SegFormerModelSize

Enums · AiDotNet.Enums

Defines the model size variants for SegFormer (Mix Transformer backbone).

For Beginners

SegFormer comes in six sizes (B0 through B5). Smaller sizes (B0) are faster and use less memory, while larger sizes (B5) are more accurate but require more compute. B0 is a great starting point for experimentation, while B2-B3 offer a good balance for production use.

How It Works

Technical Details: Each size uses a different Mix Transformer (MiT) backbone with varying embedding dimensions, transformer depths, and attention heads.

Reference: Xie et al., “SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers”, NeurIPS 2021.

Fields

Field Summary
SegFormerModelSize B0 MiT-B0: Smallest variant (3.8M params).
SegFormerModelSize B1 MiT-B1: Small variant (13.7M params).
SegFormerModelSize B2 MiT-B2: Medium variant (25.4M params).
SegFormerModelSize B3 MiT-B3: Large variant (45.2M params).
SegFormerModelSize B4 MiT-B4: Extra-large variant (62.0M params).
SegFormerModelSize B5 MiT-B5: Largest variant (82.0M params).
int value__