torchnet-vision

torchnet-vision is a plugin for torchnet which provides a set of abstractions aiming at encouraging code re-use as well as encouraging modular programming.

At the moment, torchnet-vision provides several functionalities:

• TransformImage: pre-processing image in various ways.
• Datasets: download and load dataset easily.
• Models: download and load models easily.

For an overview of the torchnet framework, please also refer to this paper.

Installation

Please install torch first, following instructions on torch.ch. If torch is already installed, make sure you have an up-to-date version of argcheck, otherwise you will get weird errors at runtime.

Assuming torch is already installed, the torchnet and torchnet-vision cores are only a set of lua files, so it is straightforward to install it with luarocks

luarocks install torchnet
git clone https://github.com/Cadene/torchnet-vision.git
cd torchnet-vision
luarocks make rocks/torchnet-vision-scm-1.rockspec

Documentation

Extraction features from lena with inceptionv3

require 'image'
tnt = require 'torchnet'
vision = require 'torchnet-vision'

augmentation = tnt.transform.compose{
   vision.image.transformimage.randomScale{minSize=299,maxSize=350},
   vision.image.transformimage.randomCrop(299),
   vision.image.transformimage.colorNormalize{
      mean = vision.models.inceptionv3.mean,
      std  = vision.models.inceptionv3.std
   },
   function(img) return img:float() end
}
img = augmentation(image.lena())

net = vision.models.inceptionv3.loadExtracting{ -- download included
   filename = 'tmp/inceptionv3.t7',
   layerid  = 30
}
net:evaluate()
print(net:forward(img:view(1,3,299,299)):size()) -- 2048

Fine tuning on MIT67 in 250 lines of code

$ CUDA_VISIBLE_DEVICES=0 th example/mit67finetuning.lua -usegpu true
$ ls demo/logs/mit67/*/
$ cat demo/logs/mit67/*/trainlog.txt
$ cat demo/logs/mit67/*/testlog.txt

Other projects using torchnet

• Wide residual network
• MultiGPU ImageNet
• Learning to Compare Image Patches via Convolutional Neural Networks (CVPR 2015)
• Facebook "A MultiPath Network for Object Detection"

Roadmap

• defining names for package and classes (vision?)
• add docs to TransformImage methods
• add test
• add a method to tnt.DataIterator to process the mean and std
• add a better system to preprocess images than tnt.transform (especially to add or remove TransformImage.colorNormalize)
• add data loaders (largscale or not)
• add video directory