Merge pull request #8 from zeroae/f/sagemaker

Add SageMaker Inference Toolkit support

.github/condarc.yml

@@ -1,6 +1,5 @@
 anaconda_upload: false
 channels:
   - zeroae
-  - defaults
   - conda-forge
-show_channel_urls: true
+show_channel_urls: true

.github/workflows/pypa-conda.yml

@@ -22,13 +22,13 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest]
-        python-version: [3.6]
+        python-version: [3.7]
         black-version: [19.10b]
         flake8-version: [3.7.9]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v1
-      - uses: actions/setup-python@v1.1.1
+      - uses: actions/setup-python@v2.2.2
         with:
           python-version: ${{ matrix.python-version }}
       - uses: actions/cache@v1
@@ -75,7 +75,7 @@ jobs:
           rm -f *.tar.gz
 
       # Create the bdist wheel file
-      - uses: actions/setup-python@v1.1.1
+      - uses: actions/setup-python@v2.2.2
       - uses: actions/cache@v1
         id: cache
         with:
@@ -152,7 +152,7 @@ jobs:
 
       #
       # Setup the test environment, python + .whl + .whl[test]
-      - uses: actions/setup-python@v1.1.1
+      - uses: actions/setup-python@v2.2.2
         with:
           python-version: ${{ matrix.python-version }}
       - name: Get pip cache
@@ -211,13 +211,16 @@ jobs:
           tar -xvf *.tar.gz --strip 1
           rm *.tar.gz
 
-      - uses: goanpeca/setup-miniconda@v1
+      - uses: conda-incubator/setup-miniconda@v2
         with:
           activate-environment: ''
           auto-activate-base: true
-          conda-build-version: 3.18
+          miniforge-variant: Mambaforge
+          use-mamba: true
+          conda-build-version: 3.21.4
           condarc-file: .github/condarc.yml
-      - run: conda install setuptools_scm conda-verify
+      - run: |
+          mamba install setuptools_scm conda-verify boa
       - uses: actions/cache@v1
         id: conda-pkgs-cache
         with:
@@ -228,7 +231,7 @@ jobs:
       - name: Run conda build
         run: |
           mkdir conda-bld
-          conda build --output-folder conda-bld .
+          conda mambabuild --output-folder conda-bld .
         env:
           ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}
       - name: Create conda-bld/manifest
@@ -338,7 +341,7 @@ jobs:
         with:
           path: /usr/share/miniconda/pkgs
           key: ${{ runner.os }}-conda-ac1.7.2
-      - uses: goanpeca/setup-miniconda@v1
+      - uses: conda-incubator/setup-miniconda@v2
         with:
           activate-environment: ''
           auto-activate-base: true

.gitignore

@@ -1,3 +1,9 @@
+# MacOS
+.DS_Store
+
+# Vagrant
+.vagrant
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

Makefile

@@ -93,10 +93,10 @@ wheels: dist  ## downloads wheel dependencies
 	ls -l wheels
 
 dist-conda:  ## builds conda-package
-	conda build --no-anaconda-upload --output-folder conda-bld \
+	conda mambabuild --no-anaconda-upload --output-folder conda-bld \
 	  -c zeroae \
 	  -c conda-forge \
-	  -c anaconda .
+	  .
 
 install: clean ## install the package to the active Python's site-packages
-	python setup.py install
+	python setup.py install

