decagon-pytorch/decagon_pytorch/weight.py

import torch
import numpy as np


def weight_variable_glorot(input_dim, output_dim):
    """Create a weight variable with Glorot & Bengio (AISTATS 2010)
    initialization.
    """
    init_range = np.sqrt(6.0 / (input_dim + output_dim))
    initial = -init_range + 2 * init_range * \
        torch.rand(( input_dim, output_dim ), dtype=torch.float32)
    initial = initial.requires_grad_(True)
    return initial