summary

implement a dynamic model selection framework to seamlessly load different model architectures without modifying core code.

background

the current model loading mechanism in run.py uses if-statements to manually check the model name and load the appropriate class. this approach is not scalable and requires modifying core code whenever a new model is added.

proposed solution

create a modular framework that dynamically discovers and loads models from the codebase:

1. register models using a decorator pattern
2. implement auto-discovery of models in the models directory
3. create a factory class for model instantiation
4. support custom models from external paths

code example

current approach:

if args.model == 'TimeMixer':
 from models.TimeMixer import Model
elif args.model == 'OtherModel':
 from models.OtherModel import Model
# etc.

proposed approach:

from utils.model_registry import get_model
model_class = get_model(args.model)
model = model_class(args)

benefits

• extensibility: new models can be added without modifying existing code
• maintainability: reduces code duplication and simplifies the codebase
• discoverability: makes it easier to understand available models
• research-friendly: simplifies adding experimental model variants

implementation details

• create utils/model_registry.py with registration mechanism
• implement model discovery functionality
• modify experiment classes to use the registry
• add documentation about extending with custom models