ibm / max-framework

Python package that can be installed to make it easier to create MAX models

Model Asset Exchange Framework

Python package that contains common code shared across all MAX models.

Dependencies

• flask-restx
• flask-cors

Installation

The package can be installed with pip. However, this is not necessary as each MAX model will get the maxfw package via the MAX-Base image.

If you want to run a MAX model outside of a Docker container then you can install it with the following command:

$ pip install -U maxfw

Usage

For an example of this package being used in a MAX model, we recommend looking at the MAX-Skeleton Repository on GitHub.

max-framework's Issues

Add image utils

Many MAX models are based on image input. They all require basic utils such as:

• load image
• handle different formats (JPG, PNG etc)
• handle errors (trying to pass non-supported image, or non-image as input)
• possibly conversion (PNG -> RGB; color -> grayscale etc)

See e.g. https://github.com/IBM/MAX-Facial-Age-Estimator/issues/13 and IBM/MAX-Human-Pose-Estimator#12 (but this applies to most if not all MAX models that operate on image input).

Add badge for PyPi

See the PyPi badge in the table in https://github.ibm.com/CODAIT/MAX-development-guide/blob/master/MODEL_OWNERS.md#other-max-repos

CORS_ENABLE flag inactive due to Debug Mode being turned off

Previously, the models used to run with Flask Debug mode turned on. This caused some of the server code to execute twice, so I put a second condition in the conditional statement here

MAX-Framework/maxfw/core/app.py

Line 29 in 1a924f2

if os.getenv('CORS_ENABLE') == 'true' and os.environ.get('WERKZEUG_RUN_MAIN') == 'true':

Now that Debug mode is turned off, the CORS_ENABLE flag will not actually enable cross origin support. I think it might be best to simply remove the second condition.

Add support for 'port' parameter on MAXApp.run() method

In MAX-Framework/maxfw/core/app.py, a Flask app is instantiated. There is a parameter for 'port' that is not available through the abstracted MAXApp.run() method.

It is documented here:
https://flask.palletsprojects.com/en/1.1.x/api/#flask.Flask.run

Perhaps something like this would add support for alternate ports when using MAXApp.run()

def run(self, host='0.0.0.0', port=5000): self.app.run(host, port)

Add common code for certain model classes

Many MAX models fall into a few common categories - classification (multi-class typically), object detection / segmentation (with bounding boxes and/or segmentation maps), etc.

We could create utilities and/or some abstract base classes that handle common code for things such as response mapping to JSON (e.g. for multi-class classification it is generic) and definitions for Swagger API formats.

Add detail to PyPi project page

Currently the PyPi project page contains no description and the project name should probably be Model Asset Exchange Framework?

Make app name configurable

This line: https://github.com/IBM/MAX-Framework/blob/master/maxfw/core/app.py#L13

means that the app name when serving in Flask looks like:

 * Serving Flask app "maxfw.core.app" (lazy loading)
 * Environment: production
...

Not too major, but would be nice to make the app name a configurable property.

Enable custom origin CORS support

Currently the ENABLE_CORS option opens up all endpoints for all domains: https://github.com/IBM/MAX-Framework/blob/master/maxfw/core/app.py#L46 We should make it possible for users to specify custom origins without the need to modify the source code: https://flask-cors.corydolphin.com/en/latest/api.html#extension

Add support for basic authentication

Currently the microservice endpoints are unprotected and can be called by anybody. We should add a simple mechanism that allows for optional basic authorization checking. A simple example of such a mechanism could be a static API token.

PredictAPI not completed

I am working with the MAX-Audio-Classifier and I am trying to understand the PredictAPI that is imported from maxfw.core in the predict.py file.
When checking the maxfw.core repo I find that the PredictAPI is not implemented.

Have I missed the PredictAPI or how does the audio-classifier import and use the PredictAPI if it has not been implemented yet?

Thanks,
Johannes

Auto-deploy new releases on PyPI

Do the same thing we did here IBM/MAX-Training-Framework#8

MAXModelWrapper: remove sbtractmethod annotations from optional methods

PR #21 enforces the @abstractmethod annotations which are defined for the following methods:

• __init__
• _pre_process
• _post_process
• _predict

As a result every model that inherits from MAXModelWrapper must implement these methods, even if they don't have any custom code.

The problematic code was published as version 1.1.2.

Add some detail to README

• installation instructions (pip install from git for now)
• usage

Add detail to PyPi project page

Currently the PyPi project page contains no description and the project name should probably be Model Asset Exchange Framework?

update image_utils.py error messages

See the discussion here:

#33 (comment)