backbone parameter freezing across models

[AhmadOudeh]

Description

There seems to be inconsistency in how different models handle backbone parameter freezing. Some models (like SwinTransformer) have a clear mechanism through frozen_stages parameter and _freeze_stages() method, while others (like FastSCNN) lack this functionality. This makes it difficult to have a consistent approach to transfer learning across different architectures.

Current State

1. Some models implement freezing through:

   • frozen_stages config parameter
   • _freeze_stages() method
   • Built-in parameter freezing logic

2. Other models (like FastSCNN) don't have these mechanisms, making it unclear how to:

   • Freeze specific layers
   • Control freezing through configs
   • Implement transfer learning consistently

Question

1. What's the recommended way to freeze backbone parameters in FastSCNN?
2. Should I manually set requires_grad=False for parameters, or is there a config-based solution?

Example

Using FastSCNN as an example:

# In SwinTransformer, we can do:
model = dict(
    backbone=dict(
        type='SwinTransformer',
        frozen_stages=2,
        ...
    )
)

# But in FastSCNN, there's no equivalent:
model = dict(
    backbone=dict(
        type='FastSCNN',
        # No frozen_stages parameter
        ...
    )
)