szagoruyko

#!/usr/bin/env python
import subprocess
import re


class bcolors:
    HEADER = '\033[95m'
    OKBLUE = '\033[94m'
    OKGREEN = '\033[92m'
    WARNING = '\033[93m'

szagoruyko

function removeDescriptors(net)
  for i,val in pairs(net.modules) do
    if tostring(val):find('cudnn') then
      for name,field in pairs(val) do
        if name:find('Desc') then
          val[name] = nil
        end
      end
      val.algType = nil
      val.iDesc = nil

szagoruyko

Using 1-th gpu
Loading ./data/ptb.train.txt, size of data = 929589
Loading ./data/ptb.valid.txt, size of data = 73760
Loading ./data/ptb.test.txt, size of data = 82430
Network parameters:
{
  layers : 2
  lr : 1
  max_max_epoch : 13
  max_grad_norm : 5

szagoruyko

Using 1-th gpu
Loading ./data/ptb.train.txt, size of data = 929589
Loading ./data/ptb.valid.txt, size of data = 73760
Loading ./data/ptb.test.txt, size of data = 82430
Network parameters:
{
  layers : 2
  lr : 1
  max_max_epoch : 13
  max_grad_norm : 5

szagoruyko

local BidirectionalSequencer, parent = torch.class('nn.BidirectionalSequencer', 'nn.Container')

function BidirectionalSequencer:__init(module_forward, module_backward, nOutputSize)
  parent.__init(self)
  self.module_forward = module_forward
  self.module_backward = module_backward
  self.modules[1] = nn.Sequencer(module_forward)
  self.modules[2] = nn.Sequencer(module_backward)

  self.output = {}

szagoruyko

require 'nn'

local vgg = nn.Sequential()
-- building block
local function ConvBNReLU(nInputPlane, nOutputPlane)
  vgg:add(nn.SpatialConvolution(nInputPlane, nOutputPlane, 3,3, 1,1, 1,1))
  vgg:add(nn.SpatialBatchNormalization(nOutputPlane,1e-3))
  vgg:add(nn.ReLU(true))
  return vgg
end

szagoruyko

-- a script to simplify trained net by incorporating every SpatialBatchNormalization to SpatialConvolution
-- and BatchNormalization to Linear

local function BNtoConv(net)
  for i,v in ipairs(net.modules) do
    if v.modules then
      BNtoConv(v)
    else
      if torch.typename(v) == 'nn.SpatialBatchNormalization' and
        (torch.typename(net:get(i-1)):find'SpatialConvolution') then

szagoruyko

local Linear, parent = torch.class('nn.NoBiasLinear', 'nn.Linear')

function Linear:__init(inputSize, outputSize)
   parent.__init(self, inputSize, outputSize)
   self.bias:fill(0)
end

function Linear:accGradParameters(input, gradOutput, scale)
   scale = scale or 1
   if input:dim() == 1 then

szagoruyko

Using Torch in Atom

Install Atom from https://atom.io or with brew cask install atom

Then install packages (example apm install hydrogen, apm is a package manager for atom):

• hydrogen
• autocomplete-paths
• language-lua

szagoruyko

Sparse tensors for Torch7

All types torch.DoubleTensor, torch.FloatTensor, etc. should have their sparse variants: torch.SparseDoubleTensor, torch.SparseFloatTensor, etc.

Copying between dense and sparse matrix should be done with :copy() function.

Underlying BLAS has to be swappable with MKL/OpenBLAS/Atlas, etc. Other math operations have to implemented with CSPARSE.