Skip to content

config_design.md

Latest commit

 

History

History
116 lines (99 loc) · 5.56 KB

config_design.md

File metadata and controls

116 lines (99 loc) · 5.56 KB

📜 Configuration Design

The design philosophy of BasicTS is "All-in-configuration".
Our goal is to let users focus on models and data, without being burdened by complex training pipelines.

🎸 New Features

Starting from version 1.0, BasicTS no longer uses .py files combined with command-line configuration paths.
Instead, it has been upgraded to a configuration class–based system.

The base configuration class is BasicTSConfig.
In addition, each specific task has its own configuration class, including:

  • BasicTSForecastingConfig
  • BasicTSClassificationConfig
  • BasicTSImputationConfig
  • BasicTSFoundationModelConfig

The base class BasicTSConfig defines shared fields and methods for saving, loading, and printing configurations,
while task-specific configuration classes contain all parameters required for executing that task.
Users can flexibly import models and set all necessary options within these configuration classes.

A configuration class typically includes the following sections:

  • General Options: Basic settings such as configuration description, gpus, and seed.
  • Environment Options: Settings such as tf32, cudnn, and deterministic.
  • Dataset Options: Specify dataset_name (dataset name, must be explicitly defined),
    dataset_type (dataset class), and dataset_params (dataset parameters).
  • Scaler Options: Specify scaler (scaler class), norm_each_channel (per-channel normalization),
    and rescale (whether to inverse-normalize outputs).
  • Model Options: Specify model (model class, must be explicitly defined) and
    model_config (model parameters, must be explicitly defined).
  • Evaluation Metrics Options: Define metrics (evaluation functions) and target_metric (primary metric).
  • Training Options:
    • General: Define settings such as num_epochs / num_steps, and loss.
    • Optimizer: Specify optimizer (optimizer class) and optimizer_params (optimizer parameters).
    • Scheduler: Specify lr_scheduler (scheduler class) and lr_scheduler_params (scheduler parameters).
    • Data: Define batch_size, num_workers, and pin_memory.
  • Validation Options:
    • General: Validation frequency val_interval.
    • Data: Define batch_size, num_workers, and pin_memory.
  • Testing Options:
    • General: Testing frequency test_interval.
    • Data: Define batch_size, num_workers, and pin_memory.

Each field in the Config class contains metadata that provides default values, detailed explanations, and usage instructions.

🏗️ Building Configuration Classes

👥 Both a Class and a Dictionary

The configuration classes in BasicTS inherit from EasyDict,
so they can be used both as Python classes and dictionaries,
making them highly flexible and extensible.

There are two main ways to use them:

  1. As a class — all parameters are accessible as class attributes.
    You can access, modify, or add new fields just like regular class variables:

    model = config.model
config.gpus = "0"
config.new_field = new_value

  2. As a dictionary — all parameters are accessible as key-value pairs.
    You can access, modify, add, or delete parameters just like with a dictionary:

    model = config["model"]
optimizer = config.get("optimizer", Adam)
config["new_key"] = new_value
config.pop("not_used")

🔨 Objects Inside Configuration Classes

You may have noticed that many configuration fields represent objects that need to be constructed later,
such as models, datasets, optimizers, and schedulers.
In BasicTS, you specify the corresponding class and initialization parameters in the configuration,
and the actual object construction is deferred until runtime.

There are two flexible ways to define these objects:

  1. Pass all initialization parameters as a dictionary.
    For example, use dataset_params to specify parameters for dataset creation:

    config = BasicTSForecastingConfig(
    dataset_params={
        'input_len': 336,
        'output_len': 336,
        ...
    },
    ...
)

  2. Pass parameters directly as configuration fields.
    For example, you can directly set input_len and output_len,
    and add custom fields later (note: only predefined fields can be passed during initialization):

    config = BasicTSForecastingConfig(
    input_len=336,
    output_len=336,
    ...
)
config.your_dataset_param_1 = param_1
config.your_dataset_param_2 = param_2

🧑‍💻 Explore Further