ltuncay commited on
Commit
3749aa0
·
verified ·
1 Parent(s): 47b2c06

fix: default to single-phase extraction without overlap

Browse files
Files changed (3) hide show
  1. README.md +4 -2
  2. config.json +1 -1
  3. configuration_audio.py +1 -1
README.md CHANGED
@@ -96,7 +96,7 @@ challenge's standardized ranking. These evaluations precede the Transformers exp
96
  | Training precision | bfloat16 mixed precision |
97
  | Training batch | 128 per GPU × 2 GPUs, DDP; effective batch 256 |
98
  | Exported weight dtype | float32 |
99
- | Extraction policy | `overlap50_two_phase` |
100
 
101
  ## Load the model
102
 
@@ -129,7 +129,7 @@ Resample to `processor.sampling_rate` and downmix stereo before preprocessing.
129
  The extractor performs padding only; spectrogram/convolution features are computed
130
  inside the model. Pass the returned sample attention mask for variable durations.
131
  Outputs include a frame attention mask and timestamps in milliseconds (-1 for padding).
132
- `pooler_output` uses the saved HEAR extraction preset, including phase-balanced pooling.
133
  Frame features average frequency patches; they are not the raw frequency-time ViT grid.
134
 
135
  ## Fine-tuning
@@ -187,3 +187,5 @@ BEST-RQ-2.2 is a variant in the BEST-RQ-2 family. Cite the family paper when usi
187
 
188
  The model weights and bundled implementation are released under the MIT license.
189
  See `LICENSE` and `CODE_LICENSE`.
 
 
 
96
  | Training precision | bfloat16 mixed precision |
97
  | Training batch | 128 per GPU × 2 GPUs, DDP; effective batch 256 |
98
  | Exported weight dtype | float32 |
99
+ | Extraction policy | `native_one_phase` |
100
 
101
  ## Load the model
102
 
 
129
  The extractor performs padding only; spectrogram/convolution features are computed
130
  inside the model. Pass the returned sample attention mask for variable durations.
131
  Outputs include a frame attention mask and timestamps in milliseconds (-1 for padding).
132
+ `pooler_output` uses the saved HEAR extraction preset, using a single phase with no context-window overlap by default.
133
  Frame features average frequency patches; they are not the raw frequency-time ViT grid.
134
 
135
  ## Fine-tuning
 
187
 
188
  The model weights and bundled implementation are released under the MIT license.
189
  See `LICENSE` and `CODE_LICENSE`.
190
+
191
+ Default inference uses `native_one_phase` (no context-window overlap and one encoding phase). Reported benchmark results retain their stated evaluation presets; to reproduce overlapping two-phase evaluation, set `model.config.extraction_preset = "overlap50_two_phase"`.
config.json CHANGED
@@ -27,7 +27,7 @@
27
  "residual_type": "standard",
28
  "rope_mode": "1d"
29
  },
30
- "extraction_preset": "overlap50_two_phase",
31
  "feature_encoder_kwargs": {},
32
  "hidden_size": 768,
33
  "model_target": "src.models.best_rq22_module.BestRQ22Module",
 
27
  "residual_type": "standard",
28
  "rope_mode": "1d"
29
  },
30
+ "extraction_preset": "native_one_phase",
31
  "feature_encoder_kwargs": {},
32
  "hidden_size": 768,
33
  "model_target": "src.models.best_rq22_module.BestRQ22Module",
configuration_audio.py CHANGED
@@ -45,7 +45,7 @@ class AudioEmbeddingConfig(PretrainedConfig):
45
  feature_encoder_kwargs: dict[str, Any] | None = None,
46
  sampling_rate: int = 16000,
47
  spectrogram_adjustment_mode: str = "pad",
48
- extraction_preset: str = "overlap50_two_phase",
49
  **kwargs: Any,
50
  ) -> None:
51
  super().__init__(**kwargs)
 
45
  feature_encoder_kwargs: dict[str, Any] | None = None,
46
  sampling_rate: int = 16000,
47
  spectrogram_adjustment_mode: str = "pad",
48
+ extraction_preset: str = "native_one_phase",
49
  **kwargs: Any,
50
  ) -> None:
51
  super().__init__(**kwargs)