Conv Generator

Abstract

This repository is a pytorch implementation of "Defects of Convolutional Decoder Networks in Frequency Representation", which has been published at ICML2023.

Requirements

1. Make sure GPU is avaible and CUDA>=11.0 has been installed on your computer. You can check it with

       nvidia-smi

2. Simply create an virtural environment with python>=3.8 and run pip install -r requirements.txt to download the required packages. If you use anaconda3 or miniconda, you can run following instructions to download the required packages in python.

       conda create -y -n Conv python=3.8
       conda activate Conv
       pip install pip --upgrade
       pip install -r requirements.txt
       conda activate Conv
       conda install pytorch=1.10.2 torchvision=0.11.3 torchaudio=0.10.2 cudatoolkit=11.1 -c pytorch -c nvidia

Document Structure

Src

--coef.py [the coeffients for calculation]

--config.py [the configuration]

--dat.py [load the data]

--hook.py [the hook on the model to register data in the propagation process]

--models.py [the models used for the exps]

--train.py [train the models]

--utils.py [useful tools]

Scripts

--bottleneck1.py [train the model to verify the bottleneck 1]

--bottleneck2.py [train the model to verify the bottleneck 2]

--corollary1.py [verify the corollary 1]

--corollary2.py [verify the corollary 2]

--plot_bottleneck1.py [visualize the bottleneck 1]

--plot_bottleneck2.py [visualize the bottleneck 2]

--plot_corollary.py [visualize the corollaries]

--remark1.py [verify the remark 1]

--remark2.py [verify the remark 2]

--remark3.py [verify the remark 3]

--remark4.py [verify the remark 4]

--remark5.py [verify the remark 5]

--theorem5.py [verify the theorem 5]

--theorem6.py [verify the theorem 6]

Contact

Please contact [tling@sjtu.edu.cn] if you have any question on the codes.

---

Shanghai Jiao Tong University - Email@[tling@sjtu.edu.cn]