This repository contains the code for the paper "Fast, Distribution-free Predictive Inference for Neural Networks with Coverage Guarantees". The paper introduces a novel, computationally-efficient algorithm for predictive inference (PI) that requires no distributional assumptions on the data and can be computed faster than existing bootstrap-type methods for neural networks.
pip3 install -r requirements.txt
python3 main.py --helpusage: main.py [-h] [--data_name DATA_NAME] [--epoch EPOCH] [--tot_trial TOT_TRIAL] [--alpha ALPHA] [--train_size TRAIN_SIZE]
[--hidden_width HIDDEN_WIDTH] [--lam LAM] [--iseed ISEED] [--psim PSIM]
options:
-h, --help show this help message and exit
--data_name DATA_NAME, -dn DATA_NAME
--epoch EPOCH, -e EPOCH
number of epochs when training the neural networks
--tot_trial TOT_TRIAL, -t TOT_TRIAL
total number of trials
--alpha ALPHA, -a ALPHA
level of significance
--train_size TRAIN_SIZE, -n TRAIN_SIZE
training size
--hidden_width HIDDEN_WIDTH, -w HIDDEN_WIDTH
--lam LAM, -l LAM ridge penalty parameter
--iseed ISEED, -s ISEED
data generating random seed
--psim PSIM, -p PSIM simulation p
dimension of data features in the simulation
python3 main.py -e 20 -n 100 -p 16 -dn sim
or
module load python/python3.9.9 #if python3 is not available
python main.py -e 20 -n 100 -p 16 -dn sim
The code is released under the MIT License.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent