pytorch

Wrappers to use torch and lua from python

What is pytorch?

• create torch tensors, call operations on them
• instantiate nn network modules, train them, make predictions
• create your own lua class, call methods on that

Create torch tensors

import PyTorch
a = PyTorch.FloatTensor(2,3).uniform()
a += 3
print('a', a)
print('a.sum()', a.sum())

Instantiate nn network modules

import PyTorch
from PyTorchAug import nn

net = nn.Sequential()
net.add(nn.SpatialConvolutionMM(1, 16, 5, 5, 1, 1, 2, 2))
net.add(nn.ReLU())
net.add(nn.SpatialMaxPooling(3, 3, 3, 3))

net.add(nn.SpatialConvolutionMM(16, 32, 3, 3, 1, 1, 1, 1))
net.add(nn.ReLU())
net.add(nn.SpatialMaxPooling(2, 2, 2, 2))

net.add(nn.Reshape(32 * 4 * 4))
net.add(nn.Linear(32 * 4 * 4, 150))
net.add(nn.Tanh())
net.add(nn.Linear(150, 10))
net.add(nn.LogSoftMax())
net.float()

crit = nn.ClassNLLCriterion()
crit.float()

net.zeroGradParameters()
input = PyTorch.FloatTensor(5, 1, 28, 28).uniform()
labels = PyTorch.ByteTensor(5).geometric(0.9).icmin(10)
output = net.forward(input)
loss = crit.forward(output, labels)
gradOutput = crit.backward(output, labels)
gradInput = net.backward(input, gradOutput)
net.updateParameters(0.02)

Write your own lua class, call methods on it

Example lua class:

require 'torch'
require 'nn'

local TorchModel = torch.class('TorchModel')

function TorchModel:__init(backend, imageSize, numClasses)
  self:buildModel(backend, imageSize, numClasses)
  self.imageSize = imageSize
  self.numClasses = numClasses
  self.backend = backend
end

function TorchModel:buildModel(backend, imageSize, numClasses)
  self.net = nn.Sequential()
  local net = self.net

  net:add(nn.SpatialConvolutionMM(1, 16, 5, 5, 1, 1, 2, 2))
  net:add(nn.ReLU())
  net:add(nn.SpatialMaxPooling(3, 3, 3, 3))
  net:add(nn.SpatialConvolutionMM(16, 32, 3, 3, 1, 1, 1, 1))
  net:add(nn.ReLU())
  net:add(nn.SpatialMaxPooling(2, 2, 2, 2))
  net:add(nn.Reshape(32 * 4 * 4))
  net:add(nn.Linear(32 * 4 * 4, 150))
  net:add(nn.Tanh())
  net:add(nn.Linear(150, numClasses))
  net:add(nn.LogSoftMax())

  self.crit = nn.ClassNLLCriterion()

  self.net:float()
  self.crit:float()
end

function TorchModel:trainBatch(learningRate, input, labels)
  self.net:zeroGradParameters()

  local output = self.net:forward(input)
  local loss = self.crit:forward(output, labels)
  local gradOutput = self.crit:backward(output, labels)
  self.net:backward(input, gradOutput)
  self.net:updateParameters(learningRate)

  local _, prediction = output:max(2)
  local numRight = labels:int():eq(prediction:int()):sum()
  return {loss=loss, numRight=numRight}  -- you can return a table, it will become a python dictionary
end

function TorchModel:predict(input)
  local output = self.net:forward(input)
  local _, prediction = output:max(2)
  return prediction:byte()
end

Python script that calls this. Assume the lua class is stored in file "torch_model.lua"

import PyTorch
import PyTorchHelpers
import numpy as np
from mnist import MNIST

batchSize = 32
numEpochs = 2
learningRate = 0.02

TorchModel = PyTorchHelpers.load_lua_class('torch_model.lua', 'TorchModel')
torchModel = TorchModel(backend, 28, 10)

mndata = MNIST('../../data/mnist')
imagesList, labelsList = mndata.load_training()
labels = np.array(labelsList, dtype=np.uint8)
images = np.array(imagesList, dtype=np.float32)
labels += 1  # since torch/lua labels are 1-based
N = labels.shape[0]

numBatches = N // batchSize
for epoch in range(numEpochs):
  epochLoss = 0
  epochNumRight = 0
  for b in range(numBatches):
    res = torchModel.trainBatch(
      learningRate,
      images[b * batchSize:(b+1) * batchSize],
      labels[b * batchSize:(b+1) * batchSize])
    numRight = res['numRight']
    epochNumRight += numRight
  print('epoch ' + str(epoch) + ' accuracy: ' + str(epochNumRight * 100.0 / N) + '%')

It's easy to modify the lua script to use CUDA, or OpenCL.

Installation

Pre-requisites

• Have installed torch, following instructions at https://github.com/torch/distro
• Have installed 'nn' torch module:

luarocks install nn

• Have installed python (tested with 2.7 and 3.4)
• lua51 headers should be installed, ie something like sudo apt-get install lua5.1 liblua5.1-dev Run:

pip install -r requirements.txt

• To be able to run tests, also do:

pip install -r test/requirements.txt

Procedure

Run:

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

Unit-tests

Run:

source ~/torch/install/bin/torch-activate
cd pytorch
./run_tests.sh

Python 2 vs Python 3?

• pytorch is developed and maintained on python 3
• you should be able to use it with python 2, but there might be the occasional oversight. Please log an issue for any python 2 incompatibilities you notice

Maintainer guidelines

Maintainer guidelines

Versioning

semantic versioning

Related projects

Examples of training models/networks using pytorch:

• pytorch-residual-networks port of Michael Wilber's torch-residual-networks, to handle data loading and preprocessing from Python, via pytorch
• cifar.pytorch pytorch implementation of Sergey's cifar.torch

Addons, for using cuda tensors and opencl tensors directly from python (no need for this to train networks. could be useful if you want to manipulate cuda tensor directly from python)

• pycltorch python wrappers for cltorch and clnn
• pycudatorch python wrappers for cutorch and cunn

Support?

Please note that currently, right now, I'm focused 100.000% on cuda-on-cl, so please be patient during this period

Recent news

12 September:

• Yannick Hold-Geoffroy added conversion of lists and tuples to Lua tables

8 September:

• added PyTorchAug.save(filename, object) and PyTorchAug.load(filename), to save/load Torch .t7 files

26 August:

• if not deploying to a virtual environment, will install with --user, into home directory

14 April:

• stack trace should be a bit more useful now :-)

17 March:

• ctrl-c works now (tested on linux)

16 March:

• uses luajit on linux now (mac os x continues to use lua)

6 March:

• all classes should be usable from nn now, without needing to explicitly register inside pytorch
  • you need to upgrade to v3.0.0 to enable this, which is a breaking change, since the nn classes are now in PyTorchAug.nn, instead of directly in PyTorchAug

5 March:

• added PyTorchHelpers.load_lua_class(lua_filename, lua_classname) to easily import a lua class from a lua file
• can pass parameters to lua class constructors, from python
• can pass tables to lua functions, from python (pass in as python dictionaries, become lua tables)
• can return tables from lua functions, to python (returned as python dictionaries)

2 March:

• removed requirements on Cython, Jinja2 for installation

28th Februrary:

• builds ok on Mac OS X now :-) See https://travis-ci.org/hughperkins/pytorch/builds/112292866

26th February:

• modified / to be the div operation for float and double tensors, and // for int-type tensors, such as byte, long, int
• since the div change is incompatible with 1.0.0 div operators, jumping radically from 1.0.0 to 2.0.0-SNAPSHOT ...
• added dependency on numpy
• added .asNumpyTensor() to convert a torch tensor to a numpy tensor

24th February:

• added support for passing strings to methods
• added require
• created prototype for importing your own classes, and calling methods on those
• works with Python 3 now :-)

Older changes