cherrypi::bos Namespace Reference

Classes
struct	ModelRunner
	Helper class for running BOS models. More...
struct	Sample
struct	StaticData

Enumerations
enum	BosFeature {   BosFeature::Undef, BosFeature::Map, BosFeature::MapId, BosFeature::Race,   BosFeature::Units, BosFeature::BagOfUnitCounts, BosFeature::BagOfUnitCountsAbs5_15_30, BosFeature::Resources5Log,   BosFeature::TechUpgradeBits, BosFeature::PendingTechUpgradeBits, BosFeature::TimeAsFrame, BosFeature::ActiveBo,   BosFeature::NextBo }
	A list of possible features that can be extracted from a Sample. More...

Functions
ag::Container	modelMakeFromCli (double dropout=0.0)
	Construct a BOS module according to command-line flags. More...
std::map< std::string, std::string >	modelFlags ()
	AUTOGRAD_CONTAINER_CLASS (IdleModel)
	AUTOGRAD_CONTAINER_CLASS (MapRaceEcoTimeFeaturize)
	AUTOGRAD_CONTAINER_CLASS (LinearModel)
	AUTOGRAD_CONTAINER_CLASS (MlpModel)
	AUTOGRAD_CONTAINER_CLASS (LstmModel)
	AUTOGRAD_CONTAINER_CLASS (ConvEncLstmModel)
std::unique_ptr< ModelRunner >	makeModelRunner (ag::Container model, std::string modelType)
std::map< std::string, int64_t > const &	buildOrderMap ()
std::vector< std::string > const &	targetBuilds ()
std::string	allowedTargetsAsFlag ()
std::string	allowedOpeningsAsFlag ()
torch::Tensor	getBuildOrderMaskByRace (char race)
torch::Tensor	getBuildOrderMaskByRace (int race)
int64_t	mapId (torch::Tensor mapFeats)
char	getOpponentRace (std::string const &opponent)
std::string	addRacePrefix (std::string buildOrder, char prefix)
std::string	addRacePrefix (std::string buildOrder, int race)
std::string	stripRacePrefix (std::string prefixedBo)
int64_t	buildOrderId (std::string const &bo)

Enumeration Type Documentation

enum cherrypi::bos::BosFeature

strong

A list of possible features that can be extracted from a Sample.

Enumerator
Undef
Map	Map features from StaticData.
MapId	Map "ID" based on sum of map features.
Race	2-dimensional: our and their race
Units	Defogger-style pooled unit types.
BagOfUnitCounts	Bag-of-words unit type counts.
BagOfUnitCountsAbs5_15_30	Bag-of-words unit type counts in future autobuild states (ours only)
Resources5Log	Ore/Gas/UsedPsi/TotalPsi: log(x / 5 + 1)
TechUpgradeBits	142-dim tech/upgrade vector: one bit for each upgrade/level/tech
PendingTechUpgradeBits	142-dim vector of pending upgrades/techs
TimeAsFrame	Numerical frame value.
ActiveBo	Id of active build order.
NextBo	Id of next build order.

Function Documentation

std::string cherrypi::bos::addRacePrefix ( std::string  buildOrder, char  prefix  )

inline

std::string cherrypi::bos::addRacePrefix ( std::string  buildOrder, int  race  )

inline

std::string cherrypi::bos::allowedOpeningsAsFlag

(

)

std::string cherrypi::bos::allowedTargetsAsFlag

(

)

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

IdleModel

)

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

MapRaceEcoTimeFeaturize

)

Expected input:

• a Bx1 tensor containing the map ID,
• a Bx2 tensor containing the races,
• a Bx4 tensor containing resource features.
• a Bx142 tensor containing upgrades and techonology features.
• a Bx142 tensor containing pending upgrades and techonology features.
• a Bx1 tensor containing the current time (in frames)
• a Bx1 tensor containing the active build

B can also be TxB.

