Sergey Zagoruyko 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

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.

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


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

	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

	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

	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

	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 = {}

	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

	-- 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

	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