sonusai 0.15.5__tar.gz → 0.15.8__tar.gz

{sonusai-0.15.5 → sonusai-0.15.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sonusai
-Version: 0.15.5
+Version: 0.15.8
 Summary: Framework for building deep neural network models for sound, speech, and voice AI
 Home-page: https://aaware.com
 License: GPL-3.0-only
@@ -16,21 +16,21 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Requires-Dist: PyYAML (>=6.0.1,<7.0.0)
 Requires-Dist: aixplain (>=0.2.6,<0.3.0)
+Requires-Dist: ctranslate2 (==4.1.0)
 Requires-Dist: dataclasses-json (>=0.6.1,<0.7.0)
 Requires-Dist: deepgram-sdk (>=3.0.0,<4.0.0)
 Requires-Dist: docopt (>=0.6.2,<0.7.0)
-Requires-Dist: faster-whisper (>=0.10.0,<0.11.0)
-Requires-Dist: ffmpeg (>=1.4,<2.0)
-Requires-Dist: greenlet (>=3.0.1,<4.0.0)
-Requires-Dist: grpcio (==1.60.0)
+Requires-Dist: faster-whisper (>=1.0.1,<2.0.0)
+Requires-Dist: h5py (>=3.11.0,<4.0.0)
 Requires-Dist: jiwer (>=3.0.3,<4.0.0)
-Requires-Dist: keras-tuner (>=1.4.5,<2.0.0)
+Requires-Dist: keras (>=3.1.1,<4.0.0)
+Requires-Dist: keras-tuner (>=1.4.7,<2.0.0)
 Requires-Dist: librosa (>=0.10.1,<0.11.0)
-Requires-Dist: lightning (>=2.1.0,<3.0.0)
+Requires-Dist: lightning (>=2.2,<2.3)
 Requires-Dist: matplotlib (>=3.8.0,<4.0.0)
-Requires-Dist: onnx (==1.14.1)
+Requires-Dist: onnx (>=1.14.1,<2.0.0)
 Requires-Dist: onnxruntime (>=1.16.1,<2.0.0)
-Requires-Dist: paho-mqtt (>=1.6.1,<2.0.0)
+Requires-Dist: paho-mqtt (>=2.0.0,<3.0.0)
 Requires-Dist: pandas (>=2.1.1,<3.0.0)
 Requires-Dist: pesq (>=0.0.4,<0.0.5)
 Requires-Dist: pyaaware (>=1.5.3,<2.0.0)
@@ -38,18 +38,14 @@ Requires-Dist: pydub (>=0.25.1,<0.26.0)
 Requires-Dist: pystoi (>=0.4.0,<0.5.0)
 Requires-Dist: python-magic (>=0.4.27,<0.5.0)
 Requires-Dist: requests (>=2.31.0,<3.0.0)
-Requires-Dist: resampy (>=0.4.2,<0.5.0)
-Requires-Dist: samplerate (>=0.1.0,<0.2.0)
-Requires-Dist: scikit-learn (>=1.3.1,<2.0.0)
-Requires-Dist: sh (>=2.0.6,<3.0.0)
+Requires-Dist: samplerate (>=0.2.1,<0.3.0)
 Requires-Dist: soundfile (>=0.12.1,<0.13.0)
 Requires-Dist: sox (>=1.4.1,<2.0.0)
-Requires-Dist: speechrecognition (>=3.10.0,<4.0.0)
+Requires-Dist: speechrecognition (>=3.10.1,<4.0.0)
 Requires-Dist: tensorflow (>=2.15.0,<3.0.0)
-Requires-Dist: tensorflow-addons (>=0.23.0,<0.24.0)
 Requires-Dist: tf2onnx (>=1.15.1,<2.0.0)
-Requires-Dist: torch (>=2.1.0,<3.0.0)
-Requires-Dist: torchaudio (>=2.1.0,<3.0.0)
+Requires-Dist: torch (>=2.2,<2.3)
+Requires-Dist: torchaudio (>=2.2,<2.3)
 Requires-Dist: torchinfo (>=1.8.0,<2.0.0)
 Requires-Dist: tqdm (>=4.66.1,<5.0.0)
 Description-Content-Type: text/x-rst
@@ -59,7 +55,7 @@ Sonus AI: Framework for simplified creation of deep NN models for sound, speech,
 Sonus AI includes functions for pre-processing training and validation data and
 creating performance metrics reports for key types of Keras models:
 - recurrent, convolutional, or a combination (i.e. RCNNs)
-- binary, multiclass single-label, multiclass multi-label, and regresssion
+- binary, multiclass single-label, multiclass multi-label, and regression
 - training with data augmentations:  noise mixing, pitch and time stretch, etc.
 
