Upload hf_clap_embedder.py with huggingface_hub

hf_clap_embedder.py

@@ -0,0 +1,34 @@
+import torch, librosa, laion_clap, functools, warnings
+from huggingface_hub import hf_hub_download  # Added for robust HF downloading
+warnings.filterwarnings('ignore')
+
+original_load = torch.load
+torch.load = functools.partial(original_load, weights_only=False)
+original_load_state_dict = torch.nn.Module.load_state_dict
+
+def tolerant_load_state_dict(self, state_dict, strict=True, assign=False):
+    return original_load_state_dict(self, state_dict, strict=False, assign=assign)
+torch.nn.Module.load_state_dict = tolerant_load_state_dict
+
+# Initialize the CLAP module with fusion disabled as requested
+model = laion_clap.CLAP_Module(enable_fusion=False)
+
+# Securely download the weights through the official Hugging Face Hub API
+# This bypasses the broken wget/urllib external download logic entirely
+local_checkpoint = hf_hub_download(
+    repo_id="lukewys/laion_clap", 
+    filename="630k-audioset-best.pt"
+)
+
+# Load the locally cached file path directly into the model
+model.load_ckpt(local_checkpoint) 
+
+torch.load = original_load
+torch.nn.Module.load_state_dict = original_load_state_dict
+
+def get_clap_embedding(path):
+    audio_data, _ = librosa.load(path, sr=48000)
+    audio_data = audio_data.reshape(1, -1)
+    with torch.no_grad():
+        audio_embed = model.get_audio_embedding_from_data(x=audio_data)
+    return audio_embed.flatten()