recap 0.1.3

Reproducible configurations for any project

recap

recap is a tool for providing REproducible Configurations for Any Project.

Research should be reproducible. Especially in deep learning, it is important to keep track of hyperparameters and configurations used in experiments. This package aims at making that easier.

Installing

Just install like any Python package:

pip install recap

Overview

Recap provides two top-level concepts that would be imported as follows:

from recap import URI, CfgNode as CN

The CfgNode is a subclass of yacs' CfgNode. It provides some additional features for parsing configurations that are inherited between files which is not possible with yacs.

Recap's URI class provides a mechanism for handling logical paths within your project more conveniently with an interface that is fully compatible with pathlib.Path.

YAML configurations

Configurations are defined just like in yacs, except that you need to import the CfgNode class from the recap package instead of yacs. Consider the following YAML configuration that sets default values for all configuration options we will use in our project. We shall name it _base.yaml because our experiments will build on these values.

SYSTEM:
  NUM_GPUS: 4
  NUM_WORKERS: 2
TRAIN:
  LEARNING_RATE: 0.001
  BATCH_SIZE: 32
  SOME_OTHER_HYPERPARAMETER: 10

The equivalent configuration can be obtained programatically like so:

from recap import CfgNode as CN

cfg = CN()
cfg.SYSTEM = CN()
cfg.SYSTEM.NUM_GPUS = 4
cfg.SYSTEM.NUM_WORKERS = 2
cfg.TRAIN = CN()
cfg.TRAIN.LEARNING_RATE = 1e-3
cfg.TRAIN.BATCH_SIZE = 32
cfg.TRAIN.SOME_OTHER_HYPERPARAMETER = 10

print(cfg)

Inheriting configurations

Recap provides functionality for inheriting configuration options from other configuration files by setting the top-level _BASE_ key. So, we could create a configuration file experiment_1.yaml for an experiment where we try a different learning rate and batch size:

_BASE_: _base.yaml

TRAIN:
  LEARNING_RATE: 1e-2
  BATCH_SIZE: 64

In our code, when we want to load the experiment configuration, we would use the recap.CfgNode.load_yaml_with_base() function:

from recap import CfgNode as CN

cfg = CN.load_yaml_with_base("experiment_1.yaml")

print(cfg)

# Will output:
"""
SYSTEM:
  NUM_GPUS: 4
  NUM_WORKERS: 2
TRAIN:
  LEARNING_RATE: 0.01
  BATCH_SIZE: 64
  SOME_OTHER_HYPERPARAMETER: 10
"""

Note that the _BASE_ keys can be arbitrarily nested; however, circular references are prohibited.

Logical URIs and the path manager

Recap includes a path manager for conveniently specifying paths to logical entities. The path strings are set up like a URI where the scheme (i.e. http in the path string http://google.com) refers to a logical entity. Each such entity needs to be set up as a PathTranslator that can translate the logical URI path to a physical path on the file system.

For example, we could set up a path translator for the data scheme to refer to the the path of a dataset on our file system located at /path/to/dataset. Then the recap URI data://train/abc.txt would be translated to /path/to/dataset/train/abc.txt.

The simplest way of setting that up is using the register_translator function (although more complex setups are possible with the recap.path_manager.PathTranslator class, allowing you to download files from the internet, for example):

from recap.path_manager import register_translator
from pathlib import Path

register_translator("data", Path("/path/to/dataset"))

Then, we can use the recap.URI class just like any pathlib.Path object:

from recap import URI

my_uri = URI("data://train/abc.txt")
# Here, str(my_uri) == "/path/to/dataset/train/abc.txt"

with my_uri.open("r") as f:
    print(f.read())

Logical URIs in inherited configurations

The recap.URI interface is fully compatible with the nested configurations. This means that you can use recap URIs within the _BASE_ field for inheriting configurations.

For example, you could register a path translator for the config scheme and then include _BASE_: config://_base.yaml in your configuration files.