Code accompanying the paper "How Tight Can PAC-Bayes be in the Small Data Regime?"

Last update: Dec 21, 2021

Overview

How Tight Can PAC-Bayes be in the Small Data Regime?

This is the code to reproduce all experiments for the following paper:

@inproceedings{Foong:2021:How_Tight_Can_PAC-Bayes_Be,
    title = {How Tight Can {PAC}-{Bayes} Be in the Small Data Regime?},
    year = {2021},
    author = {Andrew Y. K. Foong and Wessel P. Bruinsma and David R. Burt and Richard E. Turner},
    booktitle = {Advances in Neural Information Processing Systems},
    volume = {35},
    eprint = {https://arxiv.org/abs/2106.03542},
}

Every experiment creates a folder in _experiments. The names of the files in those folders should be self-explanatory.

Installation

First, create and activate a virtual environment for Python 3.8.

virtualenv venv -p python3.8 
source venv/bin/activate

Then install an appropriate GPU-accelerated version of PyTorch.

Finally, install the requirements for the project.

pip install -e .

You should now be able to run the below commands.

Generating Datasets

In order to generate the synthetic 1D datasets used, run these commands from inside classification_1d:

python gen_data.py --class_scheme balanced --num_context 30 --name 30-context --num_train_batches 5000 --num_test_batches 64
python gen_data.py --class_scheme balanced --num_context 60 --name 60-context --num_train_batches 5000 --num_test_batches 64

The generated datasets will be in pacbayes/_data_caches

Theory Experiments

See Figure 2 in Section 3 and Appendix G.

python theory_experiments.py --setting det1-1
python theory_experiments.py --setting det1-2
python theory_experiments.py --setting det2-1
python theory_experiments.py --setting det2-1

python theory_experiments.py --setting stoch1
python theory_experiments.py --setting stoch2
python theory_experiments.py --setting stoch3

python theory_experiments.py --setting random --random-seed 1 --random-better-bound maurer
python theory_experiments.py --setting random --random-seed 6 --random-better-bound catoni

GNP Classification Experiments

See Figure 3 and 4 in Section 4 and Appendices I and J. The numbers from the graphs can be found in eval_metrics_no_post_opt.txt (without post optimisation) eval_metrics_post_opt.txt (with post optimisation).

MODEL_NONDDP=maurer MODEL_DDP=maurer-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer MODEL_DDP=maurer-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=convex-nonseparable MODEL_DDP=convex-nonseparable-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=convex-nonseparable MODEL_DDP=convex-nonseparable-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-optimistic MODEL_DDP=maurer-optimistic-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-optimistic MODEL_DDP=maurer-optimistic-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-inv MODEL_DDP=maurer-inv-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-inv MODEL_DDP=maurer-inv-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-inv-optimistic MODEL_DDP=maurer-inv-optimistic-ddp NUM_CONTEXT=30 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-inv-optimistic MODEL_DDP=maurer-inv-optimistic-ddp NUM_CONTEXT=30 ./run_GNP_prop_68.sh

MODEL_NONDDP=maurer MODEL_DDP=maurer-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer MODEL_DDP=maurer-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=convex-nonseparable MODEL_DDP=convex-nonseparable-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=convex-nonseparable MODEL_DDP=convex-nonseparable-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-optimistic MODEL_DDP=maurer-optimistic-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-optimistic MODEL_DDP=maurer-optimistic-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-inv MODEL_DDP=maurer-inv-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-inv MODEL_DDP=maurer-inv-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh
MODEL_NONDDP=maurer-inv-optimistic MODEL_DDP=maurer-inv-optimistic-ddp NUM_CONTEXT=60 ./run_GNP_prop_024.sh
MODEL_NONDDP=maurer-inv-optimistic MODEL_DDP=maurer-inv-optimistic-ddp NUM_CONTEXT=60 ./run_GNP_prop_68.sh

MLP Classification Experiments

See Appendix J. The numbers from the graphs can be found in eval_metrics_no_post_opt.txt (without post optimisation) eval_metrics_post_opt.txt (with post optimisation).

MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=30 ./run_MLP.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=30 ./run_MLP.sh

MODEL_NONDDP=catoni MODEL_DDP=catoni-ddp NUM_CONTEXT=60 ./run_MLP.sh
MODEL_NONDDP=kl-val MODEL_DDP=kl-val NUM_CONTEXT=60 ./run_MLP.sh

Code accompanying the paper "How Tight Can PAC-Bayes be in the Small Data Regime?"

Related tags

Overview

How Tight Can PAC-Bayes be in the Small Data Regime?

Installation

Generating Datasets

Theory Experiments

GNP Classification Experiments

MLP Classification Experiments

Owner

An open source python library for automated feature engineering

Code for 'Self-Guided and Cross-Guided Learning for Few-shot segmentation. (CVPR' 2021)'

Melanoma Skin Cancer Detection using Convolutional Neural Networks and Transfer Learning🕵🏻‍♂️

A Number Recognition algorithm

A tool for calculating distortion parameters in coordination complexes.

Official PyTorch implementation of Segmenter: Transformer for Semantic Segmentation

Efficient training of deep recommenders on cloud.

Benchmark for Answering Existential First Order Queries with Single Free Variable

Official PyTorch implementation of our AAAI22 paper: TransMEF: A Transformer-Based Multi-Exposure Image Fusion Framework via Self-Supervised Multi-Task Learning. Code will be available soon.

Contrastive Multi-View Representation Learning on Graphs

Awesome Artificial Intelligence, Machine Learning and Deep Learning as we learn it

Official implementation of "SinIR: Efficient General Image Manipulation with Single Image Reconstruction" (ICML 2021)

Github for the conference paper GLOD-Gaussian Likelihood OOD detector

A PyTorch implementation of "Semi-Supervised Graph Classification: A Hierarchical Graph Perspective" (WWW 2019)

TrackTech: Real-time tracking of subjects and objects on multiple cameras

Tracing Versus Freehand for Evaluating Computer-Generated Drawings (SIGGRAPH 2021)

Accurate Phylogenetic Inference with Symmetry-Preserving Neural Networks

a reimplementation of LiteFlowNet in PyTorch that matches the official Caffe version

DARTS-: Robustly Stepping out of Performance Collapse Without Indicators

High level network definitions with pre-trained weights in TensorFlow