megengine.module.init¶

>>> import megengine.module as M
>>> m = M.Conv2d(16, 33, 3, stride=2)
>>> M.init.msra_normal_(m.weight, mode="fan_out", nonlinearity="relu")

注解

良好的初始化策略有助于你的模型在训练时更快地收敛。

Initialization¶

`fill_`	Fills the given `tensor` with value `val`.
`zeros_`	Fills the given `tensor` with scalar value 0.
`ones_`	Fills the given `tensor` with the scalar value 1.
`uniform_`	Fills the given `tensor` with random value sampled from uniform distribution \(\mathcal{U}(\text{a}, \text{b})\).
`normal_`	Fills the given `tensor` with random value sampled from normal distribution \(\mathcal{N}(\text{mean}, \text{std}^2)\).
`calculate_gain`	Returns a recommended gain value (see the table below) for the given nonlinearity function.
`calculate_fan_in_and_fan_out`	Calculates fan_in / fan_out value for given weight tensor.
`calculate_correct_fan`	Calculates fan_in / fan_out value for given weight tensor, depending on given `mode`.
`xavier_uniform_`	Fills tensor with random values sampled from \(\mathcal{U}(-a, a)\) where
`xavier_normal_`	Fills tensor with random values sampled from \(\mathcal{N}(0, \text{std}^2)\) where
`msra_uniform_`	Fills tensor wilth random values sampled from \(\mathcal{U}(-\text{bound}, \text{bound})\) where
`msra_normal_`	Fills tensor wilth random values sampled from \(\mathcal{N}(0, \text{std}^2)\) where

megengine.module.quantized.DequantStub.zero_grad

megengine.module.init.fill_