Conv¶

conv ¶

WeightNorm(module: Module, weight_name: str = 'weight', dim: int = 0)

Bases: Module

Weight normalization wrapper for any module with a weight parameter.

Decomposes the weight into a magnitude (g) and direction (v): weight = g * (v / ||v||)

Parameters:

Name	Type	Description	Default
`module`	`Module`	The module to wrap (e.g., nn.Conv1d, nn.Linear).	required
`weight_name`	`str`	Name of the weight parameter to normalize.	`'weight'`
`dim`	`int`	Dimension over which to compute the norm. Default 0 normalizes per output channel.	`0`

Weight normalization for MLX modules.

Implements weight normalization (Salimans & Kingma, 2016) as a wrapper that reparameterizes a weight tensor as w = g * (v / ||v||).

WeightNorm(module: Module, weight_name: str = 'weight', dim: int = 0)

Bases: Module

Weight normalization wrapper for any module with a weight parameter.

Decomposes the weight into a magnitude (g) and direction (v): weight = g * (v / ||v||)

Parameters:

Name	Type	Description	Default
`module`	`Module`	The module to wrap (e.g., nn.Conv1d, nn.Linear).	required
`weight_name`	`str`	Name of the weight parameter to normalize.	`'weight'`
`dim`	`int`	Dimension over which to compute the norm. Default 0 normalizes per output channel.	`0`

weight_norm(module: Module, weight_name: str = 'weight', dim: int = 0) -> WeightNorm

Apply weight normalization to a module.

Parameters:

Name	Type	Description	Default
`module`	`Module`	Module to wrap (e.g., nn.Conv1d, nn.Linear).	required
`weight_name`	`str`	Name of the weight parameter.	`'weight'`
`dim`	`int`	Dimension for per-channel normalization.	`0`

Returns:

Type	Description
`WeightNorm`	WeightNorm wrapper around the module.

Example::

conv = weight_norm(nn.Conv1d(16, 32, 3))
out = conv(x)