Skip to content

Conversation

@dariowsz
Copy link

The Issue

PretransformConditioner was applying nn.Linear projection to the wrong dimension. Pretransform outputs are [batch, channels, time], but Linear operates on the last dimension, so it was projecting time instead of channels.

Fix

Added a transpose before projection to convert [batch, channels, time][batch, time, channels] so the Linear layer projects channels correctly. Output is now [batch, time, output_dim], matching other text-based conditioners.

Changes

  • Transpose latents before applying proj_out
  • Updated mask generation to use the correct dimension

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant