def pt_init(self, init_var=1e-2, init_bias=0., rho=0.5, lmbd=0., l2=0., SI=15, **kwargs): """ """ # 2*self.shape[0]: precision parameters have size shape[0] pt_params = gzeros(self.m_end + self.shape[1] + 2 * self.shape[0]) if init_var is None: pt_params[:self.m_end] = gpu.garray( init_SI(self.shape, sparsity=SI)).ravel() else: pt_params[:self.m_end] = init_var * gpu.randn(self.m_end) pt_params[self.m_end:-self.shape[0]] = init_bias pt_params[-self.shape[0]:] = 1. self.pt_score = self.reconstruction self.pt_grad = self.grad_cd1 self.l2 = l2 self.rho = rho self.lmbd = lmbd self.rho_hat = None return pt_params
def pt_init(self, score=None, init_var=1e-2, init_bias=0., SI=15, **kwargs): if init_var is None: self.init_var = None self.SI = SI self.p[:self.m_end] = gpu.garray(init_SI(self.shape, sparsity=SI)).ravel() else: self.SI = SI self.init_var = init_var self.p[:self.m_end] = init_var * gpu.randn(self.m_end) self.p[self.m_end:] = init_bias self.score = score return self.p
def pt_init(self, score=None, init_var=1e-2, init_bias=0., l2=0., SI=15, **kwargs): pt_params = gzeros(self.m_end + self.shape[1] + self.shape[0]) if init_var is None: pt_params[:self.m_end] = gpu.garray(init_SI(self.shape, sparsity=SI)).ravel() else: pt_params[:self.m_end] = init_var * gpu.randn(self.m_end) pt_params[self.m_end:] = init_bias self.score = score self.l2 = l2 return pt_params
def pt_init(self, score=None, init_var=1e-2, init_bias=0., l2=0., SI=15, **kwargs): pt_params = gzeros(self.m_end + self.shape[0]) if init_var is None: pt_params[:self.m_end] = gpu.garray(init_SI(self.shape, sparsity=SI)).ravel() else: pt_params[:self.m_end] = init_var * gpu.randn(self.m_end) pt_params[self.m_end:] = init_bias self.score = score self.l2 = l2 return pt_params
def pt_init(self, init_var=1e-2, init_bias=0., rho=0.5, lmbd=0., l2=0., SI=15, **kwargs): """ """ # 2*self.shape[0]: precision parameters have size shape[0] pt_params = gzeros(self.m_end + self.shape[1] + 2*self.shape[0]) if init_var is None: pt_params[:self.m_end] = gpu.garray(init_SI(self.shape, sparsity=SI)).ravel() else: pt_params[:self.m_end] = init_var * gpu.randn(self.m_end) pt_params[self.m_end:-self.shape[0]] = init_bias pt_params[-self.shape[0]:] = 1. self.pt_score = self.reconstruction self.pt_grad = self.grad_cd1 self.l2 = l2 self.rho = rho self.lmbd = lmbd self.rho_hat = None return pt_params