Encoder Compatibility Issue Across Models Using mit_b0

[shahidul034]

Encoder Compatibility Issue Across Models Using mit_b0

Description

While using mit_b0 as an encoder, I encountered compatibility issues with several models, not just the previously mentioned Linknet and Unet++. According to the documentation, Linknet and Unet++ do not support mit_b0 as an encoder, and FPN supports it only with an encoder depth of 5. However, in practice, other models also fail to support mit_b0 as an encoder.

Models and Observed Behavior

Below is a list of models and the observed training status using mit_b0 as an encoder:

Trained Successfully

• Unet (both with imagenet and without a pretrained encoder)
• FPN (both with imagenet and without a pretrained encoder)
• PSPNet (both with imagenet and without a pretrained encoder)
• MAnet (both with imagenet and without a pretrained encoder)

Encountered Exceptions (Not Trained)

• UnetPlusPlus (both with imagenet and without a pretrained encoder)
• Linknet (both with imagenet and without a pretrained encoder)
• DeepLabV3 (both with imagenet and without a pretrained encoder)
• DeepLabV3Plus (both with imagenet and without a pretrained encoder)
• PAN (both with imagenet and without a pretrained encoder)

Expected Behavior

The documentation should either:

1. Clarify that mit_b0 is not compatible with certain models, including but not limited to Linknet and Unet++.
2. Indicate that these models are not supported with the current encoder configuration.

This clarification would help set clear expectations for users attempting to use mit_b0 across various models.

Environment

• OS: Ubuntu 24.04.1 LTS
• Framework: PyTorch
• Python version: 3.11

[brianhou0208]

Check the PR here #975