|
scale, shift = torch.chunk(emb, 2, dim=1) |
The code above, it makes: scale, shift order, This is fine and worked when both training and inference have the same module.
But when I want to load the pretrained F5-TTS module here, with huggingace/diffusers, where:
https://github.com/huggingface/diffusers/blob/20364fe5a2afbddbf2ec5097ec5d483ce0cc689a/src/diffusers/models/normalization.py#L74 that keeps shift first and scale second in both AdaLayerNorm and AdaLayerNorm_Final, it will make bad result.
suggest to keep an identical order.
F5-TTS/src/f5_tts/model/modules.py
Line 340 in 4533426
The code above, it makes: scale, shift order, This is fine and worked when both training and inference have the same module.
But when I want to load the pretrained F5-TTS module here, with huggingace/diffusers, where:
https://github.com/huggingface/diffusers/blob/20364fe5a2afbddbf2ec5097ec5d483ce0cc689a/src/diffusers/models/normalization.py#L74 that keeps shift first and scale second in both AdaLayerNorm and AdaLayerNorm_Final, it will make bad result.
suggest to keep an identical order.