Code for "The Intrinsic Dimension of Images and Its Impact on Learning" - ICLR 2021 Spotlight

Last update: Dec 10, 2022

Related tags

Overview

`dimensions`

Estimating the instrinsic dimensionality of image datasets

Code for: The Intrinsic Dimensionaity of Images and Its Impact On Learning - Phillip Pope and Chen Zhu, Ahmed Abdelkader, Micah Goldblum, Tom Goldstein (ICLR 2021, spotlight)

https://openreview.net/forum?id=XJk19XzGq2J

Environment

This code was developed in the following environment

conda create dimensions python=3.6 jupyter matplotlib scikit-learn pytorch==1.5.0 torchvision cudatoolkit=10.2 -c pytorch

To generate new data of controlled dimensionality with GANs, you must install:

pip install pytorch-pretrained-biggan

To use the shortest-path method (Granata and Carnevale 2016) you must also compile the fast graph shortest path code gsp (written by Jake VdP + Sci-Kit Learn)

cd estimators/gsp
python setup.py install

Generate data of controlled dimensionality

python generate_data/gen_images.py \
  --num_samples 1000 \
  --class_name basenji \
  --latent_dim 16 \
  --batch_size 100 \
  --save_dir samples/basenji_16

Estimate dimension of generated samples

To run the MLE (Levina and Bickel) estimator on the synthetic GAN data generated above:

python main.py \
    --estimator mle \
    --k1 25 \
    --single-k \
    --eval-every-k \
    --average-inverse \
    --dset  samples/basenji_16 \
    --max_num_samples 1000 \
    --save-path results/basenji_16.json

Use --estimators to try different estimators

Citation

If you find our paper or code useful, please cite our paper:

@inproceedings{DBLP:conf/iclr/PopeZAGG21,
  author    = {Phillip Pope and
               Chen Zhu and
               Ahmed Abdelkader and
               Micah Goldblum and
               Tom Goldstein},
  title     = {The Intrinsic Dimension of Images and Its Impact on Learning},
  booktitle = {9th International Conference on Learning Representations, {ICLR} 2021,
               Virtual Event, Austria, May 3-7, 2021},
  publisher = {OpenReview.net},
  year      = {2021},
  url       = {https://openreview.net/forum?id=XJk19XzGq2J},
  timestamp = {Wed, 23 Jun 2021 17:36:39 +0200},
  biburl    = {https://dblp.org/rec/conf/iclr/PopeZAGG21.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgements

We gratefully acknowledge use of the following codebases when developing our dimensionality estimators:

We also thank Prof. Vishnu Boddeti for clarifying comments on the graph-distance estimator.

Disclaimer

This code released as is. We will do our best to address questions/bugs, but cannot guarantee support.

Code for "The Intrinsic Dimension of Images and Its Impact on Learning" - ICLR 2021 Spotlight

Related tags

Overview

`dimensions`

Environment

Generate data of controlled dimensionality

Estimate dimension of generated samples

Citation

Acknowledgements

Disclaimer

Owner

Phil Pope

Code + pre-trained models for the paper Keeping Your Eye on the Ball Trajectory Attention in Video Transformers

PyTorch implementation of "Simple and Deep Graph Convolutional Networks"

Portfolio analytics for quants, written in Python

Implementation of DocFormer: End-to-End Transformer for Document Understanding, a multi-modal transformer based architecture for the task of Visual Document Understanding (VDU)

CLASP - Contrastive Language-Aminoacid Sequence Pretraining

An abstraction layer for mathematical optimization solvers.

DeepSpamReview: Detection of Fake Reviews on Online Review Platforms using Deep Learning Architectures. Summer Internship project at CoreView Systems.

Image Restoration Using Swin Transformer for VapourSynth

A PyTorch implementation for Unsupervised Domain Adaptation by Backpropagation(DANN), support Office-31 and Office-Home dataset

Modifications of the official PyTorch implementation of StyleGAN3. Let's easily generate images and videos with StyleGAN2/2-ADA/3!

A fast implementation of bss_eval metrics for blind source separation

UniMoCo: Unsupervised, Semi-Supervised and Full-Supervised Visual Representation Learning

This repository contains the code and models necessary to replicate the results of paper: How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective

A 10000+ hours dataset for Chinese speech recognition

ShapeGlot: Learning Language for Shape Differentiation

ECCV2020 paper: Fashion Captioning: Towards Generating Accurate Descriptions with Semantic Rewards. Code and Data.

code associated with ACL 2021 DExperts paper

Full Resolution Residual Networks for Semantic Image Segmentation

Official PyTorch Implementation of Rank & Sort Loss [ICCV2021]

Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks