sonusai 0.19.8__py3-none-any.whl → 0.19.9__py3-none-any.whl

sonusai/genmixdb.py

@@ -292,7 +292,7 @@ def genmixdb(
             augmented_targets=augmented_targets,
             targets=target_files,
             target_augmentations=target_augmentations,
-            class_balancing_augmentation=class_balancing_augmentation,
+            class_balancing_augmentation=class_balancing_augmentation,  # pyright: ignore [reportArgumentType]
             num_classes=mixdb.num_classes,
             num_ir=mixdb.num_impulse_response_files,
             mixups=mixups,
@@ -354,7 +354,7 @@ def genmixdb(
         logger.info(
             f"Feature shape:        "
             f"{mixdb.fg_stride} x {mixdb.feature_parameters} "
-            f"({mixdb.fg_stride * mixdb.feature_parameters} total params)"
+            f"({mixdb.fg_stride * mixdb.feature_parameters} total parameters)"
         )
         logger.info(f"Feature samples:      {mixdb.feature_samples} samples ({mixdb.feature_ms} ms)")
         logger.info(f"Feature step samples: {mixdb.feature_step_samples} samples ({mixdb.feature_step_ms} ms)")

sonusai/mixture/generation.py

@@ -406,6 +406,10 @@ def populate_mixture_table(
     test: bool = False,
 ) -> tuple[int, int]:
     """Generate mixtures and populate mixture table"""
+    from .helpers import from_mixture
+    from .helpers import from_target
+    from .mixdb import db_connection
+
     if noise_mix_mode == "exhaustive":
         func = _exhaustive_noise_mix
     elif noise_mix_mode == "non-exhaustive":
@@ -415,8 +419,7 @@ def populate_mixture_table(
     else:
         raise ValueError(f"invalid noise_mix_mode: {noise_mix_mode}")
 
-    used_noise_files, used_noise_samples = func(
-        location=location,
+    used_noise_files, used_noise_samples, mixtures = func(
         augmented_targets=augmented_targets,
         target_files=target_files,
         target_augmentations=target_augmentations,
@@ -428,9 +431,48 @@ def populate_mixture_table(
         num_classes=num_classes,
         feature_step_samples=feature_step_samples,
         num_ir=num_ir,
-        test=test,
     )
 
+    con = db_connection(location=location, readonly=False, test=test)
+    # Populate target table
+    for mixture in mixtures:
+        for target in mixture.targets:
+            con.execute(
+                """
+                INSERT OR IGNORE INTO target (file_id, augmentation)
+                VALUES (?, ?)
+                """,
+                from_target(target),
+            )
+
+    # Populate mixture table
+    for m_id, mixture in enumerate(mixtures):
+        con.execute(
+            """
+            INSERT INTO mixture (id, name, noise_file_id, noise_augmentation, noise_offset, noise_snr_gain, random_snr,
+            snr, samples, spectral_mask_id, spectral_mask_seed, target_snr_gain)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+            """,
+            (m_id + 1, *from_mixture(mixture)),
+        )
+
+        for target in mixture.targets:
+            target_id = con.execute(
+                """
+                SELECT target.id
+                FROM target
+                WHERE ? = target.file_id AND ? = target.augmentation
+            """,
+                from_target(target),
+            ).fetchone()[0]
+            con.execute(
+                "INSERT INTO mixture_target (mixture_id, target_id) VALUES (?, ?)",
+                (m_id + 1, target_id),
+            )
+
+    con.commit()
+    con.close()
+
     return used_noise_files, used_noise_samples
 
 
@@ -603,7 +645,6 @@ def _initialize_mixture_gains(
 
 
 def _exhaustive_noise_mix(
-    location: str,
     augmented_targets: list[AugmentedTarget],
     target_files: list[TargetFile],
     target_augmentations: list[AugmentationRule],
@@ -615,9 +656,8 @@ def _exhaustive_noise_mix(
     num_classes: int,
     feature_step_samples: int,
     num_ir: int,
-    test: bool = False,
-) -> tuple[int, int]:
-    """Use every noise/augmentation with every target/augmentation"""
+) -> tuple[int, int, list[Mixture]]:
+    """Use every noise/augmentation with every target/augmentation+interferences/augmentation"""
     from random import randint
 
