cifar.torch

For original readme, please see https://github.com/szagoruyko/cifar.torch , and Sergey's blog post at http://torch.ch/blog/2015/07/30/cifar.html

This readme is for pytorch version, which handles data loading and preprocessing in python

Installation

• If using CUDA, you need at least CUDA 7.0 and CuDNN v4
• Install Torch:

git clone https://github.com/torch/distro.git ~/torch --recursive
cd ~/torch
# install dependencies.  To install everything:
  bash install-deps
# Or, if you're on ubuntu, you only need the following dependencies:
  sudo apt-get update -y
  sudo apt-get install -y wget git gcc g++ cmake libffi-dev \
       libblas-dev liblapack-dev libatlas-base-dev gfortran libreadline-dev
# install torch
./install.sh

• install torch cudnn (only if using CUDA) and nninit:

luarocks install cudnn
luarocks install nninit

• Setup python, eg for python 3.4:

sudo apt-get install python3.4-dev
virtualenv -p python3.4 ~/env34
source ~/env34/bin/activate
pip install docopt
pip install numpy

• Install pytorch:

git clone https://github.com/hughperkins/pytorch ~/pytorch
cd ~/pytorch
source ~/torch/install/bin/torch-activate
./build.sh

Data download

./download-cifar.sh

Training

python train.py

You should see the loss gradually decrease, and the test accuracy gradually increase.

Options:

  --save SAVE                  subdirectory to save logs [default: logs]
  --batchSize BATCHSIZE        batch size [default: 128]
  --learningRate LEARNINGRATE  learning rate [default: 1]
  --learningRateDecay LRDECAY  learning rate decay [default: 1e-7]
  --weightDecay WEIGHTDECAY    weightDecay [default: 0.0005]
  --momentum MOMENTUM          momentum [default: 0.9]
  --epoch_step EPOCHSTEP       epoch step [default: 25]
  --save_every SAVEEVERY       epochs between saves [default: 50]
  --model MODEL                model name [default: vgg_bn_drop]
  --max_epoch MAXEPOCH         maximum number of iterations [default: 300]
  --backend BACKEND            backend [float|cuda|cl] [default: cudnn]
  --cudnnfastest CUDNNFASTEST  use cudnn 'fastest' mode y/n [default: y]

Differences from original lua version

• data loading in python
• preprocessing in python
• no conversion from rgb to yuv (just because... haven't added it)
• no graph for now (but... it's python... you can use all the matplot goodness you are used to using :-) )
• defaults to cudnn, and cudnn.fastest (though switchable via commandline options)

Recent changes

19 April:

• added OpenCL support