justdata 1.0.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (106) hide show
  1. justdata/__init__.py +3 -0
  2. justdata/acoustic/__init__.py +381 -0
  3. justdata/acoustic/adapters.py +329 -0
  4. justdata/acoustic/augment/__init__.py +119 -0
  5. justdata/acoustic/augment/batch.py +813 -0
  6. justdata/acoustic/augment/configs.py +204 -0
  7. justdata/acoustic/augment/patchout.py +331 -0
  8. justdata/acoustic/augment/spectrogram.py +806 -0
  9. justdata/acoustic/augment/waveform.py +879 -0
  10. justdata/acoustic/batching.py +27 -0
  11. justdata/acoustic/channel.py +53 -0
  12. justdata/acoustic/compat/__init__.py +95 -0
  13. justdata/acoustic/compat/ast.py +626 -0
  14. justdata/acoustic/compat/ced.py +152 -0
  15. justdata/acoustic/compat/efficientat.py +243 -0
  16. justdata/acoustic/compat/panns.py +115 -0
  17. justdata/acoustic/compat/passt.py +191 -0
  18. justdata/acoustic/configs.py +519 -0
  19. justdata/acoustic/corruptions/__init__.py +33 -0
  20. justdata/acoustic/corruptions/codec.py +92 -0
  21. justdata/acoustic/corruptions/datasets.py +155 -0
  22. justdata/acoustic/corruptions/device.py +39 -0
  23. justdata/acoustic/corruptions/dynamics.py +55 -0
  24. justdata/acoustic/corruptions/filters.py +105 -0
  25. justdata/acoustic/corruptions/noise.py +173 -0
  26. justdata/acoustic/corruptions/registry.py +230 -0
  27. justdata/acoustic/corruptions/reverb.py +49 -0
  28. justdata/acoustic/corruptions/timefreq.py +127 -0
  29. justdata/acoustic/datasets.py +52 -0
  30. justdata/acoustic/dcase2025.py +814 -0
  31. justdata/acoustic/decoding.py +73 -0
  32. justdata/acoustic/eval_views.py +54 -0
  33. justdata/acoustic/evals.py +71 -0
  34. justdata/acoustic/filters.py +18 -0
  35. justdata/acoustic/frontends/__init__.py +37 -0
  36. justdata/acoustic/frontends/ast.py +116 -0
  37. justdata/acoustic/frontends/kaldi.py +34 -0
  38. justdata/acoustic/frontends/log.py +94 -0
  39. justdata/acoustic/frontends/mel.py +173 -0
  40. justdata/acoustic/frontends/mfcc.py +26 -0
  41. justdata/acoustic/frontends/pcen.py +26 -0
  42. justdata/acoustic/frontends/stft.py +174 -0
  43. justdata/acoustic/jax.py +23 -0
  44. justdata/acoustic/labels.py +169 -0
  45. justdata/acoustic/layouts.py +86 -0
  46. justdata/acoustic/metadata.py +152 -0
  47. justdata/acoustic/normalization.py +196 -0
  48. justdata/acoustic/pipelines.py +112 -0
  49. justdata/acoustic/postprocessing.py +328 -0
  50. justdata/acoustic/preprocessing.py +175 -0
  51. justdata/acoustic/presets.py +257 -0
  52. justdata/acoustic/registry.py +381 -0
  53. justdata/acoustic/resampling.py +218 -0
  54. justdata/acoustic/schema.py +51 -0
  55. justdata/acoustic/segment.py +369 -0
  56. justdata/acoustic/sources.py +503 -0
  57. justdata/acoustic/stages.py +59 -0
  58. justdata/acoustic/stats.py +50 -0
  59. justdata/acoustic/tasks.py +375 -0
  60. justdata/audio/__init__.py +20 -0
  61. justdata/core/__init__.py +74 -0
  62. justdata/core/adapters.py +46 -0
  63. justdata/core/augmentations.py +34 -0
  64. justdata/core/filters.py +94 -0
  65. justdata/core/finalization.py +173 -0
  66. justdata/core/label_mixing.py +129 -0
  67. justdata/core/loader.py +384 -0
  68. justdata/core/metadata.py +250 -0
  69. justdata/core/presets.py +155 -0
  70. justdata/core/registry.py +229 -0
  71. justdata/core/sources.py +87 -0
  72. justdata/core/stats.py +227 -0
  73. justdata/corruptions.py +52 -0
  74. justdata/py.typed +0 -0
  75. justdata/vision/__init__.py +17 -0
  76. justdata/vision/augmentations/__init__.py +31 -0
  77. justdata/vision/augmentations/auto.py +480 -0
  78. justdata/vision/augmentations/color.py +163 -0
  79. justdata/vision/augmentations/composed.py +156 -0
  80. justdata/vision/augmentations/geometric.py +302 -0
  81. justdata/vision/augmentations/mixing.py +232 -0
  82. justdata/vision/augmentations/registry.py +142 -0
  83. justdata/vision/corruptions/__init__.py +16 -0
  84. justdata/vision/corruptions/blur.py +65 -0
  85. justdata/vision/corruptions/digital.py +26 -0
  86. justdata/vision/corruptions/noise.py +20 -0
  87. justdata/vision/corruptions/registry.py +113 -0
  88. justdata/vision/corruptions/weather.py +100 -0
  89. justdata/vision/datasets.py +34 -0
  90. justdata/vision/encodings/__init__.py +1 -0
  91. justdata/vision/encodings/heatmaps.py +196 -0
  92. justdata/vision/minic.py +69 -0
  93. justdata/vision/pipelines.py +62 -0
  94. justdata/vision/presets.py +509 -0
  95. justdata/vision/sources.py +1080 -0
  96. justdata/vision/stages.py +271 -0
  97. justdata/vision/tasks/__init__.py +1 -0
  98. justdata/vision/tasks/classification.py +321 -0
  99. justdata/vision/tasks/depth.py +25 -0
  100. justdata/vision/tasks/segmentation.py +194 -0
  101. justdata/vision/transforms.py +113 -0
  102. justdata/vision/utils.py +741 -0
  103. justdata-1.0.0.dist-info/METADATA +860 -0
  104. justdata-1.0.0.dist-info/RECORD +106 -0
  105. justdata-1.0.0.dist-info/WHEEL +4 -0
  106. justdata-1.0.0.dist-info/licenses/LICENSE +23 -0
@@ -0,0 +1,152 @@
1
+ from __future__ import annotations
2
+
3
+ from justdata.acoustic.compat.efficientat import DCASE_CLASSES
4
+ from justdata.acoustic.configs import (
5
+ AudioPreprocessConfig,
6
+ AudioPreset,
7
+ FrontendConfig,
8
+ LabelTransformConfig,
9
+ LogCompressionConfig,
10
+ MelConfig,
11
+ STFTConfig,
12
+ SegmentStrategyConfig,
13
+ )
14
+
15
+
16
+ def ced_frontend() -> FrontendConfig:
17
+ return FrontendConfig(
18
+ name="logmel",
19
+ stft=STFTConfig(
20
+ sample_rate=16000,
21
+ n_fft=512,
22
+ win_length=512,
23
+ hop_length=160,
24
+ window="hann",
25
+ center=True,
26
+ power=2.0,
27
+ ),
28
+ mel=MelConfig(
29
+ n_mels=64,
30
+ f_min=0.0,
31
+ f_max=8000.0,
32
+ mel_scale="htk",
33
+ mel_norm="none",
34
+ filterbank_impl="torchaudio",
35
+ ),
36
+ log=LogCompressionConfig(kind="db", top_db=120.0),
37
+ )
38
+
39
+
40
+ def _preprocess() -> AudioPreprocessConfig:
41