     import numpy as np
@@ -643,6 +683,8 @@ def _exhaustive_noise_mix(
         )
         for mixup in mixups
     ]
+
+    mixtures: list[Mixture] = []
     for noise_file_id in range(len(noise_files)):
         for noise_augmentation_rule in noise_augmentations:
             noise_augmentation = augmentation_from_rule(noise_augmentation_rule, num_ir)
@@ -665,10 +707,8 @@ def _exhaustive_noise_mix(
 
                     for spectral_mask_id in range(len(spectral_masks)):
                         for snr in all_snrs:
-                            _insert_mixture_record(
-                                location=location,
-                                m_id=m_id,
-                                mixture=Mixture(
+                            mixtures.append(
+                                Mixture(
                                     targets=targets,
                                     name=str(m_id),
                                     noise=Noise(file_id=noise_file_id + 1, augmentation=noise_augmentation),
@@ -677,19 +717,17 @@ def _exhaustive_noise_mix(
                                     snr=UniversalSNR(value=snr.value, is_random=snr.is_random),
                                     spectral_mask_id=spectral_mask_id + 1,
                                     spectral_mask_seed=randint(0, np.iinfo("i").max),  # noqa: S311
-                                ),
-                                test=test,
+                                )
                             )
                             m_id += 1
 
                             noise_offset = int((noise_offset + target_length) % noise_length)
                             used_noise_samples += target_length
 
-    return used_noise_files, used_noise_samples
+    return used_noise_files, used_noise_samples, mixtures
 
 
 def _non_exhaustive_noise_mix(
-    location: str,
     augmented_targets: list[AugmentedTarget],
     target_files: list[TargetFile],
     target_augmentations: list[AugmentationRule],
@@ -701,10 +739,9 @@ def _non_exhaustive_noise_mix(
     num_classes: int,
     feature_step_samples: int,
     num_ir: int,
-    test: bool = False,
-) -> tuple[int, int]:
-    """Cycle through every target/augmentation without necessarily using all noise/augmentation combinations
-    (reduced data set).
+) -> tuple[int, int, list[Mixture]]:
+    """Cycle through every target/augmentation+interferences/augmentation without necessarily using all
+    noise/augmentation combinations (reduced data set).
     """
     from random import randint
 
@@ -732,6 +769,8 @@ def _non_exhaustive_noise_mix(
         )
         for mixup in mixups
     ]
+
+    mixtures: list[Mixture] = []
     for mixup in augmented_target_indices_for_mixups:
         for augmented_target_indices in mixup:
             targets, target_length = _get_target_info(
@@ -763,10 +802,8 @@ def _non_exhaustive_noise_mix(
 
                     used_noise_files.add(f"{noise_file_id}_{noise_augmentation_id}")
 
-                    _insert_mixture_record(
-                        location=location,
-                        m_id=m_id,
-                        mixture=Mixture(
+                    mixtures.append(
+                        Mixture(
                             targets=targets,
                             name=str(m_id),
                             noise=Noise(file_id=noise_file_id + 1, augmentation=noise_augmentation),
@@ -775,16 +812,14 @@ def _non_exhaustive_noise_mix(
                             snr=UniversalSNR(value=snr.value, is_random=snr.is_random),
                             spectral_mask_id=spectral_mask_id + 1,
                             spectral_mask_seed=randint(0, np.iinfo("i").max),  # noqa: S311
-                        ),
-                        test=test,
+                        )
                     )
                     m_id += 1
 
-    return len(used_noise_files), used_noise_samples
+    return len(used_noise_files), used_noise_samples, mixtures
 
 
 def _non_combinatorial_noise_mix(
-    location: str,
     augmented_targets: list[AugmentedTarget],
     target_files: list[TargetFile],
     target_augmentations: list[AugmentationRule],
@@ -796,11 +831,10 @@ def _non_combinatorial_noise_mix(
     num_classes: int,
     feature_step_samples: int,
     num_ir: int,
-    test: bool = False,
-) -> tuple[int, int]:
-    """Combine a target/augmentation with a single cut of a noise/augmentation non-exhaustively
-    (each target/augmentation does not use each noise/augmentation). Cut has random start and loop back to
-    beginning if end of noise/augmentation is reached.
+) -> tuple[int, int, list[Mixture]]:
+    """Combine a target/augmentation+interferences/augmentation with a single cut of a noise/augmentation
+    non-exhaustively (each target/augmentation+interferences/augmentation does not use each noise/augmentation).
+    Cut has random start and loop back to beginning if end of noise/augmentation is reached.
     """
     from random import choice
     from random import randint
@@ -828,6 +862,8 @@ def _non_combinatorial_noise_mix(
         )
         for mixup in mixups
     ]
+
+    mixtures: list[Mixture] = []
     for mixup in augmented_target_indices_for_mixups:
         for augmented_target_indices in mixup:
             targets, target_length = _get_target_info(
@@ -857,10 +893,8 @@ def _non_combinatorial_noise_mix(
 
                     used_noise_files.add(f"{noise_file_id}_{noise_augmentation_id}")
 
-                    _insert_mixture_record(
-                        location=location,
-                        m_id=m_id,
-                        mixture=Mixture(
+                    mixtures.append(
+                        Mixture(
                             targets=targets,
                             name=str(m_id),
                             noise=Noise(file_id=noise_file_id + 1, augmentation=noise_augmentation),
@@ -869,12 +903,11 @@ def _non_combinatorial_noise_mix(
                             snr=UniversalSNR(value=snr.value, is_random=snr.is_random),
                             spectral_mask_id=spectral_mask_id + 1,
                             spectral_mask_seed=randint(0, np.iinfo("i").max),  # noqa: S311
-                        ),
-                        test=test,
+                        )
                     )
                     m_id += 1
 
-    return len(used_noise_files), used_noise_samples
+    return len(used_noise_files), used_noise_samples, mixtures
 
 
 def _get_next_noise_indices(
@@ -973,49 +1006,6 @@ def _get_target_info(
     return mixups, target_length
 
 
-def _insert_mixture_record(location: str, m_id: int, mixture: Mixture, test: bool = False) -> None:
-    from .helpers import from_mixture
-    from .helpers import from_target
-    from .mixdb import db_connection
-
-    con = db_connection(location=location, readonly=False, test=test)
-    # Populate target table
-    for target in mixture.targets:
-        con.execute(
-            """
-            INSERT OR IGNORE INTO target (file_id, augmentation)
-            VALUES (?, ?)
-            """,
-            from_target(target),
-        )
-
-    # Populate mixture table
-    con.execute(
-        """
-        INSERT INTO mixture (id, name, noise_file_id, noise_augmentation, noise_offset, noise_snr_gain, random_snr,
-        snr, samples, spectral_mask_id, spectral_mask_seed, target_snr_gain)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
-        """,
-        (m_id + 1, *from_mixture(mixture)),
-    )
-
-    for target in mixture.targets:
-        target_id = con.execute(
-            """
-            SELECT target.id
-            FROM target
-            WHERE ? = target.file_id AND ? = target.augmentation
-        """,
-            from_target(target),
-        ).fetchone()[0]
-        con.execute(
-            "INSERT INTO mixture_target (mixture_id, target_id) VALUES (?, ?)",
-            (m_id + 1, target_id),
-        )
-    con.commit()
-    con.close()
-
-
 def get_all_snrs_from_config(config: dict) -> list[UniversalSNRGenerator]:
     from .datatypes import UniversalSNRGenerator
 

sonusai/mixture/torchaudio_audio.py

@@ -4,10 +4,16 @@ from sonusai.mixture.datatypes import AudioT
 from sonusai.mixture.datatypes import ImpulseResponseData
 
 
-def read_impulse_response(name: str | Path) -> ImpulseResponseData:
+def read_impulse_response(
+    name: str | Path,
+    delay_compensation: bool = True,
+    normalize: bool = True,
+) -> ImpulseResponseData:
     """Read impulse response data using torchaudio
 
     :param name: File name
+    :param delay_compensation: Apply delay compensation
+    :param normalize: Apply normalization
     :return: ImpulseResponseData object
     """
     import numpy as np
@@ -28,14 +34,19 @@ def read_impulse_response(name: str | Path) -> ImpulseResponseData:
             raise OSError(f"Error reading {name}: {e}") from e
 
     raw = torch.squeeze(raw[0, :])
-    offset = torch.argmax(raw)
-    raw = raw[offset:]
-    # Inexplicably, torch.linalg.vector_norm() causes multiprocessing contexts to hang.
-    # Use np.linalg.norm() instead.
-    # raw = raw / torch.linalg.vector_norm(raw)
+
+    if delay_compensation:
+        offset = torch.argmax(raw)
+        raw = raw[offset:]
 
     data = np.array(raw).astype(np.float32)
-    data = data / np.linalg.norm(data)
+
+    if normalize:
+        # Inexplicably,
+        #   data = data / torch.linalg.vector_norm(data)
+        # causes multiprocessing contexts to hang.
+        # Use np.linalg.norm() instead.
+        data = data / np.linalg.norm(data)
 
     return ImpulseResponseData(name=str(name), sample_rate=sample_rate, data=data)
 

sonusai/mixture/torchaudio_augmentation.py

@@ -20,10 +20,9 @@ def apply_augmentation(audio: AudioT, augmentation: Augmentation, frame_length:
 
     effects: list[list[str]] = []
 
-    # TODO
-    #  Always normalize and remove normalize from list of available augmentations
-    #  Normalize to globally set level (should this be a global config parameter,
-    #  or hard-coded into the script?)
+    # TODO: Always normalize and remove normalize from list of available augmentations
+    # Normalize to globally set level (should this be a global config parameter, or hard-coded into the script?)
+    # TODO: Support all sox effects supported by torchaudio (torchaudio.sox_effects.effect_names())
     if augmentation.normalize is not None:
         effects.append(["norm", str(augmentation.normalize)])
 

{sonusai-0.19.8.dist-info → sonusai-0.19.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sonusai
-Version: 0.19.8
+Version: 0.19.9
 Summary: Framework for building deep neural network models for sound, speech, and voice AI
 Home-page: https://aaware.com
 License: GPL-3.0-only

