nemo_automodel.components.models.bagel.connector

View as Markdown

Projection layers that connect BAGEL vision features to text hidden states.

Module Contents

Classes

NameDescription
BagelMultiModalProjectorProject vision-tower features into the language-model hidden size.
_ActivationWrap a Transformers activation callable in an nn.Module.

API

class nemo_automodel.components.models.bagel.connector.BagelMultiModalProjector(
in_dim: int,
out_dim: int,
hidden_act: str
)

Bases: Sequential

Project vision-tower features into the language-model hidden size.

class nemo_automodel.components.models.bagel.connector._Activation(
name: str
)

Bases: Module

Wrap a Transformers activation callable in an nn.Module.

fn
= ACT2FN[name]
nemo_automodel.components.models.bagel.connector._Activation.forward(
hidden_states
)