 Sonus AI python functions are used by:

{sonusai-0.15.5 → sonusai-0.15.8}/README.rst

@@ -3,7 +3,7 @@ Sonus AI: Framework for simplified creation of deep NN models for sound, speech,
 Sonus AI includes functions for pre-processing training and validation data and
 creating performance metrics reports for key types of Keras models:
 - recurrent, convolutional, or a combination (i.e. RCNNs)
-- binary, multiclass single-label, multiclass multi-label, and regresssion
+- binary, multiclass single-label, multiclass multi-label, and regression
 - training with data augmentations:  noise mixing, pitch and time stretch, etc.
 
 Sonus AI python functions are used by:

{sonusai-0.15.5 → sonusai-0.15.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sonusai"
-version = "0.15.5"
+version = "0.15.8"
 description = "Framework for building deep neural network models for sound, speech, and voice AI"
 authors = ["Chris Eddington <chris@aaware.com>", "Jason Calderwood <jason@aaware.com>"]
 maintainers = ["Chris Eddington <chris@aaware.com>", "Jason Calderwood <jason@aaware.com>"]
@@ -16,24 +16,23 @@ aawscd_probwrite = 'sonusai.aawscd_probwrite:main'
 [tool.poetry.dependencies]
 PyYAML = "^6.0.1"
 aixplain = "^0.2.6"
+ctranslate2 = "4.1.0"
 dataclasses-json = "^0.6.1"
 deepgram-sdk = "^3.0.0"
 docopt = "^0.6.2"
-#einops = "^0"
-faster-whisper = "^0.10.0"
-ffmpeg = "^1.4"
-greenlet = "^3.0.1"
-grpcio = "1.60.0"
+faster-whisper = "^1.0.1"
+h5py = "^3.11.0"
 jiwer = "^3.0.3"
-keras-tuner = "^1.4.5"
+keras = "^3.1.1"
+keras-tuner = "^1.4.7"
 librosa = "^0.10.1"
-lightning = "^2.1.0"
+lightning = "~2.2"
 matplotlib = "^3.8.0"
-onnx = "1.14.1"
+onnx = "^1.14.1"
 #onnxruntime-gpu = "^1.16.1"
 onnxruntime = "^1.16.1"
 #openai-whisper = "^20231117"
-paho-mqtt = "^1.6.1"
+paho-mqtt = "^2.0.0"
 pandas = "^2.1.1"
 pesq = "^0.0.4"
 pyaaware = "^1.5.3"
@@ -42,18 +41,14 @@ pystoi = "^0.4.0"
 python = ">=3.9,<3.12"
 python-magic = "^0.4.27"
 requests = "^2.31.0"
-resampy = "^0.4.2"
-samplerate = "^0.1.0"
-scikit-learn = "^1.3.1"
-sh = "^2.0.6"
+samplerate = "^0.2.1"
 soundfile = "^0.12.1"
+speechrecognition = "^3.10.1"
 sox = "^1.4.1"
-speechrecognition = "^3.10.0"
 tensorflow = "^2.15.0"
-tensorflow-addons = "^0.23.0"
 tf2onnx = "^1.15.1"
-torch = "^2.1.0"
-torchaudio = "^2.1.0"
+torch = "~2.2"
+torchaudio = "~2.2"
 torchinfo = "^1.8.0"
 tqdm = "^4.66.1"
 
