Rename pad_idx to pad_value in Collater (#60)

src/seamless_communication/models/inference/translator.py

@@ -80,7 +80,7 @@ class Translator(nn.Module):
             dtype=dtype,
         )
         self.collate = Collater(
-            pad_idx=self.text_tokenizer.vocab_info.pad_idx, pad_to_multiple=2
+            pad_value=self.text_tokenizer.vocab_info.pad_idx, pad_to_multiple=2
         )
         # Load the vocoder.
         self.vocoder: Vocoder = self.load_model_for_inference(

src/seamless_communication/models/unit_extraction/unit_extraction.py

@@ -48,7 +48,7 @@ class UnitExtractor(nn.Module):
         self.model = Wav2Vec2LayerOutputModel(wav2vec2_model)
         self.device = device
         self.decode_audio = AudioDecoder(dtype=torch.float32, device=device)
-        self.collate = Collater(pad_idx=2, pad_to_multiple=2)
+        self.collate = Collater(pad_value=2, pad_to_multiple=2)
         self.kmeans_model = KmeansModel(kmeans_uri, device)
 
     @torch.inference_mode()

src/seamless_communication/models/unity/length_regulator.py

@@ -185,7 +185,7 @@ class VarianceAdaptor(Module):
         )
 
         # We need to apply the padding_mask again since we clamp by min_duration.
-        durations = apply_padding_mask(durations, padding_mask, fill_value=0)
+        durations = apply_padding_mask(durations, padding_mask, pad_value=0)
 
         # TODO: Implement pitch, energy predictors.
         # TODO: Implement GaussianUpsampling.