Upload ear/semantic.py with huggingface_hub

ear/semantic.py

@@ -71,6 +71,21 @@ BANK: dict[str, list[str]] = {
 _FLAT: list[tuple[str, str]] = [(g, t) for g, items in BANK.items() for t in items]
 
 
+def _features(raw, projection):
+    """Normalise CLAP's feature output across transformers versions.
+
+    4.x returns the projected tensor directly. 5.x returns a
+    BaseModelOutputWithPooling, so the projection has to be applied here —
+    which is exactly what 4.x did internally.
+    """
+    if hasattr(raw, "shape"):
+        return raw
+    pooled = getattr(raw, "pooler_output", None)
+    if pooled is None:
+        pooled = raw.last_hidden_state[:, 0]
+    return projection(pooled)
+
+
 class _Semantic:
     def __init__(self) -> None:
         self.ready = False
@@ -96,7 +111,7 @@ class _Semantic:
             texts = [t for _, t in _FLAT]
             with torch.no_grad():
                 inputs = processor(text=texts, return_tensors="pt", padding=True)
-                emb = model.get_text_features(**inputs)
+                emb = _features(model.get_text_features(**inputs), model.text_projection)
                 emb = emb / emb.norm(dim=-1, keepdim=True)
 
             self._model, self._processor, self._text_emb = model, processor, emb
@@ -138,7 +153,8 @@ class _Semantic:
             except TypeError:
                 inputs = self._processor(audios=clip, sampling_rate=CLAP_SR,
                                          return_tensors="pt")
-            audio_emb = self._model.get_audio_features(**inputs)
+            audio_emb = _features(self._model.get_audio_features(**inputs),
+                                  self._model.audio_projection)
             audio_emb = audio_emb / audio_emb.norm(dim=-1, keepdim=True)
             sims = (audio_emb @ self._text_emb.T).squeeze(0).cpu().numpy()