PESTO Cookbook

How to load resources only once (at server start) : I don't want to reload my model for each prediction ?

Use static variables in Process class. Ex:

class Process:
    heavy_requirement = None

    def process(self, *args, **kwargs):
        # load only once
        if Process.heavy_requirement is None:
            Process.heavy_requirement = ...

        # use heavy_requirement for processing
        ... 

How to use profiles to factorize PESTO configurations ?

Imagine you want to build two services out of one algorithm implementation :

• cpu : run on CPU with a specific model and requirements,
• gpu : run on GPU with another model and requirements.

In 'pesto/build/' you can use three files to define requirements :

• requirements.json : common requirements,
• requirements.cpu.json : specific CPU requirements,
• requirements.gpu.json : specific GPU requirements.

Then build your service with the '--profile' or '-p' option :

pesto build path/to/project -p cpu
pesto build path/to/project -p gpu

More details in the profile section of the PESTO documentation.

How to use profiles to build variants of a same algorithm ?

For exemple you want to build your algorithm with 2 variants :

• raster : return detections as an image mask,
• vector : return detections as a geometry.

Build your service with the '--profile' or '-p' option :

pesto build path/to/project -p raster cpu
pesto build path/to/project -p vector cpu

Use 'Pesto.is_profile_active(profile:str)' to check at runtime which profile was used during build.

from pesto.common.pesto import Pesto
class Process:
    def process(self, *args, **kwargs):
        mask_output = ...
        if Pesto.is_profile_active('raster'):
            return mask_output
        if Pesto.is_profile_active('vector'):
            return vectorize(mask_output)
        raise NotImplementedError()

How to install / include files in the docker image ?

Pesto will copy or pip install all your requirements in the output docker image.

You just need to define all your service requirements in the 'pesto/build/requirements.json' file.

How to test the service built with PESTO ?

Be sure to have a proper pesto-service python project (use the pesto init command). Then, go in your project pesto/tests directory and start editing files.

The pesto/tests is composed of :

• some directories (one per processing to be run)
• a expected_describe.json file

Each pesto/tests/xxx directory is composed of :

• an input directory matching pesto/api/input_schema.json,
• an output directory matching pesto/api/output_schema.json.

The input and output directories both describes a json payload (the processing input and output). Each filename key.type in those folders must match an entry in its corresponding *_schema.json :

• key is the key in the *_schema.json,
• type is the primitive type of the key :
  • string, float, int,
  • json : dictionaries
  • .tif, .jpg, *.png for images.
  • arrays can be constructed using a folder key containing its enumerted items (1.int, 2.int, ...)

{
    "key1" : "text",
    "key2" : 33,
    "key3" : 3.14
}

More examples are provided in the default pesto template.

Then, it is required to build your project (once).

pesto build /path/to/pesto-service -p p1 p2

Finally, run the tests :

pesto test /path/to/pesto-service -p p1 p2