Skip to content

Why were the self-attention layers removed from Janner's UNet? #147

@RexGLiu

Description

@RexGLiu

I was looking at the modifications made to Janner's UNet to implement FiLM conditioning. One thing I noticed was that the self-attention layers were removed. Compare for instance these:

  1. https://github.com/jannerm/diffuser/blob/7ea422860cc0106e5ca5949d980f04b799d5462c/diffuser/models/temporal.py#L85
  2. Downsample1d(dim_out) if not is_last else nn.Identity()

Apart from the missing self-attention and the introduction of FiLM conditioning, everything else is identical. Was there any reason for this design choice? Does the self-attention mess up the FiLM conditioning in the conv-1d layers?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions