XPipe 0.1.3

Standardize your ML projects

Introduction

XPipe is a library that I started developping in December 2020 for my personal use. As it might be useful for other people, I decided to publish the code as an open source project.

XPipe focuses on two principal components to make Data Science easier:

• Configuration files are a big concern in data science field and there is no standard today. XPipe facilitates your work by automatically loading python objects from a yaml configuration. You can also easily include other yaml files into another.

• Experiment tracking: The web interface enables you to easily organize your experiments into folders, to filter them and to plot different kind of graphs. You will particularly appreciate the library if you deal with a lot of experiments.

The philosophy behind the project is to be simple and customizable.

As a team, you can run a single XPipe server for everyone. It will promote exchange as everyone can easily share their work with others.

Getting started

pip install xpipe

Documentation (work in progress): https://x-pipe.readthedocs.io/en/latest/#

Configuration files

Here is a simple example of how to use yaml configuration files to seamlessly load needed objects to run your experiments.

training:
  gpu: !env CUDA_VISIBLE_DEVICES # Get the value of env variable CUDA_VISIBLE_DEVICES
  epochs: 18
  batch_size: 100

  optimizer: !obj torch.optimSGD : {lr : 0.001}

  scheduler: !obj torch.optim.lr_scheduler.MultiStepLR : {milestones: [2, 6, 10, 14]}

  loss: !obj torch.nnBCELoss : {}

model: !include "./models/my_model.yaml"

transforms:
  !obj transforms.Normalize : {}
  !obj transforms.Noise : {}
  !obj transforms.RandomFlip : {probability: 0.5}

Then you can load the configuration file:

from xpipe.config import load_config

conf = load_config("my_config.yaml")
epochs = conf.training.epochs() # 18

# Instantiate your model defined in models/my_model.yaml
my_model = params.model()

# Directly instantiate your optimizer and scheduler from configuration
# Note that you can add argument that are not in the configuration file
optimizer = params.training.optimizer(params=my_model.parameters())
scheduler = params.training.scheduler(optimizer=optimizer)

Experiment tracking

This feature is still experimental.

You have two options to start the server:

1. Run the server from the commandline. You must host a MongoDB server instance.

xpipe --db_host <db_ip_address> --db_port <db_port> --port <server_port> --artifacts-dir <artifacts_dir>

2. Run directly the docker image (no other dependancies needed)

docker pull drosos/xpipe:0.1.2
docker run -v <data_dir>:/data -p <server_port>:80 drosos/xpipe:0.1.2

The <data_dir> directory will contain the mongodb database and artifacts.

Then you can connect to http://127.0.0.1:<server_port> to access the web interface.

If you open an experiment, you can get some details and results: