SN Class Reference

Inheritance diagram for SN:

Public Member Functions
def	__init__ (self, num_svs, num_itrs, num_outputs, transpose=False, eps=1e-12)
def	u (self)
def	sv (self)
def	W_ (self)

Public Attributes
	num_itrs
	Number of power iterations per step.
	num_svs
	Number of singular values.
	transpose
	Transposed?
	eps
	Epsilon value for avoiding divide-by-0.

Detailed Description

Spectral normalization base class

Definition at line 226 of file ieagan.py.

Constructor & Destructor Documentation

__init__()

def __init__	(		self,
			num_svs,
			num_itrs,
			num_outputs,
			transpose = False,
			eps = 1e-12
	)

constructor

Reimplemented in SNConv2d, SNLinear, and SNEmbedding.

Definition at line 232 of file ieagan.py.

    def __init__(self, num_svs, num_itrs, num_outputs, transpose=False, eps=1e-12):
        """constructor"""
 
        
        self.num_itrs = num_itrs
        
        self.num_svs = num_svs
        
        self.transpose = transpose
        
        self.eps = eps
        # Register a singular vector for each sv
        for i in range(self.num_svs):
            self.register_buffer(f"u{i:d}", torch.randn(1, num_outputs))
            self.register_buffer(f"sv{i:d}", torch.ones(1))
 

Member Function Documentation

sv()

def sv

(

self

)

Singular values
note that these buffers are just for logging and are not used in training.

Definition at line 256 of file ieagan.py.

    def sv(self):
        """
        Singular values
        note that these buffers are just for logging and are not used in training.
        """
        return [getattr(self, f"sv{i:d}") for i in range(self.num_svs)]
 

u()

def u

(

self

)

Singular vectors (u side)

Definition at line 249 of file ieagan.py.

    def u(self):
        """
        Singular vectors (u side)
        """
        return [getattr(self, f"u{i:d}") for i in range(self.num_svs)]
 

W_()

def W_

(

self

)

Compute the spectrally-normalized weight

Definition at line 263 of file ieagan.py.

    def W_(self):
        """
        Compute the spectrally-normalized weight
        """
        W_mat = self.weight.view(self.weight.size(0), -1)
        if self.transpose:
            W_mat = W_mat.t()
        # Apply num_itrs power iterations
        for _ in range(self.num_itrs):
            svs, _, _ = power_iteration(
                W_mat, self.u, update=self.training, eps=self.eps
            )
            # Update the svs
        if self.training:
            # Make sure to do this in a no_grad() context or you'll get memory leaks!  # noqa
            with torch.no_grad():
                for i, sv in enumerate(svs):
                    self.sv[i][:] = sv
        return self.weight / svs[0]
 
 

Member Data Documentation

eps

eps

Epsilon value for avoiding divide-by-0.

Definition at line 242 of file ieagan.py.

num_itrs

num_itrs

Number of power iterations per step.

Definition at line 236 of file ieagan.py.

num_svs

num_svs

Number of singular values.

Definition at line 238 of file ieagan.py.

transpose

transpose

Transposed?

Definition at line 240 of file ieagan.py.

---

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

• pxd/scripts/pxd/background_generator/models/ieagan.py