def __init__( self, hidden_size: int, num_outputs: int, conditional_sigma: bool = False, tanh_squash: bool = False, ): super().__init__() self.conditional_sigma = conditional_sigma self.mu = linear_layer( hidden_size, num_outputs, kernel_init=Initialization.KaimingHeNormal, kernel_gain=0.1, bias_init=Initialization.Zero, ) self.tanh_squash = tanh_squash if conditional_sigma: self.log_sigma = linear_layer( hidden_size, num_outputs, kernel_init=Initialization.KaimingHeNormal, kernel_gain=0.1, bias_init=Initialization.Zero, ) else: self.log_sigma = nn.Parameter( torch.zeros(1, num_outputs, requires_grad=True))
def __init__(self): super().__init__() self.__global_step = nn.Parameter(torch.Tensor([0]).to(torch.int64), requires_grad=False)