+ return AudioPreprocessConfig(
42
+ target_sample_rate=16000,
43
+ channel_strategy="mono_mean",
44
+ normalize_waveform="none",
45
+ )
46
+
47
+
48
+ def _metadata(model_size: str, *, source_duration: float | None = None) -> dict:
49
+ metadata = {
50
+ "preset_version": 2,
51
+ "model_family": "ced",
52
+ "model_size": model_size,
53
+ "frontend_contract": "ced-hf-melspectrogram-v2",
54
+ "compatibility_status": "frontend_golden",
55
+ }
56
+ if source_duration is not None:
57
+ metadata["source_duration"] = source_duration
58
+ return metadata
59
+
60
+
61
+ def _audioset_label_transform() -> LabelTransformConfig:
62
+ return LabelTransformConfig(mode="multi_hot", num_classes=527)
63
+
64
+
65
+ def _dcase_label_transform() -> LabelTransformConfig:
66
+ return LabelTransformConfig(
67
+ mode="index",
68
+ num_classes=len(DCASE_CLASSES),
69
+ class_names=DCASE_CLASSES,
70
+ )
71
+
72
+
73
+ def _preset(
74
+ *,
75
+ name: str,
76
+ model_size: str,
77
+ input_duration: float,
78
+ segment: SegmentStrategyConfig,
79
+ label_transform: LabelTransformConfig,
80
+ metadata: dict | None = None,
81
+ ) -> AudioPreset:
82
+ return AudioPreset(
83
+ name=name,
84
+ input_duration=input_duration,
85
+ target_sample_rate=16000,
86
+ preprocess=_preprocess(),
87
+ segment=segment,
88
+ frontend=ced_frontend(),
89
+ layout="btf",
90
+ label_transform=label_transform,
91
+ metadata=metadata or _metadata(model_size),
92
+ )
93
+
94
+
95
+ def _audioset_preset(name: str, model_size: str) -> AudioPreset:
96
+ return _preset(
97
+ name=name,
98
+ model_size=model_size,
99
+ input_duration=10.0,
100
+ segment=SegmentStrategyConfig(
101
+ clip_duration=10.0,
102
+ train_mode="random_crop",
103
+ eval_mode="center_crop",
104
+ pad_mode="zero",
105
+ pad_position="right",
106
+ ),
107
+ label_transform=_audioset_label_transform(),
108
+ )
109
+
110
+
111
+ def ced_tiny_16k_logmel64() -> AudioPreset:
112
+ return _audioset_preset("ced_tiny_16k_logmel64", "tiny")
113
+
114
+
115
+ def ced_mini_16k_logmel64() -> AudioPreset:
116
+ return _audioset_preset("ced_mini_16k_logmel64", "mini")
117
+
118
+
119
+ def ced_small_16k_logmel64() -> AudioPreset:
120
+ return _audioset_preset("ced_small_16k_logmel64", "small")
121
+
122
+
123
+ def ced_base_16k_logmel64() -> AudioPreset:
124
+ return _audioset_preset("ced_base_16k_logmel64", "base")
125
+
126
+
127
+ def dcase2025_task1_ced_16k_1s() -> AudioPreset:
128
+ return _preset(
129
+ name="dcase2025_task1_ced_16k_1s",
130
+ model_size="base",
131
+ input_duration=1.0,
132
+ segment=SegmentStrategyConfig(
133
+ clip_duration=1.0,
134
+ train_mode="random_crop",
135
+ eval_mode="center_crop",
136
+ pad_mode="zero",
137
+ pad_position="right",
138
+ duration_policy="keep_1s",
139
+ ),
140
+ label_transform=_dcase_label_transform(),
141
+ metadata=_metadata("base", source_duration=1.0),
142
+ )
143
+
144
+
145
+ __all__ = [
146
+ "ced_base_16k_logmel64",
147
+ "ced_frontend",
148
+ "ced_mini_16k_logmel64",
