ltuncay commited on
Commit
bba843d
·
verified ·
1 Parent(s): 58910d4

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
@@ -100,7 +100,7 @@ challenge's standardized ranking. These evaluations precede the Transformers exp
100
  | Masking ratio | 40–60% |
101
  | Saved training step | 200,000, verified from checkpoint metadata |
102
  | Exported weight dtype | float32 |
103
- | Extraction policy | `overlap50_two_phase` |
104
 
105
  ## Load the model
106
 
@@ -133,7 +133,7 @@ Resample to `processor.sampling_rate` and downmix stereo before preprocessing.
133
  The extractor performs padding only; spectrogram/convolution features are computed
134
  inside the model. Pass the returned sample attention mask for variable durations.
135
  Outputs include a frame attention mask and timestamps in milliseconds (-1 for padding).
136
- `pooler_output` uses the saved HEAR extraction preset, including phase-balanced pooling.
137
  Frame features average frequency patches; they are not the raw frequency-time ViT grid.
138
 
139
  ## Fine-tuning
@@ -194,3 +194,5 @@ BEST-RQ-ViT is the ViT baseline in the BEST-RQ-2 study. For that comparison, cit
194
 
195
  The model weights and bundled implementation are released under the MIT license.
196
  See `LICENSE` and `CODE_LICENSE`.
 
 
 
100
  | Masking ratio | 40–60% |
101
  | Saved training step | 200,000, verified from checkpoint metadata |
102
  | Exported weight dtype | float32 |
103
+ | Extraction policy | `native_one_phase` |
104
 
105
  ## Load the model
106
 
 
133
  The extractor performs padding only; spectrogram/convolution features are computed
134
  inside the model. Pass the returned sample attention mask for variable durations.
135
  Outputs include a frame attention mask and timestamps in milliseconds (-1 for padding).
136
+ `pooler_output` uses the saved HEAR extraction preset, using a single phase with no context-window overlap by default.
137
  Frame features average frequency patches; they are not the raw frequency-time ViT grid.
138
 
139
  ## Fine-tuning
 
194
 
195
  The model weights and bundled implementation are released under the MIT license.
196
  See `LICENSE` and `CODE_LICENSE`.
197
+
198
+ 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
@@ -19,7 +19,7 @@
19
  "pos_embed_type": "sincos",
20
  "qkv_bias": true
21
  },
22
- "extraction_preset": "overlap50_two_phase",
23
  "feature_encoder_kwargs": {},
24
  "hidden_size": 768,
25
  "model_target": "src.models.best_rq_module.BestRQModule",
 
19
  "pos_embed_type": "sincos",
20
  "qkv_bias": true
21
  },
22
+ "extraction_preset": "native_one_phase",
23
  "feature_encoder_kwargs": {},
24
  "hidden_size": 768,
25
  "model_target": "src.models.best_rq_module.BestRQModule",
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)