Audyn 0.0.1.dev6__tar.gz → 0.0.1.dev7__tar.gz

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/Audyn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: Audyn
-Version: 0.0.1.dev6
+Version: 0.0.1.dev7
 Summary: A PyTorch toolkit for audio synthesis.
 Author: Takuya Hasumi
 License:                                  Apache License

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/Audyn.egg-info/SOURCES.txt

@@ -101,6 +101,9 @@ audyn/utils/data/audioset/composer.py
 audyn/utils/data/audioset/dataset.py
 audyn/utils/data/audioset/distributed.py
 audyn/utils/data/audioset/sampler.py
+audyn/utils/data/birdclef/_common/__init__.py
+audyn/utils/data/birdclef/birdclef2022/__init__.py
+audyn/utils/data/birdclef/birdclef2022/_download.py
 audyn/utils/data/birdclef/birdclef2024/__init__.py
 audyn/utils/data/birdclef/birdclef2024/_download.py
 audyn/utils/data/birdclef/birdclef2024/collator.py

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: Audyn
-Version: 0.0.1.dev6
+Version: 0.0.1.dev7
 Summary: A PyTorch toolkit for audio synthesis.
 Author: Takuya Hasumi
 License:                                  Apache License

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/__init__.py

@@ -9,7 +9,7 @@ from .utils.hydra import main
 
 __all__ = ["__version__", "main"]
 
-__version__ = "0.0.1.dev6"
+__version__ = "0.0.1.dev7"
 
 # for resolver
 _whitespace_re = re.compile(r"\s+")

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/optim/lr_scheduler.py

@@ -11,11 +11,17 @@ except ImportError:
 
 __all__ = [
     "_DummyLRScheduler",
+    "_DummyLR",
     "TransformerLRScheduler",
     "NoamScheduler",
+    "TransformerLR",
+    "NoamLR",
     "ExponentialWarmupLinearCooldownLRScheduler",
+    "ExponentialWarmupLinearCooldownLR",
     "MultiLRSchedulers",
+    "MultiLR",
     "GANLRScheduler",
+    "GANLR",
 ]
 
 
@@ -35,6 +41,10 @@ class _DummyLRScheduler:
         assert len(state_dict) == 0
 
 
+class _DummyLR(_DummyLRScheduler):
+    """Alias of _DummyLRScheduler."""
+
+
 class TransformerLRScheduler(_LRScheduler):
     def __init__(
         self,
@@ -63,6 +73,14 @@ class NoamScheduler(TransformerLRScheduler):
     """Alias of TransformerLRScheduler."""
 
 
+class TransformerLR(TransformerLRScheduler):
+    """Alias of TransformerLRScheduler."""
+
+
+class NoamLR(NoamScheduler):
+    """Alias of NoamLR."""
+
+
 class ExponentialWarmupLinearCooldownLRScheduler(LambdaLR):
     """Exponential warm-up + linear cool-down of learning rate.
 
@@ -71,7 +89,7 @@ class ExponentialWarmupLinearCooldownLRScheduler(LambdaLR):
     Args:
         optimizer (Optimizer): Optimizer to adjust learning rate.
         warmup_steps (int): Number of exponential warm-up steps.
-        constant_steps (int): Number of constant learning rate steps between warm-up and cool-down.
+        suspend_steps (int): Number of constant learning rate steps between warm-up and cool-down.
         cooldown_steps (int): Number of linear cool-down steps after constant learning rate.
         last_factor (float): Scale factor of learning rate at last step.
 
