.tag_constants.SERVING],
    signature_def_map={
        'predict_images':
            prediction_signature,
    },
    legacy_init_op=legacy_init_op)
builder.save()
print('Done exporting!')
```
**推理代码（keras接口和tf接口）**
在模型代码推理文件customize_service.py中，需要添加一个子类，该子类继承对应模型类型的父类，各模型类型的父类名称和导入语句如请参考[表1](https://support.huaweicloud.com/usermanual-standard-modelarts/inference-modelarts-0057.html#ZH-CN_TOPIC_0000001891635613__zh-cn_topic_0172466150_table55021545175412)。本案例中调用父类"_inference(self, data)"推理请求方法，因此下文代码中不需要重写方法。
```
from PIL import Image
import numpy as np
from model_service.tfserving_model_service import TfServingBaseService
class MnistService(TfServingBaseService):
    # 预处理中处理用户HTTPS接口输入匹配模型输入
    # 对应上述训练部分的模型输入为{"images":<array>}
    def _preprocess(self, data):
        preprocessed_data = {}
        images = []
        # 对输入数据进行迭代
        for k, v in data.items():
            for file_name, file_content in v.items():
                image1 = Image.open(file_content)
                image1 = np.array(image1, dtype=np.float32)
                image1.resize((1,784))
                images.append(image1)
        # 返回numpy array
        images = np.array(images,dtype=np.float32)
        # 对传入的多个样本做batch处理，shape保持和训练时输入一致
        images.resize((len(data), 784))
        preprocessed_data['images'] = images
        return preprocessed_data
    # 对应的上述训练部分保存模型的输出为{"scores":<array>}
    # 后处理中处理模型输出为HTTPS的接口输出
    def _postprocess(self, data):
        infer_output = {"mnist_result": []}
        # 迭代处理模型输出
        for output_name, results in data.items():
            for result in results:
                infer_output["mnist_result"].append(result.index(max(result)))
        return infer_output
```
 #### TensorFlow2.1
**训练并保存模型**
```
from __future__ import absolute_import, division, print_function, unicode_literals
import tensorflow as tf
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(256, activation='relu'),
    tf.keras.layers.Dropout(0.2),
    # 对输出层命名output，在模型推理时通过该命名取结果
    tf.keras.layers.Dense(10, activation='softmax', name="output")
])
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
model.fit(x_train, y_train, epochs=10)
tf.keras.models.save_model(model, "./mnist")
```
**推理代码**
在模型代码推理文件customize_service.py中，需要添加一个子类，该子类继承对应模型类型的父类，各模型类型的父类名称和导入语句如请参考[表1](https://support.huaweicloud.com/usermanual-standard-modelarts/inference-modelarts-0057.html#ZH-CN_TOPIC_0000001891635613__zh-cn_topic_0172466150_table55021545175412)。
```
import logging
import threading
import numpy as np
import tensorflow as tf
from PIL import Image
from model_service.tfserving_model_service import TfServingBaseService
logger = logging.getLogger()
logger.setLevel(logging.INFO)
class MnistService(TfServingBaseService):
    def __init__(self, model_name, model_path):
        self.model_name = model_name
        self.model_path = model_path
        self.model = None
        self.predict = None
        # label文件可以在这里加载,在后处理函数里使用
        # label.txt放在obs和模型包的目录
        # with open(os.path.join(self.model_path, 'label.txt')) as f:
        #     self.label = json.load(f)
        # 非阻塞方式加载saved_model模型，防止阻塞超时
        thread = threading.Thread(target=self.load_model)
        thread.start()
    def load_model(self):
        # load saved_model 格式的模型
        self.model = tf.saved_model.load(self.model_path)
        signature_defs = self.model.signatures.keys()
        signature = []
        # only one signature allowed
        for signature_def in signature_defs:
            signature.append(signature_def)
        if len(signature) == 1:
            model_signature = signature[0]
        else:
            logging.warning("signatures more than one, use serving_default signature from %s", signature)
            model_signature = tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY
        self.predict = self.model.signatures[model_signature]
    def _preprocess(self, data):
        images = []
        for k, v in data.items():
            for file_name, file_content in v.items():
                image1 = Image.open(file_content)
                image1 = np.array(image1, dtype=np.float32)
                image1.resize((28, 28, 1))
                images.append(image1)
        images = tf.convert_to_tensor(images, dtype=tf.dtypes.float32)
        preprocessed_data = images
        return preprocessed_data
    def _inference(self, data):
        return self.predict(data)
    def _postprocess(self, data):
        return {
            "result": int(data["output"].numpy()[0].argmax())
        }
```
#### Pytorch
**训练模型**
```
from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
# 定义网络结构
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        # 输入第二维需要为784
        self.hidden1 = nn.Linear(784, 5120, bias=False)
        self.output = nn.Linear(5120, 10, bias=False)
    def forward(self, x):
        x = x.view(x.size()[0], -1)
        x = F.relu((self.hidden1(x)))
        x = F.dropout(x, 0.2)
        x = self.output(x)
        return F.log_softmax(x)
def train(model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.cross_entropy(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % 10 == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                       100. * batch_idx / len(train_loader), loss.item()))
def test( model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item()  # sum up batch loss
            pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
            correct += pred.eq(target.view_as(pred)).sum().item()
    test_loss /= len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))
device = torch.device("cpu")
batch_size=64
kwargs={}
train_loader = torch.utils.data.DataLoader(
    datasets.MNIST('.', train=True, download=True,
                   transform=transforms.Compose([
                       transforms.ToTensor()
                   ])),
    batch_size=batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('.', train=False, transform=transforms.Compose([
        transforms.ToTensor()
    ])),
    batch_size=1000, shuffle=True, **kwargs)
model = Net().to(device)
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
optimizer = optim.Adam(model.parameters())
for epoch in range(1, 2 + 1):
    train(model, device, train_loader, optimizer, epoch)
    test(model, device, test_loader)
```
**保存模型**
```
# 必须采用state_dict的保存方式，支持异地部署
torch.save(model.state_dict(), "pytorch_mnist/mnist_mlp.pt")
```
**推理代码**
在模型代码推理文件customize_service.py中，需要添加一个子类，该子类继承对应模型类型的父类，各模型类型的父类名称和导入语句如请参考[表1](https://support.huaweicloud.com/usermanual-standard-modelarts/inference-modelarts-0057.html#ZH-CN_TOPIC_0000001891635613__zh-cn_topic_0172466150_table55021545175412)。
```
from PIL import Image
import log
from model_service.pytorch_model_service import PTServingBaseService
import torch.nn.functional as F
import torch.nn as nn
import torch
import json
import numpy as np
logger = log.getLogger(__name__)
import torchvision.transforms as transforms
# 定义模型预处理
infer_transformation = transforms.Compose([
    transforms.Resize((28,28)),
    # 需要处理成pytorch tensor
    transforms.ToTensor()
])
import os
class PTVisionService(PTServingBaseService):
    def __init__(self, model_name, model_path):
        # 调用父类构造方法
        super(PTVisionService, self).__init__(model_name, model_path)
        # 调用自定义函数加载模型
        self.model = Mnist(model_path)
        # 加载标签
        self.label = [0,1,2,3,4,5,6,7,8,9]
        # 亦可通过文件标签文件加载
        # model目录下放置label.json文件，此处读取
        dir_path = os.path.dirname(os.path.realpath(self.model_path))
        with open(os.path.join(dir_path, 'label.json')) as f:
            self.label = json.load(f)
    def _preprocess(self, data):
        preprocessed_data = {}
        for k, v in data.items():
            input_batch = []
            for file_name, file_content in v.items():
                with Image.open(file_content) as image1:
                    # 灰度处理
                    image1 = image1.convert("L")
                    if torch.cuda.is_available():
                        input_batch.append(infer_transformation(image1).cuda())
                    else:
                        input_batch.append(infer_transformation(image1))
            input_batch_var = torch.autograd.Variable(torch.stack(input_batch, dim=0), volatile=True)
            print(input_batch_var.shape)
            preprocessed_data[k] = input_batch_var
        return preprocessed_data
    def _postprocess(self, data):
        results = []
        for k, v in data.items():
            result = torch.argmax(v[0])
            result = {k: self.label[result]}
            results.append(result)
        return results
    def _inference(self, data):
        result = {}
        for k, v in data.items():
            result[k] = self.model(v)
        return result
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.hidden1 = nn.Linear(784, 5120, bias=False)
        self.output = nn.Linear(5120, 10, bias=False)
    def forward(self, x):
        x = x.view(x.size()[0], -1)
        x = F.relu((self.hidden1(x)))
        x = F.dropout(x, 0.2)
        x = self.output(x)
        return F.log_softmax(x)
def Mnist(model_path, **kwargs):
    # 生成网络
    model = Net()
    # 加载模型
    if torch.cuda.is_available():
        device = torch.device('cuda')
        model.load_state_dict(torch.load(model_path, map_location="cuda:0"))
    else:
        device = torch.device('cpu')
        model.load_state_dict(torch.load(model_path, map_location=device))
    # CPU或者GPU映射
    model.to(device)
    # 声明为推理模式
    model.eval()
    return model
```