RRClarC -> CSR

[joshua840]

• [x] #5
• [x] CAVs -> CSR
  • [x] Implementation
• [x] Apply RRClarC into catdog dataset
  • [x] Implementation

Running experiments

• [x] Create raw data
• [x] Run feature extractor
• [x] Train model by ERM
• [ ] GDRO
  • Runnable without ERM results
• [ ] JTT
  1. 4.1. run JTT weight generator
     1. The process itself is ERM, but cannot leverage the model in 2.
  2. 4.2. run JTT main training code with a re-weighted dataset
• [ ] DFR
  1. Train the last layer by using a balanced validation dataset
• [ ] CAVs
  • [x] run feature extractor with a concept dataset
  • [x] calculate standard CAV
    • We now have linear cavs in data/cavs
  • [x] calculate CAV with concept gradient
    • We now have TEM-578 model that is trained on three layer model_g
  • [ ] run RRClarC or CGR

[joshua840]

Signal CAV

• How to prepare W that is evaluated from CAV?
  • Option 1: create a new main file and train it without considering
  • Option 2: create a new method that can be operated in Trainer.

The option 2 seems better.

Also, consider eliminating jtt_weight_generator.py and deals fit and test functions in command line

[joshua840]

Overall process

• [x] Create raw data
• [x] Run feature extractor
• [x] Train model by ERM
• [ ] GDRO
  • Runnable without ERM results
• [ ] JTT
  1. 4.1. run JTT weight generator
     1. The process itself is ERM, but cannot leverage the model in 2.
  2. 4.2. run JTT main training code with a re-weighted dataset
• [ ] DFR
  1. Train the last layer by using a balanced validation dataset
• [ ] CAVs
  • [x] run feature extractor with a concept dataset
  • [x] calculate standard CAV
    • We now have linear cavs in data/cavs
  • [x] calculate CAV with concept gradient
    • We now have TEM-578 model that is trained on three layer model_g
  • [x] run RRClarC or CGR
    • [x] load CAVs as a weight of model_g

[joshua840]

I will close this issue and create a new issue to deals with the experiment.