149
+ "ced_small_16k_logmel64",
150
+ "ced_tiny_16k_logmel64",
151
+ "dcase2025_task1_ced_16k_1s",
152
+ ]
@@ -0,0 +1,243 @@
1
+ from __future__ import annotations
2
+
3
+ import tensorflow as tf
4
+
5
+ from justdata.acoustic.configs import (
6
+ AudioPreprocessConfig,
7
+ AudioPreset,
8
+ FeatureNormConfig,
9
+ FrontendConfig,
10
+ LabelTransformConfig,
11
+ LogCompressionConfig,
12
+ MelConfig,
13
+ STFTConfig,
14
+ SegmentStrategyConfig,
15
+ )
16
+ from justdata.acoustic.frontends.log import compress_log
17
+ from justdata.acoustic.frontends.mel import mel_weight_matrix
18
+ from justdata.acoustic.frontends.stft import stft_power_spectrogram
19
+ from justdata.acoustic.normalization import apply_feature_normalization
20
+ from justdata.acoustic.registry import register_audio_frontend
21
+
22
+
23
+ DCASE_CLASSES = (
24
+ "airport",
25
+ "shopping_mall",
26
+ "metro_station",
27
+ "street_pedestrian",
28
+ "public_square",
29
+ "street_traffic",
30
+ "tram",
31
+ "bus",
32
+ "metro",
33
+ "park",
34
+ )
35
+
36
+
37
+ def efficientat_frontend() -> FrontendConfig:
38
+ return FrontendConfig(
39
+ name="efficientat_logmel",
40
+ stft=STFTConfig(
41
+ sample_rate=32000,
42
+ n_fft=1024,
43
+ win_length=800,
44
+ hop_length=320,
45
+ window="hann",
46
+ window_periodic=False,
47
+ center=True,
48
+ power=2.0,
49
+ normalized=False,
50
+ ),
51
+ mel=MelConfig(
52
+ n_mels=128,
53
+ f_min=0.0,
54
+ f_max=15000.0,
55
+ mel_scale="htk",
56
+ mel_norm="none",
57
+ filterbank_impl="efficientat_kaldi",
58
+ ),
59
+ log=LogCompressionConfig(kind="log", log_offset=1e-5),
60
+ norm=FeatureNormConfig(kind="affine", scale=0.2, bias=0.9),
61
+ )
62
+
63
+
64
+ @register_audio_frontend("efficientat_logmel")
65
+ def efficientat_logmel(audio: tf.Tensor, config: FrontendConfig | dict) -> tf.Tensor:
66
+ config = FrontendConfig.from_dict(config)
67
+ if config.stft is None or config.mel is None:
68
+ raise ValueError("EfficientAT frontend requires STFT and mel configs")
69
+ spectrogram = stft_power_spectrogram(
70
+ audio,
71
+ config.stft,
72
+ preemphasis=0.97,
73
+ frame_length=config.stft.n_fft,
74
+ )
75
+ mel = tf.einsum("tfc,fm->tmc", spectrogram, mel_weight_matrix(config))
76
+ features = compress_log(mel, config.log)
77
+ return apply_feature_normalization(features, config.norm)
78
+
79
+
80
+ def _preprocess() -> AudioPreprocessConfig:
81
+ return AudioPreprocessConfig(
82
+ target_sample_rate=32000,
83
+ channel_strategy="mono_mean",
84
+ normalize_waveform="none",
85
+ )
86
+
87
+
88
+ def _metadata(model_family: str, *, source_duration: float | None = None) -> dict:
89
+ metadata = {
90
+ "preset_version": 2,
91
+ "model_family": model_family,
92
+ "frontend_contract": "efficientat-logmel-v1",
93
+ "compatibility_status": "frontend_golden",
94
+ }
95
+ if source_duration is not None:
96
+ metadata["source_duration"] = source_duration
97
+ return metadata
98
+
99
+
100
+ def _audioset_label_transform() -> LabelTransformConfig:
101
+ return LabelTransformConfig(mode="multi_hot", num_classes=527)
102
+
103
+
104
+ def _dcase_label_transform() -> LabelTransformConfig:
105
+ return LabelTransformConfig(
106
+ mode="index",
107
+ num_classes=len(DCASE_CLASSES),
108
+ class_names=DCASE_CLASSES,
109
+ )
110
+
111
+
112
+ def _preset(
113
+ *,
114
+ name: str,
115
+ model_family: str,
116
+ input_duration: float,
117
+ segment: SegmentStrategyConfig,
118
+ label_transform: LabelTransformConfig,
119
+ metadata: dict | None = None,
120
+ ) -> AudioPreset:
121
+ return AudioPreset(
122
+ name=name,
123
+ input_duration=input_duration,
124
+ target_sample_rate=32000,
125
+ preprocess=_preprocess(),
126
+ segment=segment,
127
+ frontend=efficientat_frontend(),
128
+ layout="bcft",
129
+ static_shape=(1, 128, round(input_duration * 100)),
130
+ label_transform=label_transform,
131
+ metadata=metadata or _metadata(model_family),
132
+ )
133
+
134
+
135
+ def _default_segment(duration: float) -> SegmentStrategyConfig:
136
+ return SegmentStrategyConfig(
137
+ clip_duration=duration,
138
+ train_mode="random_crop",
139
+ eval_mode="center_crop",
140
+ pad_mode="zero",
141
+ pad_position="right",
142
+ )
143
+
144
+
145
+ def efficientat_32k_10s_logmel128() -> AudioPreset:
146
+ return _preset(
147
+ name="efficientat_32k_10s_logmel128",
148
+ model_family="efficientat",
149
+ input_duration=10.0,
150
+ segment=_default_segment(10.0),
151
+ label_transform=_audioset_label_transform(),
152
+ )
153
+
154
+
155
+ def dymn_32k_10s_logmel128() -> AudioPreset:
156
+ return _preset(
157
+ name="dymn_32k_10s_logmel128",
158
+ model_family="dymn",
159
+ input_duration=10.0,
160
+ segment=_default_segment(10.0),
161
+ label_transform=_audioset_label_transform(),
162
+ )
163
+
164
+
165
+ def _dcase_keep_1s(name: str, model_family: str) -> AudioPreset:
166
+ return _preset(
167
+ name=name,
168
+ model_family=model_family,
169
+ input_duration=1.0,
170
+ segment=SegmentStrategyConfig(
171
+ clip_duration=1.0,
172
+ train_mode="random_crop",
173
+ eval_mode="center_crop",
174
+ pad_mode="zero",
175
+ pad_position="right",
176
+ duration_policy="keep_1s",
177
+ ),
178
+ label_transform=_dcase_label_transform(),
179
+ metadata=_metadata(model_family, source_duration=1.0),
180
+ )
181
+
182
+
183
+ def dcase2025_task1_efficientat_32k_1s() -> AudioPreset:
184
+ return _dcase_keep_1s(
185
+ "dcase2025_task1_efficientat_32k_1s",
186
+ "efficientat",
187
+ )
188
+
189
+
190
+ def dcase2025_task1_dymn_32k_1s() -> AudioPreset:
191
+ return _dcase_keep_1s(
192
+ "dcase2025_task1_dymn_32k_1s",
193
+ "dymn",
194
+ )
195
+
196
+
197
+ def _dcase_10s(name: str, *, duration_policy: str, pad_mode: str) -> AudioPreset:
198
+ metadata = _metadata("efficientat", source_duration=1.0)
199
+ metadata["model_input_duration"] = 10.0
200
+ return _preset(
201
+ name=name,
202
+ model_family="efficientat",
203
+ input_duration=10.0,
204
+ segment=SegmentStrategyConfig(
205
