Skip to content

Code for Saliency is a Possible Red Herring When Diagnosing Poor Generalization, Viviano et al ICLR 2021

Notifications You must be signed in to change notification settings

josephdviviano/saliency-red-herring

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

saliency-red-herring

NB: I NEED TO UPDATE THIS README

installation:

conda env create -n activmask --file environment.yml

source install.sh.

training:

A config file needs to be defined to run the experiments, e.g.:

gradmask train --config gradmask/config/mnist.yml

monitoring:

The code right now will log all experiments in the logs/experiments.csv file. The time of saving, git hash, config, and best accuracy on the valid set is saved.

skopt

Steps to launch bayesian hyperparameters search:

  1. In your .yml config file, choose the parameters you want to optimize (i.e. learning rate).
  2. Replace the value with the search parameters. For example:
    # Optimizer
    optimizer:
      Adam:
        lr: "Real(10**-4, 10**-2, 'log-uniform')"
    
    search the learning rate in the range (0.01, 0.0001), on a log scale. Examples..
  3. Launch your config file with activmask train-skopt --config config/path.yml

An config example can be found in config/mnist_skopt.yml

About

Code for Saliency is a Possible Red Herring When Diagnosing Poor Generalization, Viviano et al ICLR 2021

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •