From 4926a993cf94d03aad4d24be2515bbe70c03c42a Mon Sep 17 00:00:00 2001 From: Carlos Lordelo Date: Sun, 6 Oct 2024 19:07:53 +0200 Subject: [PATCH 1/3] Add finetune as CLAPAudioConditioner argument --- stable_audio_tools/models/conditioners.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/stable_audio_tools/models/conditioners.py b/stable_audio_tools/models/conditioners.py index e998ab10..a365a3a6 100644 --- a/stable_audio_tools/models/conditioners.py +++ b/stable_audio_tools/models/conditioners.py @@ -179,9 +179,11 @@ def __init__(self, clap_ckpt_path, audio_model_type="HTSAT-base", enable_fusion=True, - project_out: bool = False): + project_out: bool = False, + finetune: bool = False): super().__init__(512, output_dim, project_out=project_out) + self.finetune = finetune device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Suppress logging from transformers From 0f09c6e5dd341ebecd66711c9acaba30ae89e846 Mon Sep 17 00:00:00 2001 From: Carlos Lordelo Date: Wed, 27 May 2026 14:23:57 +0200 Subject: [PATCH 2/3] Fix rolled_metadata for contrastive disc during ARC training --- stable_audio_tools/training/arc.py | 13 ++++--------- 1 file changed, 4 insertions(+), 9 deletions(-) diff --git a/stable_audio_tools/training/arc.py b/stable_audio_tools/training/arc.py index e3e71fd3..e7ca5af1 100644 --- a/stable_audio_tools/training/arc.py +++ b/stable_audio_tools/training/arc.py @@ -1319,15 +1319,10 @@ def training_step(self, batch, batch_idx): logit_center_penalty = 0 if self.do_contrastive_disc: - - rolled_metadata = [] - - for i in range(reals.shape[0]): - rolled_keys = ["prompt"] - rolled_metadata.append(metadata[i]) - for rolled_key in rolled_keys: - rolled_metadata[i][rolled_key] = metadata[(i + 1) % reals.shape[0]][rolled_key] - + + n = reals.shape[0] + rolled_metadata = [{**md, "prompt": metadata[(i + 1) % n]["prompt"]} for i, md in enumerate(metadata)] + rolled_conditioning = self.discriminator.conditioner(rolled_metadata, self.device) if self.inpainting_config is not None: From 6baa57318e35fe6bebf65c0a66d0b1b509290266 Mon Sep 17 00:00:00 2001 From: Carlos Lordelo Date: Wed, 27 May 2026 14:40:16 +0200 Subject: [PATCH 3/3] Revert "Add finetune as CLAPAudioConditioner argument" This reverts commit 4926a993cf94d03aad4d24be2515bbe70c03c42a. --- stable_audio_tools/models/conditioners.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/stable_audio_tools/models/conditioners.py b/stable_audio_tools/models/conditioners.py index 8b92508e..c69b948d 100644 --- a/stable_audio_tools/models/conditioners.py +++ b/stable_audio_tools/models/conditioners.py @@ -365,11 +365,9 @@ def __init__(self, clap_ckpt_path, audio_model_type="HTSAT-base", enable_fusion=True, - project_out: bool = False, - finetune: bool = False): + project_out: bool = False): super().__init__(512, output_dim, project_out=project_out) - self.finetune = finetune device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Suppress logging from transformers