Ramp Sagemaker CookieCutter

Usage

This repo acts as centralised starting point for all Sagemaker Models at ramp.

To use run the following:

pip install cookiecutter==2.3.0
cookiecutter gh:rampgrowth/ds-sagemaker-template

This will generate the following package structure on your local machine:

<project_name>
├── README.md
├── main.py
├── src
│   ├── __init__.py
│   ├── base.py
│   ├── data.py
│   ├── loss.py
│   └── model.py
├── requirements.txt
├── requirements_dev.txt
└── config.json

Overview

The following modules are shipped with the base cookiecutter template. All files can be modfied based on user prefference or updated to implementation specific requirements.

NB: However, at the very miniumum the following modules need to be present:

<project_name>
├── src
│   ├── base.py
│   ├── data.py
│   ├── loss.py
│   └── model.py
├── main.py
└── requirements.txt

train.py

This is the main entypoint used by sagemaker to train your defined model

models.py and base.py

Custom torch models need to defined in this module. All models need to inherit from base.BaseNet which impliments the following:

1. Pytorch Categorical Embeddings
2. training and validation steps
3. Weights and Biases logging
4. Optimisers
5. ONNX:
   • example_input_array -> Used to generate the ONNX model
   • export_model -> Model export definitions

data.py

Dataloader and data-prep defintions

features.py

Datatransformation operations. Please bear in mind all modules imported in this folder, need to exist on the inference machine at production runtime.

For instance the base template makes use of:

import pandas as pd
import numpy as np
import sklearn

This implies that a target machine (engine/runway) will have these modules installed. Please ensure this is the case.

Relative imports are strictly not allowed. For instance the following operation in illegal:

from .utils import some_util_function
from src.utils import some_util_function