+ clip_duration=10.0,
206
+ train_mode="pad_or_crop",
207
+ eval_mode="center_crop",
208
+ pad_mode=pad_mode,
209
+ pad_position="right",
210
+ duration_policy=duration_policy,
211
+ ),
212
+ label_transform=_dcase_label_transform(),
213
+ metadata=metadata,
214
+ )
215
+
216
+
217
+ def dcase2025_task1_efficientat_32k_1s_zero_pad_to_10s() -> AudioPreset:
218
+ return _dcase_10s(
219
+ "dcase2025_task1_efficientat_32k_1s_zero_pad_to_10s",
220
+ duration_policy="pad_to_model_duration",
221
+ pad_mode="zero",
222
+ )
223
+
224
+
225
+ def dcase2025_task1_efficientat_32k_1s_repeat_to_10s() -> AudioPreset:
226
+ return _dcase_10s(
227
+ "dcase2025_task1_efficientat_32k_1s_repeat_to_10s",
228
+ duration_policy="repeat_pad_to_model_duration",
229
+ pad_mode="repeat",
230
+ )
231
+
232
+
233
+ __all__ = [
234
+ "DCASE_CLASSES",
235
+ "dcase2025_task1_dymn_32k_1s",
236
+ "dcase2025_task1_efficientat_32k_1s",
237
+ "dcase2025_task1_efficientat_32k_1s_repeat_to_10s",
238
+ "dcase2025_task1_efficientat_32k_1s_zero_pad_to_10s",
239
+ "dymn_32k_10s_logmel128",
240
+ "efficientat_32k_10s_logmel128",
241
+ "efficientat_frontend",
242
+ "efficientat_logmel",
243
+ ]
@@ -0,0 +1,115 @@
1
+ from __future__ import annotations
2
+
3
+ from justdata.acoustic.configs import (
4
+ AudioPreprocessConfig,
5
+ AudioPreset,
6
+ FrontendConfig,
7
+ LabelTransformConfig,
8
+ LogCompressionConfig,
9
+ MelConfig,
10
+ STFTConfig,
11
+ SegmentStrategyConfig,
12
+ )
13
+
14
+
15
+ def panns_frontend(
16
+ *,
17
+ sample_rate: int,
18
+ window_size: int,
19
+ hop_size: int,
20
+ mel_bins: int,
21
+ fmin: float,
22
+ fmax: float,
23
+ ) -> FrontendConfig:
24
+ return FrontendConfig(
25
+ name="logmel",
26
+ stft=STFTConfig(
27
+ sample_rate=sample_rate,
28
+ n_fft=window_size,
29
+ win_length=window_size,
30
+ hop_length=hop_size,
31
+ center=True,
32
+ power=2.0,
33
+ ),
34
+ mel=MelConfig(
35
+ n_mels=mel_bins,
36
+ f_min=fmin,
37
+ f_max=fmax,
38
+ mel_scale="slaney",
39
+ mel_norm="slaney",
40
+ filterbank_impl="librosa",
41
+ ),
42
+ log=LogCompressionConfig(kind="db"),
43
+ )
44
+
45
+
46
+ def _preset(
47
+ *,
48
+ name: str,
49
+ sample_rate: int,
50
+ frontend: FrontendConfig,
51
+ ) -> AudioPreset:
52
+ return AudioPreset(
53
+ name=name,
54
+ input_duration=10.0,
55
+ target_sample_rate=sample_rate,
56
+ preprocess=AudioPreprocessConfig(
57
+ target_sample_rate=sample_rate,
58
+ channel_strategy="mono_mean",
59
+ normalize_waveform="none",
60
+ ),
61
+ segment=SegmentStrategyConfig(
62
+ clip_duration=10.0,
63
+ train_mode="random_crop",
64
+ eval_mode="center_crop",
65
+ pad_mode="zero",
66
+ pad_position="right",
67
+ ),
68
+ frontend=frontend,
69
+ layout="btf",
70
+ label_transform=LabelTransformConfig(mode="multi_hot", num_classes=527),
71
+ metadata={
72
+ "preset_version": 2,
73
+ "model_family": "panns",
74
+ "model_name": "cnn14",
75
+ "frontend_contract": "panns-cnn14-logmel-v1",
76
+ "compatibility_status": "frontend_golden",
77
+ },
78
+ )
79
+
80
+
81
+ def panns_cnn14_32k_10s_logmel64() -> AudioPreset:
82
+ return _preset(
83
+ name="panns_cnn14_32k_10s_logmel64",
84
+ sample_rate=32000,
85
+ frontend=panns_frontend(
86
+ sample_rate=32000,
87
+ window_size=1024,
88
+ hop_size=320,
89
+ mel_bins=64,
90
+ fmin=50.0,
91
+ fmax=14000.0,
92
+ ),
93
+ )
94
+
95
+
96
+ def panns_cnn14_16k_10s_logmel64() -> AudioPreset:
97
+ return _preset(
98
+ name="panns_cnn14_16k_10s_logmel64",
99
+ sample_rate=16000,
100
+ frontend=panns_frontend(
101
+ sample_rate=16000,
102
+ window_size=512,
103
+ hop_size=160,
104
+ mel_bins=64,
105
+ fmin=50.0,
106
+ fmax=8000.0,
107
+ ),
108
+ )
109
+
110
+
111
+ __all__ = [
112
+ "panns_cnn14_16k_10s_logmel64",
113
+ "panns_cnn14_32k_10s_logmel64",
114
+ "panns_frontend",
115
+ ]
@@ -0,0 +1,191 @@
1
+ from __future__ import annotations
2
+
3
+ from justdata.acoustic.compat.efficientat import DCASE_CLASSES
4
+ from justdata.acoustic.configs import (
5
+ AudioPreprocessConfig,
6
+ AudioPreset,
7
+ FeatureNormConfig,
8
+ FrontendConfig,
9
+ LabelTransformConfig,
10
+ LogCompressionConfig,
11
+ MelConfig,
12
+ STFTConfig,
13
+ SegmentStrategyConfig,
14
+ )
15
+
16
+
17
+ PATCH_GRID = {
18
+ "input_fdim": 128,
19
+ "input_tdim": 998,
20
+ "fstride": 10,
21
+ "tstride": 10,
22
+ }
23
+
24
+ PATCHOUT = {
25
+ "structured_frequency": 4,
26
+ "structured_time": 40,
27
+ "unstructured": 0,
28
+ }
29
+
30
+
31
+ def passt_frontend() -> FrontendConfig:
32
+ return FrontendConfig(
33
+ name="efficientat_logmel",
34
+ stft=STFTConfig(
35
+ sample_rate=32000,
36
+ n_fft=1024,
37
+ win_length=800,
38
+ hop_length=320,
39
+ window_periodic=False,
40
+ center=True,
41
+ power=2.0,
42
+ ),
43
+ mel=MelConfig(
44
+ n_mels=128,
45
+ f_min=0.0,
46
+ f_max=15000.0,
47
+ mel_scale="htk",
48
+ mel_norm="none",
49
+ filterbank_impl="efficientat_kaldi",
50
+ ),
51
+ log=LogCompressionConfig(kind="log", log_offset=1e-5),
52
+ norm=FeatureNormConfig(kind="affine", scale=0.2, bias=0.9),
53
+ )
54
+
55
+
56
+ def _preprocess() -> AudioPreprocessConfig:
57
+ return AudioPreprocessConfig(
58
+ target_sample_rate=32000,
59
+ channel_strategy="mono_mean",
60
+ normalize_waveform="none",
61
+ )
62
+
63
+
64
+ def _metadata(*, source_duration: float | None = None) -> dict:
65
+ metadata = {
66
+ "preset_version": 2,
67
+ "model_family": "passt",
68
+ "frontend_contract": "passt-logmel-v1",
69
+ "compatibility_status": "frontend_golden",
70
+ "patch_grid": PATCH_GRID,
71
+ }
72
+ if source_duration is not None:
73
+ metadata["source_duration"] = source_duration
74
+ return metadata
75
+
76
+
77
+ def _audioset_label_transform() -> LabelTransformConfig:
78
+ return LabelTransformConfig(mode="multi_hot", num_classes=527)
79
+
80
+
81
+ def _dcase_label_transform() -> LabelTransformConfig:
82
+ return LabelTransformConfig(
83
+ mode="index",
84
+ num_classes=len(DCASE_CLASSES),
85
+ class_names=DCASE_CLASSES,
86
+ )
87
+
88
+
89
+ def _preset(
90
+ *,
91
