vis4d.op.layer.weight_init¶

Model weight initialization.

Functions

`bias_init_with_prob`(prior_prob)	Initialize conv/fc bias value according to a given probability value.
`constant_init`(module, val[, bias])	Initialize module with constant value.
`kaiming_init`(module[, negative_slope, mode, ...])	Initialize module with Kaiming initialization.
`normal_init`(module[, mean, std, bias])	Initialize module with normal distribution.
`uniform_init`(module[, lower, upper, bias])	Initialize module with uniform distribution.
`xavier_init`(module[, gain, bias, distribution])	Initialize module with Xavier initialization.

bias_init_with_prob(prior_prob)[source]¶

Initialize conv/fc bias value according to a given probability value.

constant_init(module, val, bias=0.0)[source]¶

Initialize module with constant value.

kaiming_init(module, negative_slope=0.0, mode='fan_out', nonlinearity='relu', bias=0.0, distribution='normal')[source]¶

Initialize module with Kaiming initialization.

Parameters:

module (nn.Module) – Module to initialize.
negative_slope (float, optional) – The negative slope of the rectifier used after this layer (only used with 'leaky_relu'). Defaults to 0.0.
mode (str, optional) – Either 'fan_in' (default) or 'fan_out'. Choosing 'fan_in' preserves the magnitude of the variance of the weights in the forward pass. Choosing 'fan_out' preserves magnitudes in the backwards pass. Defaults to “fan_out”.
nonlinearity (str, optional) – The non-linear function (nn.functional name). Defaults to “relu”.
bias (float, optional) – The bias to use. Defaults to 0.0.
distribution (str, optional) – Either 'uniform' or 'normal'. Defaults to “normal”.

Return type:

None

normal_init(module, mean=0.0, std=1.0, bias=0)[source]¶

Initialize module with normal distribution.

uniform_init(module, lower=0.0, upper=1.0, bias=0.0)[source]¶

Initialize module with uniform distribution.

xavier_init(module, gain=1.0, bias=0.0, distribution='normal')[source]¶

Initialize module with Xavier initialization.