本文為英文版的機器翻譯版本，如內容有任何歧義或不一致之處，概以英文版為準。

# 先決條件
<a name="neo-deployment-hosting-services-prerequisites"></a>

**注意**  
如果您使用 適用於 Python (Boto3) 的 AWS SDK、 或 SageMaker AI 主控台編譯模型 AWS CLI，請遵循本節中的指示。

若要建立 SageMaker Neo 編譯的模型，您需要下列項目：

1. 一個 Docker 映像 Amazon ECR URI。您可以從[這個清單](https://docs.aws.amazon.com/sagemaker/latest/dg/neo-deployment-hosting-services-container-images.html)選取一個符合需求的 URL。

1. 進入點指令碼檔案：

   1. **若為 PyTorch 和 MXNet 模型：**

      *如果您使用 SageMaker AI 訓練模型*，則訓練指令碼必須實作下述功能。訓練指令碼可當成推論期間的進入點指令碼。在[使用 MXNet 模組和 SageMaker Neo 進行 MNIST 訓練、編譯和部署](https://sagemaker-examples.readthedocs.io/en/latest/sagemaker_neo_compilation_jobs/mxnet_mnist/mxnet_mnist_neo.html)詳細說明的範例中，訓練指令碼 (`mnist.py`) 會實作必要的函式。

      *如果您未使用 SageMaker AI 訓練模型*，則必須提供可在推論時使用的進入點指令碼 (`inference.py`) 檔案。根據架構 (MxNet 或 PyTorch)，推論指令碼位置必須符合 [MxNet 的 SageMaker Python SDK 模型目錄結構](https://sagemaker.readthedocs.io/en/stable/frameworks/mxnet/using_mxnet.html#model-directory-structure)，或是 [PyTorch 的模型目錄結構](https://sagemaker.readthedocs.io/en/stable/frameworks/pytorch/using_pytorch.html#model-directory-structure)。

      在 CPU 和 GPU 執行個體類型上搭配 **PyTorch** 和 **MXNet** 使用 Neo 推論最佳化容器映像時，推論指令碼必須實作下列功能：
      + `model_fn`：載入模型。(選用)
      + `input_fn`：將傳入請求承載轉換為 numpy 陣列。
      + `predict_fn`：執行預測。
      + `output_fn`：將預測輸出轉換為回應承載。
      + 或者，您也可以定義 `transform_fn`，合併 `input_fn`、`predict_fn` 與 `output_fn`。

      以下是名為 `code` (`code/inference.py`) 的目錄中的 `inference.py` 指令碼範例，用於 **PyTorch 和 MXNet (Gluon 和模組)。**這些範例會先載入模型，然後將其提供給 GPU 上的映像資料：

------
#### [ MXNet Module ]

      ```
      import numpy as np
      import json
      import mxnet as mx
      import neomx  # noqa: F401
      from collections import namedtuple
      
      Batch = namedtuple('Batch', ['data'])
      
      # Change the context to mx.cpu() if deploying to a CPU endpoint
      ctx = mx.gpu()
      
      def model_fn(model_dir):
          # The compiled model artifacts are saved with the prefix 'compiled'
          sym, arg_params, aux_params = mx.model.load_checkpoint('compiled', 0)
          mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
          exe = mod.bind(for_training=False,
                         data_shapes=[('data', (1,3,224,224))],
                         label_shapes=mod._label_shapes)
          mod.set_params(arg_params, aux_params, allow_missing=True)
          
          # Run warm-up inference on empty data during model load (required for GPU)
          data = mx.nd.empty((1,3,224,224), ctx=ctx)
          mod.forward(Batch([data]))
          return mod
      
      
      def transform_fn(mod, image, input_content_type, output_content_type):
          # pre-processing
          decoded = mx.image.imdecode(image)
          resized = mx.image.resize_short(decoded, 224)
          cropped, crop_info = mx.image.center_crop(resized, (224, 224))
          normalized = mx.image.color_normalize(cropped.astype(np.float32) / 255,
                                        mean=mx.nd.array([0.485, 0.456, 0.406]),
                                        std=mx.nd.array([0.229, 0.224, 0.225]))
          transposed = normalized.transpose((2, 0, 1))
          batchified = transposed.expand_dims(axis=0)
          casted = batchified.astype(dtype='float32')
          processed_input = casted.as_in_context(ctx)
      
          # prediction/inference
          mod.forward(Batch([processed_input]))
      
          # post-processing
          prob = mod.get_outputs()[0].asnumpy().tolist()
          prob_json = json.dumps(prob)
          return prob_json, output_content_type
      ```

------
#### [ MXNet Gluon ]