+ name: str,
92
+ input_duration: float,
93
+ segment: SegmentStrategyConfig,
94
+ label_transform: LabelTransformConfig,
95
+ train_augment: dict | None = None,
96
+ metadata: dict | None = None,
97
+ ) -> AudioPreset:
98
+ return AudioPreset(
99
+ name=name,
100
+ input_duration=input_duration,
101
+ target_sample_rate=32000,
102
+ preprocess=_preprocess(),
103
+ segment=segment,
104
+ frontend=passt_frontend(),
105
+ layout="bcft",
106
+ static_shape=(1, 128, round(input_duration * 100)),
107
+ label_transform=label_transform,
108
+ train_augment=train_augment or {},
109
+ metadata=metadata or _metadata(),
110
+ )
111
+
112
+
113
+ def passt_32k_10s_logmel128() -> AudioPreset:
114
+ return _preset(
115
+ name="passt_32k_10s_logmel128",
116
+ input_duration=10.0,
117
+ segment=SegmentStrategyConfig(
118
+ clip_duration=10.0,
119
+ train_mode="random_crop",
120
+ eval_mode="center_crop",
121
+ pad_mode="zero",
122
+ pad_position="right",
123
+ ),
124
+ label_transform=_audioset_label_transform(),
125
+ train_augment={"patchout": PATCHOUT},
126
+ )
127
+
128
+
129
+ def dcase2025_task1_passt_32k_1s() -> AudioPreset:
130
+ return _preset(
131
+ name="dcase2025_task1_passt_32k_1s",
132
+ input_duration=1.0,
133
+ segment=SegmentStrategyConfig(
134
+ clip_duration=1.0,
135
+ train_mode="random_crop",
136
+ eval_mode="center_crop",
137
+ pad_mode="zero",
138
+ pad_position="right",
139
+ duration_policy="keep_1s",
140
+ ),
141
+ label_transform=_dcase_label_transform(),
142
+ train_augment={"patchout": PATCHOUT},
143
+ metadata=_metadata(source_duration=1.0),
144
+ )
145
+
146
+
147
+ def _dcase_10s(name: str, *, duration_policy: str, pad_mode: str) -> AudioPreset:
148
+ metadata = _metadata(source_duration=1.0)
149
+ metadata["model_input_duration"] = 10.0
150
+ return _preset(
151
+ name=name,
152
+ input_duration=10.0,
153
+ segment=SegmentStrategyConfig(
154
+ clip_duration=10.0,
155
+ train_mode="pad_or_crop",
156
+ eval_mode="center_crop",
157
+ pad_mode=pad_mode,
158
+ pad_position="right",
159
+ duration_policy=duration_policy,
160
+ ),
161
+ label_transform=_dcase_label_transform(),
162
+ train_augment={"patchout": PATCHOUT},
163
+ metadata=metadata,
164
+ )
165
+
166
+
167
+ def dcase2025_task1_passt_32k_1s_zero_pad_to_10s() -> AudioPreset:
168
+ return _dcase_10s(
169
+ "dcase2025_task1_passt_32k_1s_zero_pad_to_10s",
170
+ duration_policy="pad_to_model_duration",
171
+ pad_mode="zero",
172
+ )
173
+
174
+
175
+ def dcase2025_task1_passt_32k_1s_repeat_to_10s() -> AudioPreset:
176
+ return _dcase_10s(
177
+ "dcase2025_task1_passt_32k_1s_repeat_to_10s",
178
+ duration_policy="repeat_pad_to_model_duration",
179
+ pad_mode="repeat",
180
+ )
181
+
182
+
183
+ __all__ = [
184
+ "PATCH_GRID",
185
+ "PATCHOUT",
186
+ "dcase2025_task1_passt_32k_1s",
187
+ "dcase2025_task1_passt_32k_1s_repeat_to_10s",
188
+ "dcase2025_task1_passt_32k_1s_zero_pad_to_10s",
189
+ "passt_32k_10s_logmel128",
190
+ "passt_frontend",
191
+ ]