Standardisation

Standardisation#

A driving principle of the molflux package is standardisation across all of its five submodules: datasets, features, splits, modelzoo, metrics. In this section, we demonstrate and describe the unified API of the package. Learning to use the following functionality gives the user immediate knowledge of how to use all parts of the package. Each of the following methods can be imported from the relevant submodule (for example from molflux.datasets import list_datasets).

Browsing#

To start, we first introduce the basic browsing functionality of the submodules. Each submodule has a list_* function that returns a dictionary of available objects (datasets, representations, models, etc…). These are

  1. list_datasets

  2. list_representations

  3. list_splits

  4. list_models

  5. list_metrics

The dictionaries returned are grouped by the optional dependency required for the objects (key) and the list of available objects (value). For more information, see the respective browsing sections in the
documentation: datasets, features, splits, modelzoo, metrics.

Loading#

load_*#

The load_* is the most straightforward loading method of the API. It can be used to load the objects of any of the five main submodules

  1. load_datasets

  2. load_representation

  3. load_splitting_strategy

  4. load_model

  5. load_metric

The pattern is load_*(name: str, **kwargs) where name is a string name of the object and kwargs are optional object specific kwargs.

load_from_dict#

Although the load_* function can load the required objects, specifying kwargs directly can become tedious. In general, we recommend using a config driven approach to load objects. This is done via the load_from_dict method which expects a config dictionary in the following format

{
  "name": <string name of object>,
  "config": <dictionary of kwargs used at instantiation>,
  "presets": <dictionary of kwargs used at the relevant method call>,
}

load_from_dicts#

It is not uncommon to want to load multiple objects of the same flavour at the same type (for example loading multiple datasets or multiple representations). For this, we have the load_from_dicts (plural) which takes in a list of config dictionaries as specified in load_from_dict.

load_from_yaml#

A complete machine learning pipeline consists of multiple stages each of which requires a config. To streamline the process of specifying configs for pipelines, we provide the option to use a unified yaml file to store the configs of all the stages and the load_from_yaml for each submodule which knows how to pick out the relevant parts of the config for its purposes.

The general form of the yaml config looks like

---
version: v1
kind: datasets
specs:
  - name:
    config:
---
version: v1
kind: representations
specs:
  - name:
    config:
---
version: v1
kind: splits
specs:
  - name:
    config:
---
version: v1
kind: models
specs:
  - name:
    config:
---
version: v1
kind: metrics
specs:
  - name:
    config:

where each submodule can take a list of configs for each object to load.

See also

For an explicit example of using a single yaml file, see the ESOL regression example using yaml.

For more information, see the respective loading sections in the
documentation: datasets, features, splits, modelzoo, metrics.

Contents