      ```
      import numpy as np
      import json
      import mxnet as mx
      import neomx  # noqa: F401
      
      # Change the context to mx.cpu() if deploying to a CPU endpoint
      ctx = mx.gpu()
      
      def model_fn(model_dir):
          # The compiled model artifacts are saved with the prefix 'compiled'
          block = mx.gluon.nn.SymbolBlock.imports('compiled-symbol.json',['data'],'compiled-0000.params', ctx=ctx)
          
          # Hybridize the model & pass required options for Neo: static_alloc=True & static_shape=True
          block.hybridize(static_alloc=True, static_shape=True)
          
          # Run warm-up inference on empty data during model load (required for GPU)
          data = mx.nd.empty((1,3,224,224), ctx=ctx)
          warm_up = block(data)
          return block
      
      
      def input_fn(image, input_content_type):
          # pre-processing
          decoded = mx.image.imdecode(image)
          resized = mx.image.resize_short(decoded, 224)
          cropped, crop_info = mx.image.center_crop(resized, (224, 224))
          normalized = mx.image.color_normalize(cropped.astype(np.float32) / 255,
                                        mean=mx.nd.array([0.485, 0.456, 0.406]),
                                        std=mx.nd.array([0.229, 0.224, 0.225]))
          transposed = normalized.transpose((2, 0, 1))
          batchified = transposed.expand_dims(axis=0)
          casted = batchified.astype(dtype='float32')
          processed_input = casted.as_in_context(ctx)
          return processed_input
      
      
      def predict_fn(processed_input_data, block):
          # prediction/inference
          prediction = block(processed_input_data)
          return prediction
      
      def output_fn(prediction, output_content_type):
          # post-processing
          prob = prediction.asnumpy().tolist()
          prob_json = json.dumps(prob)
          return prob_json, output_content_type
      ```

------
#### [ PyTorch 1.4 and Older ]

      ```
      import os
      import torch
      import torch.nn.parallel
      import torch.optim
      import torch.utils.data
      import torch.utils.data.distributed
      import torchvision.transforms as transforms
      from PIL import Image
      import io
      import json
      import pickle
      
      
      def model_fn(model_dir):
          """Load the model and return it.
          Providing this function is optional.
          There is a default model_fn available which will load the model
          compiled using SageMaker Neo. You can override it here.
      
          Keyword arguments:
          model_dir -- the directory path where the model artifacts are present
          """
      
          # The compiled model is saved as "compiled.pt"
          model_path = os.path.join(model_dir, 'compiled.pt')
          with torch.neo.config(model_dir=model_dir, neo_runtime=True):
              model = torch.jit.load(model_path)
              device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
              model = model.to(device)
      
          # We recommend that you run warm-up inference during model load
          sample_input_path = os.path.join(model_dir, 'sample_input.pkl')
          with open(sample_input_path, 'rb') as input_file:
              model_input = pickle.load(input_file)
          if torch.is_tensor(model_input):
              model_input = model_input.to(device)
              model(model_input)
          elif isinstance(model_input, tuple):
              model_input = (inp.to(device) for inp in model_input if torch.is_tensor(inp))
              model(*model_input)
          else:
              print("Only supports a torch tensor or a tuple of torch tensors")
              return model
      
      
      def transform_fn(model, request_body, request_content_type,
                       response_content_type):
          """Run prediction and return the output.
          The function
          1. Pre-processes the input request
          2. Runs prediction
          3. Post-processes the prediction output.
          """
          # preprocess
          decoded = Image.open(io.BytesIO(request_body))
          preprocess = transforms.Compose([
              transforms.Resize(256),
              transforms.CenterCrop(224),
              transforms.ToTensor(),
              transforms.Normalize(
                  mean=[
                      0.485, 0.456, 0.406], std=[
                      0.229, 0.224, 0.225]),
          ])
          normalized = preprocess(decoded)
          batchified = normalized.unsqueeze(0)
          # predict
          device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
          batchified = batchified.to(device)
          output = model.forward(batchified)
      
          return json.dumps(output.cpu().numpy().tolist()), response_content_type
      ```

------
#### [ PyTorch 1.5 and Newer ]

      ```
      import os
      import torch
      import torch.nn.parallel
      import torch.optim
      import torch.utils.data
      import torch.utils.data.distributed
      import torchvision.transforms as transforms
      from PIL import Image
      import io
      import json
      import pickle
      
      
      def model_fn(model_dir):
          """Load the model and return it.
          Providing this function is optional.
          There is a default_model_fn available, which will load the model
          compiled using SageMaker Neo. You can override the default here.
          The model_fn only needs to be defined if your model needs extra
          steps to load, and can otherwise be left undefined.
      