@@ -63,7 +58,7 @@ ipython = "^8.16.1"
 jupyter = "^1.0.0"
 mypy = "^1.6.0"
 mypy-extensions = "^1.0.0"
-pytest = "^7.4.2"
+pytest = "^8.1.1"
 types-pyyaml = "^6.0.12.12"
 types-requests = "^2.31.0.8"
 yappi = "^1.4.0"

{sonusai-0.15.5 → sonusai-0.15.8}/sonusai/mixture/generation.py

@@ -433,7 +433,7 @@ def _initialize_targets_audio(mixdb: MixtureDatabase, mixture: Mixture) -> tuple
         # target_gain is used to back out the gain augmentation in order to return the target audio
         # to its normalized level when calculating truth (if needed).
         if target.augmentation.gain is not None:
-            target.gain = 10 ** (target.augmentation.gain / 20)
+            target.gain = round(10 ** (target.augmentation.gain / 20), ndigits=5)
         else:
             target.gain = 1
 
@@ -507,6 +507,8 @@ def _initialize_mixture_gains(mixdb: MixtureDatabase,
         mixture.target_snr_gain *= gain_adjustment
         mixture.noise_snr_gain *= gain_adjustment
 
+    mixture.target_snr_gain = round(mixture.target_snr_gain, ndigits=5)
+    mixture.noise_snr_gain = round(mixture.noise_snr_gain, ndigits=5)
     return mixture
 
 

{sonusai-0.15.5 → sonusai-0.15.8}/sonusai/mixture/mixdb.py

@@ -1,5 +1,6 @@
 from functools import cached_property
 from functools import lru_cache
+from functools import partial
 from sqlite3 import Connection
 from sqlite3 import Cursor
 from typing import Any
@@ -78,8 +79,6 @@ class SQLiteContextManager:
 
 class MixtureDatabase:
     def __init__(self, location: str, test: bool = False) -> None:
-        from functools import partial
-
         self.location = location
         self.db = partial(SQLiteContextManager, self.location, test)
 
@@ -380,24 +379,13 @@ class MixtureDatabase:
                                  t_num=spectral_mask[3],
                                  t_max_percent=spectral_mask[4]) for spectral_mask in results.fetchall()]
 
-    @lru_cache
     def spectral_mask(self, sm_id: int) -> SpectralMask:
         """Get spectral mask with ID from db
 
         :param sm_id: Spectral mask ID
         :return: Spectral mask
         """
-        with self.db() as c:
-            spectral_mask = c.execute(
-                "SELECT spectral_mask.f_max_width, f_num, t_max_width, t_num, t_max_percent " +
-                "FROM spectral_mask " +
-                "WHERE ? = spectral_mask.id",
-                (sm_id,)).fetchone()
-            return SpectralMask(f_max_width=spectral_mask[0],
-                                f_num=spectral_mask[1],
-                                t_max_width=spectral_mask[2],
-                                t_num=spectral_mask[3],
-                                t_max_percent=spectral_mask[4])
+        return _spectral_mask(self.db, sm_id)
 
     @cached_property
     def target_files(self) -> TargetFiles:
@@ -439,37 +427,13 @@ class MixtureDatabase:
         with self.db() as c:
             return [int(item[0]) for item in c.execute("SELECT target_file.id FROM target_file").fetchall()]
 
-    @lru_cache
     def target_file(self, t_id: int) -> TargetFile:
         """Get target file with ID from db
 
