How can we specify nested python class in .s2i/environment?

[yu-iskw]

Motivation

We want to manage the training code and the serving code as well. But, we want to manage the files in a directory. So, our question is can we specify such a nested python class in .s2i/environment or not.

Expectaed Behavior

We want to specify a nested class as below.

Example .s2i/environment

MODEL_NAME=src.v1.IrisClassifier
API_TYPE=REST
SERVICE_TYPE=MODEL
PERSISTENCE=0

Example Directory Structure

.
├── IrisClassifier.sav
├── contract.json
├── requirements.txt
├── sklearn_iris.ipynb
├── sklearn_iris_deployment.json
├── src
│   └── v1
│       └── IrisClassifier.py
│       └── train_iris.py

Error Message

starting microservice
2019-03-13 17:44:26,507 - seldon_core.microservice:main:261 - INFO:  Starting microservice.py:main
2019-03-13 17:44:26,510 - seldon_core.microservice:main:292 - INFO:  Annotations: {}
Traceback (most recent call last):
  File "/usr/local/bin/seldon-core-microservice", line 10, in <module>
    sys.exit(main())
  File "/usr/local/lib/python3.6/site-packages/seldon_core/microservice.py", line 295, in main
    user_class = getattr(interface_file, args.interface_name)
AttributeError: module 'src.v1.IrisClassifier' has no attribute 'src.v1.IrisClassifier'

[ukclivecox]

We use Importlib specifically importlib.import_module which should handle the dot notation. We have not tested this case, but we definitely need a story to provide this.

[jlewi]

Won't we have problems here

seldon-core/python/seldon_core/microservice.py

Line 188 in 6b30fb6

user_class = getattr(interface_file, args.interface_name)

if interface_name contains a dotted path.

e.g. in this case suppose interface_name = "v1.IrisClassifier"

So we end up calling

    interface_file = importlib.import_module("v1.IrisClassifier")
    user_class = getattr(interface_file, "v1.IrisClassifier)

The call to getattr is using the full module dotted name of the package which doesn't match the class name; I think it should be

  user_class = getattr(interface_file, "IrisClassifier")

Would it make sense to allow the module path and class to be specified separately?

One option would be to just assume everything after the last "." is the class name and everything before is the module you could then parse it doing something like

full_module, user_class = args.interface_name.rsplit(".", 1)