          Keyword arguments:
          model_dir -- the directory path where the model artifacts are present
          """
      
          # The compiled model is saved as "model.pt"
          model_path = os.path.join(model_dir, 'model.pt')
          device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
          model = torch.jit.load(model_path, map_location=device)
          model = model.to(device)
      
          return model
      
      
      def transform_fn(model, request_body, request_content_type,
                          response_content_type):
          """Run prediction and return the output.
          The function
          1. Pre-processes the input request
          2. Runs prediction
          3. Post-processes the prediction output.
          """
          # preprocess
          decoded = Image.open(io.BytesIO(request_body))
          preprocess = transforms.Compose([
                                      transforms.Resize(256),
                                      transforms.CenterCrop(224),
                                      transforms.ToTensor(),
                                      transforms.Normalize(
                                          mean=[
                                              0.485, 0.456, 0.406], std=[
                                              0.229, 0.224, 0.225]),
                                          ])
          normalized = preprocess(decoded)
          batchified = normalized.unsqueeze(0)
          
          # predict
          device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
          batchified = batchified.to(device)
          output = model.forward(batchified)
          return json.dumps(output.cpu().numpy().tolist()), response_content_type
      ```

------

   1.  **針對 inf1 執行個體或 onnx、xgboost、keras 容器映像檔** 

      針對所有其他 Neo 推論最佳化容器映像或 inferentia 執行個體類型，進入點指令碼必須為 Neo 深度學習執行期實作以下函式：
      + `neo_preprocess`：將傳入請求承載轉換為 numpy 陣列。
      + `neo_postprocess`：將 Neo 深度學習執行期的預測輸出轉換為回應內文。
**注意**  
前面這兩個函式都未使用 MXNet、PyTorch 或 TensorFlow 的任何功能。

      如需如何使用這些函式的範例，請參閱 [Neo 模型編譯範例筆記本](https://docs.aws.amazon.com//sagemaker/latest/dg/neo.html#neo-sample-notebooks)。

   1. **針對 TensorFlow 模型**

      如果您的模型在將資料傳送至模型之前需要自訂的預處理和後處理邏輯，則您必須指定可在推論時使用的進入點指令碼 `inference.py` 檔案。指令碼應該實作一對 `input_handler` 和 `output_handler` 函式或單一處理常式函式。
**注意**  
請注意，如果已實作處理常式函式，則會忽略 `input_handler` 和 `output_handler`。

      以下是 `inference.py` 指令碼的程式碼範例，您可以將其與編譯模型結合在一起，在映像分類模型上執行自訂預處理和後處理。SageMaker AI 用戶端會將映像檔案當成 `application/x-image` 內容類型傳送至 `input_handler` 函式，在這個函式將該檔案轉換為 JSON。然後使用 REST API 將轉換後的映像檔案傳送至 [Tensorflow 模型伺服器 (TFX)](https://www.tensorflow.org/tfx/serving/api_rest)。

      ```
      import json
      import numpy as np
      import json
      import io
      from PIL import Image
      
      def input_handler(data, context):
          """ Pre-process request input before it is sent to TensorFlow Serving REST API
          
          Args:
          data (obj): the request data, in format of dict or string
          context (Context): an object containing request and configuration details
          
          Returns:
          (dict): a JSON-serializable dict that contains request body and headers
          """
          f = data.read()
          f = io.BytesIO(f)
          image = Image.open(f).convert('RGB')
          batch_size = 1
          image = np.asarray(image.resize((512, 512)))
          image = np.concatenate([image[np.newaxis, :, :]] * batch_size)
          body = json.dumps({"signature_name": "serving_default", "instances": image.tolist()})
          return body
      
      def output_handler(data, context):
          """Post-process TensorFlow Serving output before it is returned to the client.
          
          Args:
          data (obj): the TensorFlow serving response
          context (Context): an object containing request and configuration details
          
          Returns:
          (bytes, string): data to return to client, response content type
          """
          if data.status_code != 200:
              raise ValueError(data.content.decode('utf-8'))
      
          response_content_type = context.accept_header
          prediction = data.content
          return prediction, response_content_type
      ```

      如果沒有自訂的預處理或後處理，SageMaker AI 用戶端會以類似的方式將檔案映像轉換為 JSON，然後再將其傳送到 SageMaker AI 端點。

      如需更多資訊，請參閱[部署至在 SageMaker Python SDK 服務端點的 TensorFlow](https://sagemaker.readthedocs.io/en/stable/frameworks/tensorflow/deploying_tensorflow_serving.html#providing-python-scripts-for-pre-pos-processing)。

1. 包含已編譯模型成品的 Amazon S3 儲存貯體 URI。