Command Line Arguments¶

In addition to the Compile-Time Variables, we make use of the JSL::Argument class, allowing easy integration of command-line arguments for a number of properties.

Arguments & Triggers¶

Variable Name	Type	Trigger	Purpose	Default Value
`ConfigFile`	`std::string`	config	The first argument read: if set away from the default, reverts to reading all arguments from the chosen configuration file instead of the command line.	“__null_location__”
`ConfigDelimiter`	`char`	config-delim	The delimiter character used to parse the config file	‘ ‘
`RandomSeed`	`int`	random-seed	The value passed to srand() at startup for reproducible randomness	0
`DataSource`	`std::string`	data	The directory for the stellar data lists	“../../Data/ShuffledData”
`OutputDirectory`	`std::string`	output	The director for the output data (created if it doesn’t already exist)	“Output”
`StartVectorLocation`	std::string`	restart	The directory to search in for a valid savefile for vector initialisation (if set to default, uses random vector)	“__null_location__”
`GradLim`	`double`	gradlim	The maximum value of \(\nabla\mathcal{L}\) which will be considered ‘converged’	0
`MaxSteps`	`int`	max-steps	The maximum number of epochs the optimiser may use before exiting	1000
`SaveSteps`	`int`	save-steps	The number of steps between saving locations	1
`SaveAllSteps`	`bool`	unique-temp-save	If true, the temporary, raw vectors are saved uniquely. Recommended to set this to false to prevent huge amounts of data generation.	0
`Minibatches`	`int`	minibatch	The maximum number of batches used per epoch in the SGD prescription	64
`HarnessSlowDown`	`double`	harness-slow	The factor by which step sizes are reduced when the harness is active	10
`HarnessRelease`	`int`	harness-release	The number of full epochs over which the step size recovers from the HarnessSlowDown	5

Usage¶

Commands can be passed into the code in one of two ways: either through the commandline interface:

> ./theia -minibatch 16 -gradlim 0.1 -harness-release 12

Or through a configuration file, written as:

//configuration.txt
minibatch 16
gradlim 0.1
data ../new/data/location
random-seed 299

And launched as:

> ./theia -config configuration.txt

Note that if a configuration file is attached, all other command line arguments will be ignored.

CommandArgs Holder Class¶

Once loaded, the command-line arguments are stored within a central CommandArgs object

class CommandArgs¶

Public Functions

inline CommandArgs()¶: Default constructor….doesn’t do anything as the arguments self-initialise.

inline void ReadArguments(int argc, char *argv[], int ProcessRank)¶

Initialise the command-line arguments, and check if a configuration file is requested. Note that there is no checking for repeat arguments or multiply defined trigger strings, so multiple assignments are perfectly possible.

Parameters

argc – The system-provided command argument count
*argv[] – The system-provided list of command arguments
ProcessRank – The MPI-provided ID of the current process

Returns

Initialises the object against the provided parameters

Public Members

Argument<int> RandomSeed = Argument<int>(0, "random-seed")¶: The value passed to srand() <>_ for reproducible randomness.

Argument<std::string> StartVectorLocation = Argument<std::string>("__null_location__", "restart")¶: The directory to search in for a valid savefile configuration for relaunch.

Argument<double> GradLim = Argument<double>(0, "gradlim")¶: The maximum value of :math:\nabla\mathcal{L} which will be considered ‘converged’

Argument<int> MaxSteps = Argument<int>(1000, "max-steps")¶: The maximum number of epochs the optimiser may use before exiting.

Argument<int> SaveSteps = Argument<int>(1, "save-steps")¶: The number of steps between saving locations.

Argument<bool> SaveAllSteps = Argument<bool>(false, "unique-temp-save")¶: If true, the temporary, raw vectors are saved uniquely. Recommended to set this to false to prevent huge amounts of data generation.

Argument<int> Minibatches = Argument<int>(64, "minibatch")¶: The maximum number of batches used per epoch in the SGD prescription.

Argument<double> HarnessSlowDown = Argument<double>(10, "harness-slow")¶: The factor by which step sizes are reduced when the harness is active.

Argument<int> HarnessRelease = Argument<int>(5, "harness-release")¶: The number of full epochs over which the step size recovers from the HarnessSlowDown.

Argument<std::string> DataSource = Argument<std::string>("../../Data/ShuffledData", "data")¶: The directory for the stellar data lists.

Argument<std::string> OutputDirectory = Argument<std::string>("Output", "output")¶: The director for the output data (created if it doesn’t already exist)

std::vector<JSL::ArgumentInterface*> argPointers = {&RandomSeed, &StartVectorLocation, &GradLim, &MaxSteps, &DataSource, &OutputDirectory, &SaveSteps, &SaveAllSteps, &Minibatches, &HarnessRelease, &HarnessSlowDown}¶: Pointers list so can easily loop over the (heterogenous) array for assigments.