deepity.DKPPCN.DKPPCN Class Reference

Inheritance diagram for deepity.DKPPCN.DKPPCN:

Public Member Functions
None	__init__ (self, *Layer|Activation architecture, Optional[int] batch_size=None, str device="cpu")
"DKPPCN"	configure (self, float learning_rate=1e-6, float inference_rate=0.1, float feedback_rate=1e-4, float lmbda=1e-2, str optimizer="SGD", str psi_optimizer="SGD")
None	set_optimizer (self, str optimizer)
None	set_psi_optimizer (self, str optimizer)
None	set_learning_rate (self, float learning_rate)
None	set_feedback_rate (self, float feedback_rate)
None	set_inference_rate (self, float inference_rate)
None	compile (self)
None	randomize_weights (self, Optional[int] seed=None, str dist="normal")
float	train_step (self, npt.NDArray[np.float32] X, npt.NDArray[np.float32] Y, int inference_steps=1)
npt.NDArray[np.float32]	predict (self, npt.NDArray[np.float32] X, int inference_steps=1)
"DKPPCN"	fit (self, npt.NDArray[np.float32] X, npt.NDArray[np.float32] Y, int epochs, int inference_steps=1, Optional[float] initial_lr=None, float decay_rate=1.0, bool shuffle=True)

Public Attributes
	architecture = architecture
	device = device
	layers

Protected Member Functions
None	_validate_architecture (self)
None	_build_backend (self)
int	_terminal_size (self)
None	_require_configured (self)

Protected Attributes
Optional[float]	_learning_rate = None
Optional[float]	_inference_rate = None
Optional[float]	_feedback_rate = None
Optional[float]	_lambda = None
Optional[str]	_optimizer = None
Optional[str]	_psi_optimizer = None

Detailed Description

Predictive Coding Network using Direct Kolen-Pollack (DKP)
feedback alignment.

DKP maintains independently learned forward weights (W) and
direct feedback weights (Psi). The feedback pathway connects
hidden layers directly to the terminal error rather than
relaying errors layer-by-layer.

Unlike the other PCN implementations, DKP performs a direct
feedback alignment update before the ordinary settling loop.
This allows inference_steps=1 to be the default operating point.

Constructor & Destructor Documentation

__init__()

None deepity.DKPPCN.DKPPCN.__init__	(		self,
		*Layer | Activation	architecture,
		Optional[int]	batch_size = None,
		str	device = "cpu" )

Member Function Documentation

_build_backend()

None deepity.DKPPCN.DKPPCN._build_backend ( self )

protected

_require_configured()

None deepity.DKPPCN.DKPPCN._require_configured ( self )

protected

_terminal_size()

int deepity.DKPPCN.DKPPCN._terminal_size ( self )

protected

_validate_architecture()

None deepity.DKPPCN.DKPPCN._validate_architecture ( self )

protected

compile()

None deepity.DKPPCN.DKPPCN.compile

(

self

)

Allocate backend buffers for the network built by
_build_backend(). Called automatically by configure().

configure()

"DKPPCN" deepity.DKPPCN.DKPPCN.configure	(		self,
		float	learning_rate = 1e-6,
		float	inference_rate = 0.1,
		float	feedback_rate = 1e-4,
		float	lmbda = 1e-2,
		str	optimizer = "SGD",
		str	psi_optimizer = "SGD" )

Configure and initialize the DKP network.

Parameters
----------
learning_rate:
    Learning rate for the forward weights W.

inference_rate:
    Inference/settling rate.

feedback_rate:
    Learning rate for the direct feedback weights Psi.

lmbda:
    Regularization parameter.

optimizer:
    Optimizer used for W.

psi_optimizer:
    Independent optimizer used for Psi.

fit()

"DKPPCN" deepity.DKPPCN.DKPPCN.fit	(		self,
		npt.NDArray[np.float32]	X,
		npt.NDArray[np.float32]	Y,
		int	epochs,
		int	inference_steps = 1,
		Optional[float]	initial_lr = None,
		float	decay_rate = 1.0,
		bool	shuffle = True )

Train the network for multiple epochs.

DKP defaults to one inference step because direct feedback
alignment is specifically intended to reduce the amount of
iterative settling required.

predict()

npt.NDArray[np.float32] deepity.DKPPCN.DKPPCN.predict	(		self,
		npt.NDArray[np.float32]	X,
		int	inference_steps = 1 )

Run inference on `X`, settling for `inference_steps` iterations.
@return The terminal layer's beliefs.

randomize_weights()

None deepity.DKPPCN.DKPPCN.randomize_weights	(		self,
		Optional[int]	seed = None,
		str	dist = "normal" )

Randomize all network weights. `seed` gives a reproducible run;
`dist` e.g. "normal(0,1)" or "uniform(-0.3,0.3)" (the default
"normal" has no "(a,b)" and so deliberately falls back to each
layer's own existing init).

set_feedback_rate()

None deepity.DKPPCN.DKPPCN.set_feedback_rate	(		self,
		float	feedback_rate )

Set the learning rate for the feedback weights Psi.

set_inference_rate()

None deepity.DKPPCN.DKPPCN.set_inference_rate	(		self,
		float	inference_rate )

Set the inference/settling rate.

set_learning_rate()

None deepity.DKPPCN.DKPPCN.set_learning_rate	(		self,
		float	learning_rate )

Set the learning rate for the forward weights W.

set_optimizer()

None deepity.DKPPCN.DKPPCN.set_optimizer	(		self,
		str	optimizer )

Set the optimizer used for the forward weights W.

set_psi_optimizer()

None deepity.DKPPCN.DKPPCN.set_psi_optimizer	(		self,
		str	optimizer )

Set the optimizer used for the feedback weights Psi.

train_step()

float deepity.DKPPCN.DKPPCN.train_step	(		self,
		npt.NDArray[np.float32]	X,
		npt.NDArray[np.float32]	Y,
		int	inference_steps = 1 )

Run one training step on a single batch: a direct feedback
alignment update, then settling for `inference_steps`
iterations before updating weights.
@return Total settling energy accumulated over `inference_steps`.

Member Data Documentation

_feedback_rate

Optional[float] deepity.DKPPCN.DKPPCN._feedback_rate = None

protected

_inference_rate

Optional[float] deepity.DKPPCN.DKPPCN._inference_rate = None

protected

_lambda

Optional[float] deepity.DKPPCN.DKPPCN._lambda = None

protected

_learning_rate

Optional[float] deepity.DKPPCN.DKPPCN._learning_rate = None

protected

_optimizer

Optional[str] deepity.DKPPCN.DKPPCN._optimizer = None

protected

_psi_optimizer

Optional[str] deepity.DKPPCN.DKPPCN._psi_optimizer = None

protected

architecture

deepity.DKPPCN.DKPPCN.architecture = architecture

device

deepity.DKPPCN.DKPPCN.device = device

layers

deepity.DKPPCN.DKPPCN.layers

---

The documentation for this class was generated from the following file:

• deepity/DKPPCN.py