Skip to content

Latest commit

 

History

History
188 lines (143 loc) · 9.78 KB

README.md

File metadata and controls

188 lines (143 loc) · 9.78 KB
Raw

Torch Model archiver for TorchServe

Contents of this Document

Overview

A key feature of TorchServe is the ability to package all model artifacts into a single model archive file. It is a separate command line interface (CLI), torch-model-archiver, that can take model checkpoints or model definition file with state_dict, and package them into a .mar file. This file can then be redistributed and served by anyone using TorchServe. It takes in the following model artifacts: a model checkpoint file in case of torchscript or a model definition file and a state_dict file in case of eager mode, and other optional assets that may be required to serve the model. The CLI creates a .mar file that TorchServe's server CLI uses to serve the models.

Important: Make sure you try the Quick Start: Creating a Model Archive tutorial for a short example of using torch-model-archiver.

The following information is required to create a standalone model archive:

  1. Model name
  2. Model file
  3. Serialized file

Installation

Install torch-model-archiver as follows:

pip install torch-model-archiver

Installation from source

Install torch-model-archiver as follows:

git clone https://github.com/pytorch/serve.git
cd serve/model-archiver
pip install .

Torch Model Archiver Command Line Interface

Now let's cover the details on using the CLI tool: model-archiver.

Here is an example usage with the densenet161 model archive following the example in the examples README:

torch-model-archiver --model-name densenet161 --version 1.0 --model-file examples/image_classifier/densenet_161/model.py --serialized-file densenet161-8d451a50.pth --extra-files examples/image_classifier/index_to_name.json --handler image_classifier

Arguments

$ torch-model-archiver -h
usage: torch-model-archiver [-h] --model-name MODEL_NAME  --version MODEL_VERSION_NUMBER
                      --model-file MODEL_FILE_PATH --serialized-file MODEL_SERIALIZED_PATH
                      --handler HANDLER [--runtime {python,python2,python3}]
                      [--export-path EXPORT_PATH] [-f] [--requirements-file]

Model Archiver Tool

optional arguments:
  -h, --help            show this help message and exit
  --model-name MODEL_NAME
                        Exported model name. Exported file will be named as
                        model-name.mar and saved in current working directory
                        if no --export-path is specified, else it will be
                        saved under the export path
  --serialized-file SERIALIZED_FILE
                        Path to .pt or .pth file containing state_dict in
                        case of eager mode or an executable ScriptModule
                        in case of TorchScript.
  --model-file MODEL_FILE
                        Path to python file containing model architecture.
                        This parameter is mandatory for eager mode models.
                        The model architecture file must contain only one
                        class definition extended from torch.nn.modules.
  --handler HANDLER     TorchServe's default handler name  or handler python
                        file path to handle custom TorchServe inference logic.
  --extra-files EXTRA_FILES
                        Comma separated path to extra dependency files.
  --runtime {python,python2,python3}
                        The runtime specifies which language to run your
                        inference code on. The default runtime is
                        RuntimeType.PYTHON. At the present moment we support
                        the following runtimes python, python2, python3
  --export-path EXPORT_PATH
                        Path where the exported .mar file will be saved. This
                        is an optional parameter. If --export-path is not
                        specified, the file will be saved in the current
                        working directory.
  --archive-format {tgz,default}
                        The format in which the model artifacts are archived.
                        "tgz": This creates the model-archive in <model-name>.tar.gz format.
                        If platform hosting requires model-artifacts to be in ".tar.gz"
                        use this option.
                        "no-archive": This option creates an non-archived version of model artifacts
                        at "export-path/{model-name}" location. As a result of this choice,
                        MANIFEST file will be created at "export-path/{model-name}" location
                        without archiving these model files
                        "default": This creates the model-archive in <model-name>.mar format.
                        This is the default archiving format. Models archived in this format
                        will be readily hostable on TorchServe.
  -f, --force           When the -f or --force flag is specified, an existing
                        .mar file with same name as that provided in --model-
                        name in the path specified by --export-path will
                        overwritten
  -v, --version         Model's version.
  -r, --requirements-file
                        Path to requirements.txt file containing a list of model specific python
                        packages to be installed by TorchServe for seamless model serving.

Artifact Details

MAR-INF

MAR-INF is a reserved folder name that will be used inside .mar file. This folder contains the model archive metadata files. Users should avoid using MAR-INF in their model path.

Runtime

Model name

A valid model name must begin with a letter of the alphabet and can only contains letters, digits, underscores _, dashes - and periods ..

Note: The model name can be overridden when you register the model with Register Model API.

Model file

A model file should contain the model architecture. This file is mandatory in case of eager mode models.

Serialized file

A serialized file (.pt or .pth) should be a checkpoint in case of torchscript and state_dict in case of eager mode.

Handler

Handler can be TorchServe's inbuilt handler name or path to a py file to handle custom TorchServe inference logic. TorchServe supports following handlers out or box:

  1. image_classifier
  2. object_detector
  3. text_classifier
  4. image_segmenter

In case of custom handler, if you plan to provide just module_name or module_name:entry_point_function_name then make sure that it is prefixed with absolute or relative path of python file. e.g. if your custom handler custom_image_classifier.py is in /home/serve/examples then --handler /home/serve/examples/custom_image_classifier or if it has my_entry_point module level function then --handler /home/serve/examples/custom_image_classifier:my_entry_point_func

For more details refer default handler documentation or custom handler documentation

Creating a Model Archive

1. Download the torch model archiver source

git clone https://github.com/pytorch/serve.git

2. Package your model

With the model artifacts available locally, you can use the torch-model-archiver CLI to generate a .mar file that can be used to serve an inference API with TorchServe.

In this next step we'll run torch-model-archiver and tell it our model's name is densenet_161 and its version is 1.0 with the model-name and version parameter respectively and that it will use TorchServe's default image_classifier handler with the handler argument . Then we're giving it the model-file and serialized-file to the model's assets.

For torchscript:

torch-model-archiver --model-name densenet_161 --version 1.0 --serialized-file model.pt --handler image_classifier

For eagermode:

torch-model-archiver --model-name densenet_161 --version 1.0 --model-file model.py --serialized-file model.pt --handler image_classifier

This will package all the model artifacts files and output densenet_161.mar in the current working directory. This .mar file is all you need to run TorchServe, serving inference requests for a simple image recognition API. Go back to the Serve a Model tutorial and try to run this model archive that you just created!

Model specific custom python requirements

Custom models/handlers may depend on different python packages which are not installed by-default as a part of TorchServe setup. Supply a python requirements file containing the list of required python packages to be installed by TorchServe for seamless model serving using --requirements-file parameter while creating the model-archiver.

Example:

torch-model-archiver --model-name densenet_161 --version 1.0 --model-file model.py --serialized-file model.pt --handler image_classifier --requirements-file <path_to_custom_requiremnets_file>

Note: This feature is by-default disabled in TorchServe and needs to be enabled through configuration. For more details refer TorchServe's configuration documentation