MCPcopy Create free account
hub / github.com/apple/axlearn / _create_layer_parameter_specs

Method _create_layer_parameter_specs

axlearn/common/layers.py:666–685  ·  view source on GitHub ↗
(self)

Source from the content-addressed store, hash-verified

664 def _create_layer_parameter_specs(self) -> dict[str, ParameterSpec]:
665 cfg = self.config
666 return {
667 "scale": ParameterSpec(shape=[cfg.input_dim], mesh_axes=(None,)),
668 "bias": ParameterSpec(shape=[cfg.input_dim], mesh_axes=(None,)),
669 "moving_mean": ParameterSpec(
670 shape=[cfg.input_dim],
671 dtype=jnp.float32,
672 mesh_axes=(None,),
673 initializer=constant_initializer(0.0),
674 weight_decay_scale=0,
675 ),
676 "moving_variance": ParameterSpec(
677 shape=[cfg.input_dim],
678 dtype=jnp.float32,
679 mesh_axes=(None,),
680 initializer=constant_initializer(1.0),
681 weight_decay_scale=0,
682 ),
683 }
684
685 def forward(self, x: Tensor, *, segment_ids: Optional[Tensor] = None) -> Tensor:
686 cfg = self.config
687 x_dtype = x.dtype
688 if cfg.forward_dtype is not None:

Callers

nothing calls this directly

Calls 2

ParameterSpecClass · 0.90
constant_initializerFunction · 0.90

Tested by

no test coverage detected