Returns:

• a BxN tensor, with N being the sum of all embeddings

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

LinearModel

)

Expected input: Dictionary with "features" as tensor vector:

• a BxF tensor containing the unit type counts (F = n_unit_types)
• a BxF tensor containing our future unit type counts (F = n_unit_types / 2
• 3)
• a Bx1 tensor containing the map ID.
• a Bx2 tensor containing the races,
• a Bx4 tensor containing resource features.
• a Bx142 tensor containing upgrades and techonology features.
• a Bx142 tensor containing pending upgrades and techonology features.
• a Bx1 tensor containing the current time (in frames)
• a Bx1 tensor containing the active build

Returns an unordered_map:

• "vHeads": predicted values p(win) for all builds, masked wrt opponent race
• "Pi": softmax over p(win), masked wrt opponent race
• "V": overall value function – currently just zeros

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

MlpModel

)

Expected input: Dictionary with "features" as tensor vector:

• a BxF tensor containing the unit type counts (F = n_unit_types)
• a BxF tensor containing our future unit type counts (F = n_unit_types / 2
• 3)
• a Bx1 tensor containing the map ID.
• a Bx2 tensor containing the races,
• a Bx4 tensor containing resource features.
• a Bx142 tensor containing upgrades and techonology features.
• a Bx142 tensor containing pending upgrades and techonology features.
• a Bx1 tensor containing the current time (in frames)
• a Bx1 tensor containing the active build

Returns an unordered_map:

• "vHeads": predicted values p(win) for all builds, masked wrt opponent race
• "Pi": softmax over p(win), masked wrt opponent race
• "V": overall value function – currently just zeros

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

LstmModel

)

Expected input: Dictionary with "features" as tensor vector:

• a TxBxF tensor containing the unit type counts (F = n_unit_types)
• a TxBxF tensor containing our future unit type counts (F = n_unit_types / 2 * 3)
• a TxBx1 tensor containing the map ID.
• a 1xBx2 tensor containing the races,
• a TxBx4 tensor containing resource features.
• a TxBx142 tensor containing upgrades and techonology features.
• a TxBx142 tensor containing pending upgrades and techonology features.
• a TxBx1 tensor containing the current time (in frames)
• a TxBx1 tensor containing the active build And optionally
• "hidden": hidden activations for the LSTMs

The 'T' dimension may be omitted.

Returns an unordered_map:

• "vHeads": predicted values p(win) for all builds, masked wrt opponent race
• "Pi": softmax over p(win), masked wrt opponent race
• "V": overall value function – currently just zeros
• "hidden": the hidden state

cherrypi::bos::AUTOGRAD_CONTAINER_CLASS

(

ConvEncLstmModel

)

Expected input: Dictionary with "features" as tensor vector:

• a 1xBxCxhxw tensor containing the map features,
• a TxBx1 tensor containing the map ID.
• a 1xBx2 tensor containing the races,
• a TxBxFxHxW tensor containing the units features,
• a TxBxF tensor containing the unit type counts (F = n_unit_types)
• a TxBxF tensor containing our future unit type counts (F = n_unit_types / 2 * 3)
• a TxBx4 tensor containing resource features.
• a TxBx142 tensor containing upgrades and techonology features.
• a TxBx142 tensor containing pending upgrades and techonology features.
• a TxBx1 tensor containing the current frame number
• a TxBx1 tensor containing the active build And optionally
• "hidden": hidden activations for the LSTMs

The 'T' dimension may be omitted.

Returns an unordered_map:

• "vHeads": predicted values p(win) for all builds, masked wrt opponent race
• "Pi": softmax over p(win), masked wrt opponent race
• "V": overall value function – currently just zeros
• "hidden": the hidden state

int64_t cherrypi::bos::buildOrderId ( std::string const &  bo )

inline

std::map< std::string, int64_t > const & cherrypi::bos::buildOrderMap

(

)

torch::Tensor cherrypi::bos::getBuildOrderMaskByRace

(

char

race

)

torch::Tensor cherrypi::bos::getBuildOrderMaskByRace

(

int

race

)

char cherrypi::bos::getOpponentRace ( std::string const &  opponent )

inline

std::unique_ptr< ModelRunner > cherrypi::bos::makeModelRunner	(	ag::Container	model,
		std::string	modelType
	)

int64_t cherrypi::bos::mapId

(

torch::Tensor

mapFeats

)

std::map< std::string, std::string > cherrypi::bos::modelFlags

(

)

ag::Container cherrypi::bos::modelMakeFromCli

(

double

dropout

)

Construct a BOS module according to command-line flags.

std::string cherrypi::bos::stripRacePrefix ( std::string  prefixedBo )

inline

std::vector< std::string > const & cherrypi::bos::targetBuilds

(

)