@@ -81,7 +99,7 @@ class ExponentialWarmupLinearCooldownLRScheduler(LambdaLR):
         self,
         optimizer: Optimizer,
         warmup_steps: int,
-        constant_steps: int,
+        suspend_steps: int,
         cooldown_steps: int,
         last_factor: float = 1,
         last_epoch: int = -1,
@@ -89,14 +107,13 @@ class ExponentialWarmupLinearCooldownLRScheduler(LambdaLR):
     ) -> None:
         def _lr_scheduler_lambda(step: int) -> float:
             if step < warmup_steps:
-                step = min(step, warmup_steps)
                 normalized_step = 1 - step / warmup_steps
                 factor = math.exp(-5.0 * normalized_step**2)
-            elif step < warmup_steps + constant_steps:
+            elif step < warmup_steps + suspend_steps:
                 factor = 1
-            elif step < warmup_steps + constant_steps + cooldown_steps:
-                step_after_constant = step - (warmup_steps + constant_steps)
-                normalized_step = step_after_constant / cooldown_steps
+            elif step < warmup_steps + suspend_steps + cooldown_steps:
+                step_after_suspend = step - (warmup_steps + suspend_steps)
+                normalized_step = step_after_suspend / cooldown_steps
                 factor = last_factor + (1 - last_factor) * normalized_step
             else:
                 factor = last_factor
@@ -111,6 +128,10 @@ class ExponentialWarmupLinearCooldownLRScheduler(LambdaLR):
         )
 
 
+class ExponentialWarmupLinearCooldownLR(ExponentialWarmupLinearCooldownLRScheduler):
+    """Alias of ExponentialWarmupLinearCooldownLRScheduler."""
+
+
 class MultiLRSchedulers:
     """Module to manage multiple learning rate schedulers."""
 
@@ -172,6 +193,10 @@ class MultiLRSchedulers:
                 lr_scheduler.load_state_dict(state_dict[name])
 
 
+class MultiLR(MultiLRSchedulers):
+    """Alias of MultiLRSchedulers."""
+
+
 class GANLRScheduler:
     def __init__(self, generator: _LRScheduler, discriminator: _LRScheduler) -> None:
         self.generator = generator
@@ -180,3 +205,7 @@ class GANLRScheduler:
     def step(self, *args, **kwargs) -> None:
         self.generator.step(*args, **kwargs)
         self.discriminator.step(*args, **kwargs)
+
+
+class GANLR(GANLRScheduler):
+    """Alias of GANLRScheduler."""

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/utils/data/__init__.py

@@ -6,7 +6,7 @@ from omegaconf import DictConfig
 from torch.utils.data import DataLoader
 
 from .collator import Collator, default_collate_fn, rename_webdataset_keys
-from .composer import AudioFeatureExtractionComposer, Composer
+from .composer import AudioFeatureExtractionComposer, Composer, SequentialComposer
 from .dataloader import (
     DistributedDataLoader,
     DistributedDynamicBatchDataLoader,
@@ -28,6 +28,7 @@ __all__ = [
     "DistributedDynamicBatchDataLoader",
     "Composer",
     "AudioFeatureExtractionComposer",
+    "SequentialComposer",
     "Collator",
     "slice_feautures",
     "take_log_features",