         :param t_id: Target file ID
         :return: Target file
         """
-        import json
-
-        from .datatypes import TruthSetting
-        from .datatypes import TruthSettings
-
-        with self.db() as c:
-            target = c.execute("SELECT target_file.name, samples, level_type FROM target_file WHERE ? = target_file.id",
-                               (t_id,)).fetchone()
-
-            truth_settings: TruthSettings = []
-            for ts in c.execute(
-                    "SELECT truth_setting.setting " +
-                    "FROM truth_setting, target_file_truth_setting " +
-                    "WHERE ? = target_file_truth_setting.target_file_id " +
-                    "AND truth_setting.id = target_file_truth_setting.truth_setting_id",
-                    (t_id,)).fetchall():
-                entry = json.loads(ts[0])
-                truth_settings.append(TruthSetting(config=entry.get('config', None),
-                                                   function=entry.get('function', None),
-                                                   index=entry.get('index', None)))
-            return TargetFile(name=target[0],
-                              samples=target[1],
-                              level_type=target[2],
-                              truth_settings=truth_settings)
+        return _target_file(self.db, t_id)
 
     @cached_property
     def num_target_files(self) -> int:
@@ -499,17 +463,13 @@ class MixtureDatabase:
         with self.db() as c:
             return [int(item[0]) for item in c.execute("SELECT noise_file.id FROM noise_file").fetchall()]
 
-    @lru_cache
     def noise_file(self, n_id: int) -> NoiseFile:
         """Get noise file with ID from db
 
         :param n_id: Noise file ID
         :return: Noise file
         """
-        with self.db() as c:
-            noise = c.execute("SELECT noise_file.name, samples FROM noise_file WHERE ? = noise_file.id",
-                              (n_id,)).fetchone()
-            return NoiseFile(name=noise[0], samples=noise[1])
+        return _noise_file(self.db, n_id)
 
     @cached_property
     def num_noise_files(self) -> int:
@@ -540,17 +500,13 @@ class MixtureDatabase:
             return [int(item[0]) for item in
                     c.execute("SELECT impulse_response_file.id FROM impulse_response_file").fetchall()]
 
-    @lru_cache
     def impulse_response_file(self, ir_id: int) -> str:
         """Get impulse response file with ID from db
 
         :param ir_id: Impulse response file ID
         :return: Noise
         """
-        with self.db() as c:
-            return str(c.execute(
-                "SELECT impulse_response_file.file FROM impulse_response_file WHERE ? = impulse_response_file.id",
-                (ir_id + 1,)).fetchone()[0])
+        return _impulse_response_file(self.db, ir_id)
 
     @cached_property
     def num_impulse_response_files(self) -> int:
@@ -593,31 +549,13 @@ class MixtureDatabase:
         with self.db() as c:
             return [int(item[0]) - 1 for item in c.execute("SELECT mixture.id FROM mixture").fetchall()]
 
-    @lru_cache
     def mixture(self, m_id: int) -> Mixture:
         """Get mixture record with ID from db
 
         :param m_id: Zero-based mixture ID
         :return: Mixture record
         """
-        from .helpers import to_mixture
-        from .helpers import to_target
-
-        with self.db() as c:
-            mixture = c.execute(
-                "SELECT mixture.name, noise_file_id, noise_augmentation, noise_offset, noise_snr_gain, " +
-                "random_snr, snr, samples, spectral_mask_id, spectral_mask_seed, target_snr_gain, id " +
-                "FROM mixture " +
-                "WHERE ? = mixture.id",
-                (m_id + 1,)).fetchone()
-
-            targets = [to_target(target) for target in c.execute(
-                "SELECT target.file_id, augmentation, gain " +
-                "FROM target, mixture_target " +
-                "WHERE ? = mixture_target.mixture_id AND target.id = mixture_target.target_id",
-                (mixture[11],)).fetchall()]
-
-            return to_mixture(mixture, targets)
+        return _mixture(self.db, m_id)
 
     @cached_property
     def mixid_width(self) -> int:
@@ -1127,3 +1065,114 @@ class MixtureDatabase:
             class_count[cl] = int(np.sum(truth_t[:, cl] >= self.class_weights_thresholds[cl]))
 
         return class_count
+
+
+@lru_cache
+def _spectral_mask(db: partial, sm_id: int) -> SpectralMask:
+    """Get spectral mask with ID from db
+
+    :param db: Database context
+    :param sm_id: Spectral mask ID
+    :return: Spectral mask
+    """
+    with db() as c:
+        spectral_mask = c.execute(
+            "SELECT spectral_mask.f_max_width, f_num, t_max_width, t_num, t_max_percent " +
+            "FROM spectral_mask " +
+            "WHERE ? = spectral_mask.id",
+            (sm_id,)).fetchone()
+        return SpectralMask(f_max_width=spectral_mask[0],
+                            f_num=spectral_mask[1],
+                            t_max_width=spectral_mask[2],
+                            t_num=spectral_mask[3],
+                            t_max_percent=spectral_mask[4])
+
+
+@lru_cache
+def _target_file(db: partial, t_id: int) -> TargetFile:
+    """Get target file with ID from db
+
+    :param db: Database context
+    :param t_id: Target file ID
+    :return: Target file
+    """
+    import json
+
+    from .datatypes import TruthSetting
+    from .datatypes import TruthSettings
+
+    with db() as c:
+        target = c.execute("SELECT target_file.name, samples, level_type FROM target_file WHERE ? = target_file.id",
+                           (t_id,)).fetchone()
+
+        truth_settings: TruthSettings = []
+        for ts in c.execute(
+                "SELECT truth_setting.setting " +
+                "FROM truth_setting, target_file_truth_setting " +
+                "WHERE ? = target_file_truth_setting.target_file_id " +
+                "AND truth_setting.id = target_file_truth_setting.truth_setting_id",
+                (t_id,)).fetchall():
+            entry = json.loads(ts[0])
+            truth_settings.append(TruthSetting(config=entry.get('config', None),
+                                               function=entry.get('function', None),
+                                               index=entry.get('index', None)))
+        return TargetFile(name=target[0],
+                          samples=target[1],
+                          level_type=target[2],
+                          truth_settings=truth_settings)
+
+
+@lru_cache
+def _noise_file(db: partial, n_id: int) -> NoiseFile:
+    """Get noise file with ID from db
+
+    :param db: Database context
+    :param n_id: Noise file ID
+    :return: Noise file
+    """
+    with db() as c:
+        noise = c.execute("SELECT noise_file.name, samples FROM noise_file WHERE ? = noise_file.id",
+                          (n_id,)).fetchone()
+        return NoiseFile(name=noise[0], samples=noise[1])
+
+
+@lru_cache
+def _impulse_response_file(db: partial, ir_id: int) -> str:
+    """Get impulse response file with ID from db
+
+    :param db: Database context
+    :param ir_id: Impulse response file ID
+    :return: Noise
+    """
+    with db() as c:
+        return str(c.execute(
+            "SELECT impulse_response_file.file FROM impulse_response_file WHERE ? = impulse_response_file.id",
+            (ir_id + 1,)).fetchone()[0])
+
+
+@lru_cache
+def _mixture(db: partial, m_id: int) -> Mixture:
+    """Get mixture record with ID from db
+
+    :param db: Database context
+    :param m_id: Zero-based mixture ID
+    :return: Mixture record
+    """
+    from .helpers import to_mixture
+    from .helpers import to_target
+
+    with db() as c:
+        mixture = c.execute(
+            "SELECT mixture.name, noise_file_id, noise_augmentation, noise_offset, noise_snr_gain, " +
+            "random_snr, snr, samples, spectral_mask_id, spectral_mask_seed, target_snr_gain, id " +
+            "FROM mixture " +
+            "WHERE ? = mixture.id",
+            (m_id + 1,)).fetchone()
+
+        targets = [to_target(target) for target in c.execute(
+            "SELECT target.file_id, augmentation, gain " +
+            "FROM target, mixture_target " +
+            "WHERE ? = mixture_target.mixture_id AND target.id = mixture_target.target_id",
+            (mixture[11],)).fetchall()]
+
+        return to_mixture(mixture, targets)

{sonusai-0.15.5 → sonusai-0.15.8}/sonusai/mixture/truth_functions/data.py

@@ -23,6 +23,8 @@ class Data:
                               num_classes=config.num_classes,
                               truth_mutex=config.mutex)
 
+        self.num_bands = fg.num_bands
+        self.ttype = fg.ftransform_ttype
         self.frame_size = fg.ftransform_R
 
         if len(target_audio) % self.frame_size != 0:

{sonusai-0.15.5 → sonusai-0.15.8}/sonusai/mixture/truth_functions/target.py

@@ -13,12 +13,13 @@ Calculates the true transform of the target using the STFT
 configuration defined by the feature. This will include a
 forward transform window if defined by the feature.
 
-Output shape: [:, 2 * bins] (stacked real, imag)
+Output shape: [:, num_classes]
+                    (target stacked real, imag; or real only for tdac-co)
     """
 
     from sonusai import SonusAIError
 
-    if data.config.num_classes != 2 * data.target_fft.bins:
+    if data.config.num_classes != data.num_bands:
         raise SonusAIError(f'Invalid num_classes for target_f truth: {data.config.num_classes}')
 
     target_freq = _execute_fft(data.target_audio, data.target_fft, len(data.offsets))
@@ -28,6 +29,7 @@ Output shape: [:, 2 * bins] (stacked real, imag)
                                       frame_size=data.frame_size,
                                       zero_based_indices=data.zero_based_indices,
                                       bins=data.target_fft.bins,
+                                      ttype=data.ttype,
                                       start=0,
                                       truth=data.truth)
 
@@ -43,11 +45,13 @@ using the STFT configuration defined by the feature. This
 will include a forward transform window if defined by the
 feature.
 
-Output shape: [:, 4 * bins] (target stacked real, imag; mixture stacked real, imag)
+Output shape: [:, 2 * num_classes]
+                    (target stacked real, imag; or real only for tdac-co)
+                    (mixture stacked real, imag; or real only for tdac-co)
     """
     from sonusai import SonusAIError
 
-    if data.config.num_classes != 2 * data.target_fft.bins + 2 * data.mixture_fft.bins:
+    if data.config.num_classes != 2 * data.num_bands:
         raise SonusAIError(f'Invalid num_classes for target_mixture_f truth: {data.config.num_classes}')
 
     target_freq = _execute_fft(data.target_audio, data.target_fft, len(data.offsets))
@@ -59,6 +63,7 @@ Output shape: [:, 4 * bins] (target stacked real, imag; mixture stacked real, im
                                       frame_size=data.frame_size,
                                       zero_based_indices=data.zero_based_indices,
                                       bins=data.target_fft.bins,
+                                      ttype=data.ttype,
                                       start=0,
                                       truth=data.truth)
 
@@ -67,6 +72,7 @@ Output shape: [:, 4 * bins] (target stacked real, imag; mixture stacked real, im
                                       frame_size=data.frame_size,
                                       zero_based_indices=data.zero_based_indices,
                                       bins=data.target_fft.bins,
+                                      ttype=data.ttype,
                                       start=data.target_fft.bins * 2,
                                       truth=data.truth)
 
@@ -125,6 +131,7 @@ def _stack_real_imag(data: AudioF,
                      frame_size: int,
                      zero_based_indices: list[int],
                      bins: int,
+                     ttype: str,
                      start: int,
                      truth: Truth) -> Truth:
     import numpy as np
@@ -134,7 +141,8 @@ def _stack_real_imag(data: AudioF,
         b = _get_bin_slice(index + start, bins)
         truth[i, b] = np.real(data)
 
-        b = _get_bin_slice(b.stop, bins)
-        truth[i, b] = np.imag(data)
+        if ttype != 'tdac-co':
+            b = _get_bin_slice(b.stop, bins)
+            truth[i, b] = np.imag(data)
 
     return truth