Skip to content

Latest commit

 

History

History
38 lines (24 loc) · 943 Bytes

readme.md

File metadata and controls

38 lines (24 loc) · 943 Bytes

v1.0 Release Note

"regularizeCNN" provides essential tools for evaluating the effects of different regularizers/optimizers when training a neural network.

Structure

Module diagram

System requirements

Hardware (recommended)

  • i7-950 or equivalent
  • 8G+ DRAM
  • Nvidia Geforce GPU with CUDA capability. 8G graphics memory.

Hardware (only for small batch evaluation)

  • i7-950 or equivalent
  • 8G+ DRAM

OS/Build Environment

** Listed is my working environment. Some other may also work.

  • Win7/8/10 /Ubuntu 18.04+/ Debian 9.7.0+
  • Python 3.7
  • Pytorch 1.5.0 + torchvision 0.6.0
  • numpy 1.18.5
  • matplotlib 3.2.2
  • tqdm 4.46.1
  • pillow 7.2.0

Usage

demoTrain.py : Train a new neural-net with specified dataset and save to file "._outputs/{timestamp}.model".

demoAdversary.py : Load a pre-trained net from file and generate adversary samples that can fool the net.