Vagrantfile

@@ -0,0 +1,94 @@
+# -*- mode: ruby -*-
+# vi: set ft=ruby :
+
+# All Vagrant configuration is done below. The "2" in Vagrant.configure
+# configures the configuration version (we support older styles for
+# backwards compatibility). Please don't change it unless you know what
+# you're doing.
+Vagrant.configure("2") do |config|
+  # The most common configuration options are documented and commented below.
+  # For a complete reference, please see the online documentation at
+  # https://docs.vagrantup.com.
+
+  # Every Vagrant development environment requires a box. You can search for
+  # boxes at https://vagrantcloud.com/search.
+  config.vm.box = "hashicorp/bionic64"
+
+  # Disable automatic box update checking. If you disable this, then
+  # boxes will only be checked for updates when the user runs
+  # `vagrant box outdated`. This is not recommended.
+  # config.vm.box_check_update = false
+
+  # Create a forwarded port mapping which allows access to a specific port
+  # within the machine from a port on the host machine. In the example below,
+  # accessing "localhost:8080" will access port 80 on the guest machine.
+  # NOTE: This will enable public access to the opened port
+  config.vm.network "forwarded_port", guest: 8888, host: 8888
+
+  # Create a forwarded port mapping which allows access to a specific port
+  # within the machine from a port on the host machine and only allow access
+  # via 127.0.0.1 to disable public access
+  # config.vm.network "forwarded_port", guest: 80, host: 8080, host_ip: "127.0.0.1"
+
+  # Create a private network, which allows host-only access to the machine
+  # using a specific IP.
+  # config.vm.network "private_network", ip: "192.168.33.10"
+
+  # Create a public network, which generally matched to bridged network.
+  # Bridged networks make the machine appear as another physical device on
+  # your network.
+  # config.vm.network "public_network"
+
+  # Share an additional folder to the guest VM. The first argument is
+  # the path on the host to the actual folder. The second argument is
+  # the path on the guest to mount the folder. And the optional third
+  # argument is a set of non-required options.
+  # config.vm.synced_folder "../data", "/vagrant_data"
+
+  # Provider-specific configuration so you can fine-tune various
+  # backing providers for Vagrant. These expose provider-specific options.
+  # Example for VirtualBox:
+  #
+  # config.vm.provider "virtualbox" do |vb|
+  #   # Display the VirtualBox GUI when booting the machine
+  #   vb.gui = true
+  #
+  #   # Customize the amount of memory on the VM:
+  #   vb.memory = "1024"
+  # end
+  #
+  # View the documentation for the provider you are using for more
+  # information on available options.
+  config.vm.provider "vmware_desktop" do |v|
+    v.vmx["memsize"] = "2048"
+  end
+
+  # Enable provisioning with a shell script. Additional provisioners such as
+  # Ansible, Chef, Docker, Puppet and Salt are also available. Please see the
+  # documentation for more information about their specific syntax and use.
+  config.vm.provision "shell", name: "Install Mambaforge", privileged: false, reset: true, inline: <<-SHELL
+    MAMBA_FORGE_FILE=Mambaforge-$(uname)-$(uname -m).sh
+    if ! [ -d ~/mambaforge ]; then
+      if ! [ -f $MAMBA_FORGE_FILE ]; then
+        wget -q https://github.com/conda-forge/miniforge/releases/latest/download/$MAMBA_FORGE_FILE
+      fi
+      bash $MAMBA_FORGE_FILE -b -u
+      rm -f $MAMBA_FORGE_FILE
+
+      mambaforge/bin/conda init --all
+    fi
+  SHELL
+
+  config.vm.provision "shell", name: "Install OS Packages", inline: <<-SHELL
+    ### Add OpenJDK 8
+    apt-get update
+    apt-get --yes install openjdk-8-jre-headless
+  SHELL
+
+  config.vm.provision "shell", name: "Create Development Environment", privileged: false, inline: <<-SHELL
+    ### Create the DarkNet Environment
+    source mambaforge/etc/profile.d/conda.sh
+    mamba env update --name darknet-cpu -f /vagrant/environment.yml
+    echo 'conda activate darknet-cpu' >> ~/.bashrc
+  SHELL
+end

environment.yml

@@ -1,4 +1,3 @@
-
 name: darknet.py-dev
 channels:
   - zeroae
@@ -9,6 +8,8 @@ dependencies:
   - pip
   - pip:
     - -e .
+    - multi-model-server
+    - sagemaker-inference
 
   # Setup Requirements (setup.py:setup_requirements)
   - compilers
@@ -21,7 +22,7 @@ dependencies:
   # Install Requirements (setup.py:requirements)
   - click >=7.0
   - click-plugins
-  - darknet
+  - darknet-cpu
   - entrypoints
   - fsspec <=0.7.5
   - numpy
@@ -30,6 +31,12 @@ dependencies:
   # Zoo Optional Requirements
   - intake
 
+  # MMS Requirements
+  - enum-compat
+  - future
+  - retrying
+  - scipy
+
   # Test Requirements (setup.py:test_requirements)
   - pytest >=3
   - pytest-cov

setup.py

@@ -47,6 +47,14 @@
     # fmt: on
 ]
 
+mms_requirements = [
+    # fmt: off
+    "future",
+    "multi-model-server",
+    "retrying",
+    "sagemaker-inference",
+    # fmt: on
+]
 zoo_requirements = [
     # fmt: off
     "intake",
@@ -130,7 +138,7 @@
     long_description_content_type="text/x-rst",
     include_package_data=True,
     keywords="py darknet",
-    name="darknet-py",
+    name="darknet.py",
     package_dir={"": "src"},
     packages=find_namespace_packages(where="./src"),
     setup_requires=setup_requirements,
@@ -141,6 +149,7 @@
         "test": test_requirements,
         "doc": doc_requirements,
         "zoo": zoo_requirements,
+        "mms": mms_requirements,
         # fmt: on
     },
     url="https://github.com/zeroae/darknet.py",