{sonusai-0.19.8.dist-info → sonusai-0.19.9.dist-info}/RECORD

@@ -15,7 +15,7 @@ sonusai/doc.py,sha256=zSmXpioB0YS_5-7kqfS5cr--veSaXkxRKzldId9Hyoc,878
 sonusai/genft.py,sha256=TqtmexKw7tZsugU1DbZ3fY7_YZ2hFlIU-ema6f0LRgQ,5586
 sonusai/genmetrics.py,sha256=RiEYmkRl0yVCIr9PHkKRqEG68WR77pP9WpWAYeKnMLE,5723
 sonusai/genmix.py,sha256=pugp74prksf4ZiJYMXWg2l3v48Qqi9KpP2WhPEBpZOU,6702
-sonusai/genmixdb.py,sha256=QB9iARnAutuPPxML-g-kc3C-W1Sodabr68e4GSRD21Q,18680
+sonusai/genmixdb.py,sha256=gUF_9dGfuWRCmHtCoXZRrgCMejz46A5KhWoEh31laqc,18724
 sonusai/lsdb.py,sha256=0HOGDDndB3LT9cz9AaxKIpt9vslAoSP4F239gply4Xg,5149
 sonusai/main.py,sha256=HbnEia1B1-Z-mlHkLfojH8aj9GIpL1Btw3oH60T_CCQ,2590
 sonusai/metrics/__init__.py,sha256=ssV6JEK_oklRSocsp6HMcG-GtJvV8IkRQtdKhHHmwU8,878
@@ -46,7 +46,7 @@ sonusai/mixture/datatypes.py,sha256=xNDBWFTVQ3plJ7qHKzrXyV4pffPYuf1xMVqBsR40n4o,
 sonusai/mixture/db_datatypes.py,sha256=kvdUOMS6Pkkj9AmxCiq6zM8x7jbPPi933tVaXRxbTdQ,1534
 sonusai/mixture/eq_rule_is_valid.py,sha256=O3gCAs_0hpxENK5b7kxxpDmOpKHlXGBWuLGT_97ARSM,1210
 sonusai/mixture/feature.py,sha256=L0bPFG0RO-CrrtTStUMt_14euYsVo8_TWTP2IKSFKaA,2335
-sonusai/mixture/generation.py,sha256=sPsl-r5d9mkSglHOG5_VbmoDNqOscuWTdu7F6OJa-Qw,38751
+sonusai/mixture/generation.py,sha256=Okmyc7LVVdyt7UDzLSR4XdK-Q92vSQHX0RI__RJbJM8,38551
 sonusai/mixture/helpers.py,sha256=UEN_9LttIZwvms9nOkjMr_Yf3lpb7sMPmDVUoX8D7yQ,15225
 sonusai/mixture/log_duration_and_sizes.py,sha256=qhgl87C2KbjxLdKEpjYOoqNL6rc-8-PB4R7Gx_7UG8g,1240
 sonusai/mixture/mixdb.py,sha256=W8B54U6f-8szMylXSnU0YzOR-a9M3XoCRpyqOgBIJuQ,73498
@@ -57,8 +57,8 @@ sonusai/mixture/spectral_mask.py,sha256=U9XJ_SAoI9b67K_3SE7bNw6U8cPGFOBttaZAxMjA
 sonusai/mixture/target_class_balancing.py,sha256=o_TZ8kVYq10lgeXHh3GUFfflfdUvRt0FekFu2eaNkDs,4251
 sonusai/mixture/targets.py,sha256=6emo2fxxp9ZhSpHuUM9xIjYMz8zeIHAw684jT3l7fAs,6442
 sonusai/mixture/tokenized_shell_vars.py,sha256=lXTzUDutuBWGV1zIsqeIxWmy-eKm0Vx1y8-iLdsL1gQ,4921
-sonusai/mixture/torchaudio_audio.py,sha256=169VXKEoOHc5nyiHJwaihkcQ_a1ZH_O-nnk9Gq4JtaQ,2887
-sonusai/mixture/torchaudio_augmentation.py,sha256=68QpJ4JcAH2fsL4qVck7VvBv615UbCBPEnNe4HMvA_8,3844
+sonusai/mixture/torchaudio_audio.py,sha256=72Hxo5TKAW7mYpRy15QFfD7AYDORBk6bVCcHENniWGw,3116
+sonusai/mixture/torchaudio_augmentation.py,sha256=uFAKxIfs50J5FR-WXodsEACm2Ao-t5dZRSJ0DwTAfBg,3930
 sonusai/mixture/truth.py,sha256=-CwwawFRGjqodR2yKvAMGL1XaYLct-tli7wZ2gbhLtQ,2121
 sonusai/mixture/truth_functions/__init__.py,sha256=0mlOFChPnXG5BC0eKOe4n9VH17jY4iOqZFLuF6Gprdk,1505
 sonusai/mixture/truth_functions/crm.py,sha256=iidcffXfqV8k9O5wt5KTWIAFaTSjmhV5ucKZPbTgpvQ,3809
@@ -121,7 +121,7 @@ sonusai/utils/stratified_shuffle_split.py,sha256=d7WLUirywSvgZWkt_5a0F8YvTnJjuXl
 sonusai/utils/write_audio.py,sha256=0lKdaX57N6H-UWdioqmXCJMjwT1eBz5B-bSGqDvloAc,838
 sonusai/utils/yes_or_no.py,sha256=0h1okjXmDNbJp7rZJFR2V-HFU1GJDm3YFTUVmYExkOU,263
 sonusai/vars.py,sha256=kBBzuvC8szmdIZEEDA7XXmD765addZKdM2aFipeGO1w,933
-sonusai-0.19.8.dist-info/METADATA,sha256=KQEqF0XgSi8xV6rSYzn-bync-wXfEjzGyWXgb909gDo,2535
-sonusai-0.19.8.dist-info/WHEEL,sha256=Nq82e9rUAnEjt98J6MlVmMCZb-t9cYE2Ir1kpBmnWfs,88
-sonusai-0.19.8.dist-info/entry_points.txt,sha256=zMNjEphEPO6B3cD1GNpit7z-yA9tUU5-j3W2v-UWstU,92
-sonusai-0.19.8.dist-info/RECORD,,
+sonusai-0.19.9.dist-info/METADATA,sha256=rCXEzWOsVKZdZB8E5iPt0fachIQ_-zkmM1UNAe8JZc4,2535
+sonusai-0.19.9.dist-info/WHEEL,sha256=Nq82e9rUAnEjt98J6MlVmMCZb-t9cYE2Ir1kpBmnWfs,88
+sonusai-0.19.9.dist-info/entry_points.txt,sha256=zMNjEphEPO6B3cD1GNpit7z-yA9tUU5-j3W2v-UWstU,92
+sonusai-0.19.9.dist-info/RECORD,,