change: add serving script for mxnet (#882)

* add inference script

Author: GaryTu1020

.gitignore

@@ -0,0 +1,28 @@
+.idea
+build
+src/*.egg-info
+.cache
+.coverage
+sagemaker_venv*
+*.egg-info
+.tox
+**/__pycache__
+**/.ipynb_checkpoints
+dist/
+**/tensorflow-examples.tar.gz
+**/*.pyc
+**.pyc
+scratch*.py
+.eggs
+*.egg
+examples/tensorflow/distributed_mnist/data
+*.iml
+doc/_build
+doc/_static
+doc/_templates
+**/.DS_Store
+venv/
+*~
+.pytest_cache/
+*.swp
+.docker/

README.md

@@ -1,2 +1,2 @@
-# python-sdk-testing
+# test-branch-git-config
 It's a repo for testing the sagemaker Python SDK Git support

mxnet/mnist_hosting_with_custom_handlers.py

@@ -0,0 +1,127 @@
+# Copyright 2017-2018 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+from __future__ import absolute_import
+
+import gzip
+import json
+import mxnet as mx
+import numpy as np
+import os
+import struct
+
+
+# --- this example demonstrates how to extend default behavior during model hosting ---
+
+# --- Model preparation ---
+# it is possible to specify own code to load the model, otherwise a default model loading takes place
+def model_fn(path_to_model_files):
+    from mxnet.io import DataDesc
+
+    loaded_symbol = mx.symbol.load(os.path.join(path_to_model_files, "symbol"))
+    created_module = mx.mod.Module(symbol=loaded_symbol)
+    created_module.bind([DataDesc("data", (1, 1, 28, 28))])
+    created_module.load_params(os.path.join(path_to_model_files, "params"))
+    return created_module
+
+
+# --- Option 1 - provide just 1 entry point for end2end prediction ---
+# if this function is specified, no other overwriting described in Option 2 will have effect
+# returns serialized data and content type it has used
+def transform_fn(model, request_data, input_content_type, requested_output_content_type):
+    # for demonstration purposes we will be calling handlers from Option2
+    return (
+        output_fn(
+            process_request_fn(model, request_data, input_content_type),
+            requested_output_content_type,
+        ),
+        requested_output_content_type,
+    )
+
+
+# --- Option 2 - overwrite container's default input/output behavior with handlers ---
+# there are 2 data handlers: input and output, you need to conform to their interface to fit into default execution
+def process_request_fn(model, data, input_content_type):
+    if input_content_type == "text/s3_file_path":
+        prediction_input = handle_s3_file_path(data)
+    elif input_content_type == "application/json":
+        prediction_input = handle_json_input(data)
+    else:
+        raise NotImplementedError(
+            "This model doesnt support requested input type: " + input_content_type
+        )
+
+    return model.predict(prediction_input)
+
+
+# for this example S3 path points to a file that is same format as in test/images.gz
+def handle_s3_file_path(path):
+    import sys
+
+    if sys.version_info.major == 2:
+        import urlparse
+
+        parse_cmd = urlparse.urlparse
+    else:
+        import urllib
+
+        parse_cmd = urllib.parse.urlparse
+
+    import boto3
+    from botocore.exceptions import ClientError
+
+    # parse the path
+    parsed_url = parse_cmd(path)
+
+    # get S3 client
+    s3 = boto3.resource("s3")
+
+    # read file content and pass it down
+    obj = s3.Object(parsed_url.netloc, parsed_url.path.lstrip("/"))
+    print("loading file: " + str(obj))
+
+    try:
+        data = obj.get()["Body"]
+    except ClientError as ce:
+        raise ValueError(
+            "Can't download from S3 path: " + path + " : " + ce.response["Error"]["Message"]
+        )
+
+    import StringIO
+
+    buf = StringIO(data.read())
+    img = gzip.GzipFile(mode="rb", fileobj=buf)
+
+    _, _, rows, cols = struct.unpack(">IIII", img.read(16))
+    images = np.fromstring(img.read(), dtype=np.uint8).reshape(10000, rows, cols)
+    images = images.reshape(images.shape[0], 1, 28, 28).astype(np.float32) / 255
+
+    return mx.io.NDArrayIter(images, None, 1)
+
+
+# for this example it is assumed that the client is passing data that can be "directly" provided to the model
+def handle_json_input(data):
+    nda = mx.nd.array(json.loads(data))
+    return mx.io.NDArrayIter(nda, None, 1)
+
+
+def output_fn(prediction_output, requested_output_content_type):
+    # output from the model is NDArray
+
+    data_to_return = prediction_output.asnumpy()
+
+    if requested_output_content_type == "application/json":
+        json.dumps(data_to_return.tolist), requested_output_content_type
+
+    raise NotImplementedError(
+        "Model doesn't support requested output type: " + requested_output_content_type
+    )