src/darknet/sagemaker/__init__.py

@@ -0,0 +1,4 @@
+from .default_inference_handler import DefaultDarknetInferenceHandler, Network
+from ..py.util import image_to_3darray
+
+__all__ = ["DefaultDarknetInferenceHandler", "Network", "image_to_3darray"]

src/darknet/sagemaker/__main__.py

@@ -0,0 +1,26 @@
+from retrying import retry
+from subprocess import CalledProcessError
+from sagemaker_inference import model_server
+
+# TODO: from .classifier import handler_service as classifier_service
+from .detector import handler_service as detector_service
+
+
+def _retry_if_error(exception):
+    return isinstance(exception, CalledProcessError or OSError)
+
+
+@retry(stop_max_delay=1000 * 50, retry_on_exception=_retry_if_error)
+def _start_mms():
+    # by default the number of workers per model is 1, but we can configure it through the
+    # environment variable below if desired.
+    # os.environ['SAGEMAKER_MODEL_SERVER_WORKERS'] = '2'
+    # TODO: Start Classifier *or* Detector Service
+    model_server.start_model_server(handler_service=detector_service.__name__)
+
+
+def main():
+    _start_mms()
+
+
+main()

src/darknet/sagemaker/classifier/__init__.py

@@ -0,0 +1,3 @@
+from .handler_service import HandlerService
+
+__all__ = ["HandlerService"]

src/darknet/sagemaker/classifier/__main__.py

@@ -0,0 +1,25 @@
+from retrying import retry
+from subprocess import CalledProcessError
+from sagemaker_inference import model_server
+
+from . import handler_service as classifier_service
+
+
+def _retry_if_error(exception):
+    return isinstance(exception, CalledProcessError or OSError)
+
+
+@retry(stop_max_delay=1000 * 50, retry_on_exception=_retry_if_error)
+def _start_mms():
+    # by default the number of workers per model is 1, but we can configure it through the
+    # environment variable below if desired.
+    # os.environ['SAGEMAKER_MODEL_SERVER_WORKERS'] = '2'
+    # TODO: Start Classifier *or* Detector Service
+    model_server.start_model_server(handler_service=classifier_service.__name__)
+
+
+def main():
+    _start_mms()
+
+
+main()

src/darknet/sagemaker/classifier/default_inference_handler.py

@@ -0,0 +1,36 @@
+from typing import Tuple, List
+
+from sagemaker_inference.errors import UnsupportedFormatError
+
+from .. import DefaultDarknetInferenceHandler, image_to_3darray, Network
+
+
+class DefaultDarknetClassifierInferenceHandler(DefaultDarknetInferenceHandler):
+    def default_predict_fn(self, data, model: Tuple[Network, List[str]]):
+        """A default predict_fn for DarkNet. Calls a model on data deserialized in input_fn.
+        Args:
+            data: input data (PIL.Image) for prediction deserialized by input_fn
+            model: Darknet model loaded in memory by model_fn
+
+        Returns: a prediction
+        """
+        network, labels = model
+        max_labels = data.get("MaxLabels", 5)
+        # TODO: min_confidence = data.get("MinConfidence", 55)
+
+        if "NDArray" in data:
+            probabilities = network.predict(data["NDArray"])
+        elif "Image" in data:
+            image, _ = image_to_3darray(data["Image"], network.shape)
+            probabilities = network.predict_image(image)
+        else:
+            raise UnsupportedFormatError("Expected an NDArray or an Image")
+
+        rv = [
+            {
+                "Name": label,
+                "Confidence": prob * 100,
+            }
+            for label, prob in sorted(zip(labels, probabilities), key=lambda x: x[1], reverse=True)
+        ]
+        return {"Labels": rv[0:max_labels] if max_labels else rv}

src/darknet/sagemaker/classifier/handler_service.py

@@ -0,0 +1,22 @@
+from sagemaker_inference.default_handler_service import DefaultHandlerService
+from sagemaker_inference.transformer import Transformer
+
+from .default_inference_handler import DefaultDarknetClassifierInferenceHandler
+
+
+class HandlerService(DefaultHandlerService):
+    """Handler service that is executed by the model server.
+    Determines specific default inference handlers to use based on the type MXNet model being used.
+    This class extends ``DefaultHandlerService``, which define the following:
+        - The ``handle`` method is invoked for all incoming inference requests to the model server.
+        - The ``initialize`` method is invoked at model server start up.
+    Based on: https://github.com/awslabs/mxnet-model-server/blob/master/docs/custom_service.md
+    """
+
+    def __init__(self):
+        self._initialized = False
+
+        transformer = Transformer(
+            default_inference_handler=DefaultDarknetClassifierInferenceHandler()
+        )
+        super(HandlerService, self).__init__(transformer=transformer)