Skip to main content

Intro

CLI app provides functionality to deploy custom models onto fma infrastructure. It employs features of pydantic library extended with custom field types to describe model's io. It interacts with fma services to send your models code and to run it on fma's hardware.

Quickstart​

Installation​

pip install git+https://github.com/flymyai/fma-custom-models.git

Initialization​

You have to create a new folder and run fma init <your_model_name> inside of it. This will create one more folder with your model name and a model.py file inside of it. The model.py will be prepolutated with:

from typing import List

from pydantic import BaseModel

from fma.toolkit import model as fma_model


class Input(BaseModel):
    pass


class Output(BaseModel):
    pass


class Model(fma_model.Model):
    requirements: List[str] = []

    def initialize(self):
        pass

    def predict(self, input: Input) -> Output:
        return {}

IMPORTANT NOTE: even though io is described via pydantic models they are not actually supported in the predict method. Instead predict method accepts and returns python dicts. Support for pydantic models will be added eventually.

Model development​

During this step you simply have to populate the initialize and predict methods with inferences logic. You can set instance attributes via accessing self, all the imports should be inside of the methods. You can also specify required libraries and packages in the requirements field of Model class.

Deploying the model​

If you have finished all the steps above all you need to do is to:

Log in​

You can do this via fma login --username <username> --password <password> or by providing one/none, missing ones will be prompted.

Actually deploy​

By running fma deploy After this the data about your model will be written into metadata.yaml file.

Extra info​

Additional fields​

You can specify image or file fields in the io classes by:

from fma.toolkit.fields.image import Image

When you use this field in your code you treat it as a pillow's Image.Image object. So when you access it from the input it is an instance of Image.Image class and if you want to return an image the value of the field that is specified as image should be an instance of Image.Image class

or

from fma.toolkit.fields.file import File

The previous logic applies for File field, but instead of pillow's object you should io.BytesIO

Note that File is currently supported only as the output

Project structure​

Your project structure should be follow the next schema:

<your_model_name>
    | model.py
metadata.yaml

So there should not be any other folder in the same directory with the <your_model_name> folder, otherwise it will cause an error.