Torch7-tools

This would be a collection of useful routines and function currently missing in Torch7.
Most of the time the scripts are just usefull routines or functions I am using for my job and I am allowed to extract and publish here.

Tools description

Table of contents

• Network lightener
• Gradient repopulation
• Command line parser
• Cuda to Float network converter

Important!

The new net-toolkit Torch7 package will be released soon. Basically it will provide a way for easily handling networks and it will include the following two functions, which are to be considered therefore deprecated.

Network lightener

netLighter library provides a saveNet(name, network) function which saves a lighter version of your current network, removing all unnecessary data from it (such as gradients, temporal data and etc...). The default location is ./; if a different one is preferred instead, you may want to specify it with a global option opt.save = 'my-path/'. Usage:

-- if './' is not ok:
opt = opt or {}
opt.save = 'my-path/'
-- otherwise, it will be sufficient only:
require 'netLighter'
-- net = nn.Sequential() and other stuff
saveNet('myNet.net',net)

Gradient repopulation

Let's say we would like to load a network we have previously saved with saveNet() for continuing a training > session on it. Some inner parameters (something about gradients) have to be restored, since saveNet() did a pruning operation on the network in order to save space. Here is how we can handle this case:

local networkFile = '/path/To/MyNet.net'
model = torch.load(networkFile)
repopulateGrad(model)

Now we can keep training, perhaps without forgetting to define a criterion loss (the criterion is not saved with the network, so we have to re-define it, if we don't already do it somewhere else in the code).

Command line parser

penlightTest shows a great deal of command line parser options that could turn helpful when we need to send some initial configuration values to the script in a compact manner. Running the script in lua (or torch) with no argument will print on screen the help screen (usually reachable with the option --help or -h, which in this case has been deliberately overwritten to be the height handle).

lua penlightTest.lua

Moreover, you can notice the presence of default false and default true flags and short multi-letter handles as well, which are not iterpreted as in the Unix environment (say like ls -al) because of slack = true of lapp has been (intentionally) set to true. You can run the test like this (it will show everything it can do that I care)

lua penlightTest.lua -v abc --time 40 -h 20 -sk 15 --flag1 -f3

The same two lines above could have been written by sustituting lua with torch (and this is what I usually do). Further documentation on the lapp command line parser can be found here.

Cuda to Float network converter

netConverter converts a cuda network into an equivalent float one.

./netConverter.lua ../results/multinet-cuda.net

The scripts automatically changes the extension to -float.net and saves the new network in the same location of the input one (which is specified as argument of the routine, as you can see from the example above).