PEFT in SAM

1. Theory

1.1. Addition-based method

1.1.1. Adapter tuning

Adapter former

• Conclusion
  • Parallel insertion is more efficacious than sequential insertion

Convpass

• Conclusion
  • Highlights that current adapters are hindered by a lack of strong inductive bias, limiting their performance
• Method
  • Incoporate conv blocks

AIM

ST Adapter

Rob Adapter

1.1.2. Prompt tuning

VPT

DePT

CVP

LPT

Pro tuning

LION

ViPT

VP, EVP, DAM-VP, ProSFDA, P2P, ILM-VP,

1.1.3. Prefix tuning

PATT

eTT

LAM

VQT

1.1.4. Side tuning

Side tuning

SAN

ViT Adapter

LST

DTL

1.2. Partial based tuning

1.2.1. Specification tuning

Linear Probe

BitFit

DP-BitFit

DiffFit

AdapterBias

LN Tune

1.2.2. Reparameter tuning

LoRA

KronA

KAdaptation

FaCT

EFFT

SSF

RepAdapter

1.3. Unified based tuning

NOAH

LAM

V-PEFT

U-tuning

2. Implementation

2.1. Description

2.2. Dataset

2.3. Modeling

2.4. Benchmark

2.5. Inference

2.6. Training

TODO

• Develop SAM variants
  • MedSA
    • LoRA
    • Adapter
    • Adapter + LoRA
  • SAM LST
  • SAM Adapter
  • SAMed
  • MASAM
  • SAM IHS
  • SAM US
• Develop PEFT variants
  • LoRA
  • Adapter
  • LST
  • FacT
• Promptings
  • Points
  • Boxes
  • Masks
• Benchmarking
  • SAM
  • MedSAM