Trusting the Untrustworthy: A Cautionary Tale on the Pitfalls of Training-based Rejection Option

Fig 1. Global coverage (left) refers to the proportion of the entire dataset that is covered by a model's predictions. Low variance in global coverage means that the method consistently covers the target portion of the dataset across different independent runs. However, this consistency in overall coverage may result in a large variance in the coverage for each individual class (right). In other words, the methods may perform well in covering some classes while not performing as well in covering others.

Submitted to ICML 2023.

Requirements

To reproduce the results, you will need to install the packages listed in requirements.txt and have python>=3.7 installed.

Please run pip install -r requirements.txt to install them.

Usage

This section will explain how to reproduce results.

1. Training the models from scratch

Simply run the code:

python -m src.${method}.train \
    --model_name ${model}_${dataset} \
    --seed ${seed}

Where the variables you can choose from:

• method:

  • GiniSelector, which will train the model with the cross-entropy (CE) loss.
  • SelectiveNet.
  • ConfidNet.
  • DeepGamblers.
  • SelfAdaptiveTraining.

• model:

  • vgg16
  • densenet121
  • resnet34

• dataset:

  • cifar-10
  • cifar-100
  • svhn

• seed: an integer

The training will save the best weights as well as logs, logits, and scores in the checkpoints/${method}/${model}_${dataset}/${seed} folder.

Or you can download the checkpoints

Abiding by the principles of open science, we provide the checkpoints for most methods and models on the release page. PS.: The checkpoints for SelectiveNet are not available due to the size of the files (~30GB in total).

Run the following code to download and unzip a checkpoint from the release:

mkdir checkpoints/
cd checkpoints/
wget -c https://github.com/giniselector/giniselector/releases/download/checkpoints/${method}-${model}_${dataset}.zip
unzip ./${method}-${model}_${dataset}.zip

Or go to the release page and download the checkpoints manually.

Then you can save the logits and scores from a specific method and model by running:

python -m src.${method}.test \
    --model_name ${model}_${dataset} \
    --seed ${seed}

2. Evaluating

To evaluate the benchmark, please run:

python -m src.analysis.save_results \
    --model_name ${model}_${dataset}

The results will be exported to .csv in the results/ repository, and the evaluation metrics are defined in src/utils/eval.py

3. Plotting

Run the code:

python -m src.analysis.plots \
    --model_name ${model}_${dataset}

The images will be saved to the images/ folder.

Our Method

The post-hoc Gini score can be plugged into any deep classifier trained with the cross-entropy loss. The function implemented in python that takes the logits as input is defined below.

import torch


def gini(logits: torch.Tensor):
    g = torch.sum(torch.softmax(logits, 1) ** 2, 1)
    return 1 - g

Results

(Optional) Environmental Variables

Please, place the following lines in your .env file if you want to modify any of the default folders.

#.env
export DATA_DIR=""
export IMAGENET_ROOT=""
export CHECKPOINTS_DIR="checkpoints/"

Where:

• DATA_DIR is the directory where the datasets will be downloaded.
• IMAGENET_ROOT is the directory where the ImageNet dataset is located.
• CHECKPOINTS_DIR is the directory where the pre-trained models will be placed.