Upload configuration_prosody.py with huggingface_hub

configuration_prosody.py

@@ -0,0 +1,68 @@
+"""Configuration for the prosody-to-text pipeline."""
+
+from __future__ import annotations
+
+from transformers import PretrainedConfig
+
+
+class ProsodyToTextConfig(PretrainedConfig):
+    """Everything needed to rebuild the pipeline.
+
+    The mel contract is the part that matters. `mel_bins` and `length_ratio`
+    must match what the conformer was trained with: the loader falls back to
+    defaults when they are absent, and a wrong `length_ratio` mis-times the
+    mel against the audio without failing, which is the worst way for it to
+    be wrong.
+    """
+
+    model_type = "prosody_to_text"
+
+    def __init__(
+        self,
+        # --- the conformer's mel contract -------------------------------
+        mel_bins: int = 12,
+        length_ratio: float = 0.5,
+        conformer_subfolder: str = "conformer",
+        # --- feature extraction -----------------------------------------
+        # 200 Hz, and the physical-range normalisation the conformer was
+        # trained with. These mirror the training preprocessor; changing one
+        # without the others silently moves the model off its own
+        # distribution.
+        feature_rate_hz: float = 200.0,
+        f0_scale: float = 500.0,
+        intensity_scale: float = 100.0,
+        # --- the decoder -------------------------------------------------
+        whisper_subfolder: str = "whisper",
+        tokenizer_subfolder: str = "tokenizer",
+        # Ten, matching the published inference runs. Wider was tried and
+        # made the transcriptions worse: on a signal this weak a wider beam
+        # searches harder for the highest-likelihood sequence, which is the
+        # caption boilerplate the checkpoint overlearned.
+        num_beams: int = 10,
+        language: str = "en",
+        no_repeat_ngram_size: int = 3,
+        max_new_tokens: int = 225,
+        # On, matching the dataset runs the published numbers come from: ban
+        # numerals, and take the token budget from the utterance's own
+        # duration rather than a flat limit. Set False to decode
+        # unconstrained, which is what the drawing demo does.
+        constrain: bool = True,
+        **kwargs,
+    ) -> None:
+        self.mel_bins = mel_bins
+        self.length_ratio = length_ratio
+        self.conformer_subfolder = conformer_subfolder
+
+        self.feature_rate_hz = feature_rate_hz
+        self.f0_scale = f0_scale
+        self.intensity_scale = intensity_scale
+
+        self.whisper_subfolder = whisper_subfolder
+        self.tokenizer_subfolder = tokenizer_subfolder
+        self.num_beams = num_beams
+        self.language = language
+        self.no_repeat_ngram_size = no_repeat_ngram_size
+        self.max_new_tokens = max_new_tokens
+        self.constrain = constrain
+
+        super().__init__(**kwargs)