audyn-0.0.1.dev7/audyn/utils/data/birdclef/_common/__init__.py

@@ -0,0 +1,81 @@
+import ast
+import os
+from typing import Any, Dict, List
+
+__all__ = [
+    "decode_csv_line",
+]
+
+
+def decode_csv_line(line: List[str]) -> Dict[str, Any]:
+    """Decode line of train_metadata.csv.
+
+    Args:
+        line (list): One line of train_metadata.csv split by comma (,).
+
+    Returns:
+        dict: Dictionary containing metadata of given line.
+
+    .. note::
+
+        Returned dictionary contains following values.
+
+            - filename (str): Filename with out extension. e.g. ``asbfly/XC134896``.
+            - primary_label (str): Primary label of bird species.
+            - secondary_label (list): Secondary labels of bird species.
+            - type (list): Chirp types.
+            - latitude (float, optional): Latitude of recording.
+            - longitude (float, optional): Longitude of recording.
+            - scientific_name (str): Scientific name of bird.
+            - common_name (str): Common name of bird.
+            - rating (float): Rating.
+            - path (str): Path to audio file equivalent to ``filename`` + ``.ogg``.
+                e.g. ``asbfly/XC134896.ogg``.
+
+    """
+    (
+        primary_label,
+        secondary_labels,
+        chirp_types,
+        latitude,
+        longitude,
+        scientific_name,
+        common_name,
+        _,
+        _,
+        rating,
+        _,
+        path,
+    ) = line
+
+    secondary_labels = ast.literal_eval(secondary_labels)
+    chirp_types = ast.literal_eval(chirp_types)
+    secondary_labels = [secondary_label.lower() for secondary_label in secondary_labels]
+    chirp_types = [chirp_type.lower() for chirp_type in chirp_types]
+
+    filename, _ = os.path.splitext(path)
+
+    if len(latitude) > 0:
+        latitude = float(latitude)
+    else:
+        latitude = None
+
+    if len(longitude) > 0:
+        longitude = float(longitude)
+    else:
+        longitude = None
+
+    data = {
+        "filename": filename,
+        "primary_label": primary_label,
+        "secondary_label": secondary_labels,
+        "type": chirp_types,
+        "latitude": latitude,
+        "longitude": longitude,
+        "scientific_name": scientific_name,
+        "common_name": common_name,
+        "rating": float(rating),
+        "path": path,
+    }
+
+    return data

audyn-0.0.1.dev7/audyn/utils/data/birdclef/birdclef2022/__init__.py

@@ -0,0 +1,67 @@
+import csv
+from typing import List, Tuple
+
+import torch
+
+from .._common import decode_csv_line
+from ._download import download_birdclef2022_primary_labels
+
+__all__ = [
+    "primary_labels",
+    "num_primary_labels",
+    "stratified_split",
+    "decode_csv_line",  # for compatibility with birdclef2024
+]
+
+primary_labels = download_birdclef2022_primary_labels()
+num_primary_labels = len(primary_labels)
+
+
+def stratified_split(
+    path: str,
+    train_ratio: float,
+    seed: int = 0,
+) -> Tuple[List[str], List[str]]:
+    """Split dataset into training and validation.
+
+    Args:
+        path (str): Path to csv file.
+        train_ratio (float): Ratio of training set.
+        seed (int): Random seed.
+
+    Returns:
+        tuple: Splits of filenames.
+
+            - list: List of training filenames.
+            - list: List of validation filenames.
+
+    """
+    g = torch.Generator()
+    g.manual_seed(seed)
+
+    filenames = {primary_label: [] for primary_label in primary_labels}
+    train_filenames = []
+    validation_filenames = []
+
+    with open(path) as f:
+        reader = csv.reader(f)
+
+        for idx, line in enumerate(reader):
+            if idx < 1:
+                continue
+
+            primary_label, *_, filename = line
+            filenames[primary_label].append(filename)
+
+    # split dataset
+    for primary_label, _filenames in filenames.items():
+        num_files = len(_filenames)
+        indices = torch.randperm(num_files, generator=g).tolist()
+
+        for idx in indices[: int(train_ratio * num_files)]:
+            train_filenames.append(_filenames[idx])
+
+        for idx in indices[int(train_ratio * num_files) :]:
+            validation_filenames.append(_filenames[idx])
+
+    return train_filenames, validation_filenames

audyn-0.0.1.dev7/audyn/utils/data/birdclef/birdclef2022/_download.py

@@ -0,0 +1,36 @@
+import os
+from typing import List, Optional
+
+from .....utils import audyn_cache_dir
+from ....github import download_file_from_github_release
+
+__all__ = [
+    "download_birdclef2022_primary_labels",
+]
+
+
+def download_birdclef2022_primary_labels(
+    root: Optional[str] = None, url: Optional[str] = None
+) -> List[str]:
+    filename = "primary-labels.txt"
+
+    if root is None:
+        root = os.path.join(audyn_cache_dir, "data", "birdclef2022")
+
+    if url is None:
+        url = "https://github.com/tky823/Audyn/releases/download/v0.0.1.dev6/birdclef2022-primary-labels.txt"  # noqa: E501
+
+    path = os.path.join(root, filename)
+
+    if not os.path.exists(path):
+        os.makedirs(root, exist_ok=True)
+        download_file_from_github_release(url, path)
+
+    labels = []
+
+    with open(path) as f:
+        for line in f:
+            label = line.strip()
+            labels.append(label)
+
+    return labels

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/utils/data/birdclef/birdclef2024/__init__.py

@@ -1,11 +1,11 @@
-import ast
 import csv
 import glob
 import os
-from typing import Any, Dict, List, Tuple
+from typing import List, Tuple
 
 import torch
 
+from .._common import decode_csv_line
 from ._download import download_birdclef2024_primary_labels
 
 __all__ = [
@@ -13,7 +13,7 @@ __all__ = [
     "num_primary_labels",
     "stratified_split",
     "split",
-    "decode_csv_line",
+    "decode_csv_line",  # for backward compatibility
 ]
 
 primary_labels = download_birdclef2024_primary_labels()
@@ -120,77 +120,3 @@ def split(
     validation_filenames = [filenames[idx] for idx in indices[int(num_files * train_ratio) :]]
 
     return train_filenames, validation_filenames
-
-
-def decode_csv_line(line: List[str]) -> Dict[str, Any]:
-    """Decode line of train_metadata.csv.
-
-    Args:
-        line (list): One line of train_metadata.csv split by comma (,).
-
-    Returns:
-        dict: Dictionary containing metadata of given line.
-
-    .. note::
-
-        Returned dictionary contains following values.
-
-            - filename (str): Filename with out extension. e.g. ``asbfly/XC134896``.
-            - primary_label (str): Primary label of bird species.
-            - secondary_label (list): Secondary labels of bird species.
-            - type (list): Chirp types.
-            - latitude (float, optional): Latitude of recording.
-            - longitude (float, optional): Longitude of recording.
-            - scientific_name (str): Scientific name of bird.
-            - common_name (str): Common name of bird.
-            - rating (float): Rating.
-            - path (str): Path to audio file equivalent to ``filename`` + ``.ogg``.
-                e.g. ``asbfly/XC134896.ogg``.
-
-    """
-    (
-        primary_label,
-        secondary_labels,
-        chirp_types,
-        latitude,
-        longitude,
-        scientific_name,
-        common_name,
-        _,
-        _,
-        rating,
-        _,
-        path,
-    ) = line
-
-    secondary_labels = ast.literal_eval(secondary_labels)
-    chirp_types = ast.literal_eval(chirp_types)
-    secondary_labels = [secondary_label.lower() for secondary_label in secondary_labels]
-    chirp_types = [chirp_type.lower() for chirp_type in chirp_types]
-
-    filename, _ = os.path.splitext(path)
-
-    if len(latitude) > 0:
-        latitude = float(latitude)
-    else:
-        latitude = None
-
-    if len(longitude) > 0:
-        longitude = float(longitude)
-    else:
-        longitude = None
-
-    data = {
-        "filename": filename,
-        "primary_label": primary_label,
-        "secondary_label": secondary_labels,
-        "type": chirp_types,
-        "latitude": latitude,
-        "longitude": longitude,
-        "scientific_name": scientific_name,
-        "common_name": common_name,
-        "rating": float(rating),
-        "path": path,
-    }
-
-    return data

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/utils/data/birdclef/birdclef2024/_download.py

@@ -4,17 +4,21 @@ from typing import List, Optional
 from .....utils import audyn_cache_dir
 from ....github import download_file_from_github_release
 
+__all__ = [
+    "download_birdclef2024_primary_labels",
+]
+
 
 def download_birdclef2024_primary_labels(
     root: Optional[str] = None, url: Optional[str] = None
 ) -> List[str]:
-    filename = "primary_labels.txt"
+    filename = "primary-labels.txt"
 
     if root is None:
         root = os.path.join(audyn_cache_dir, "data", "birdclef2024")
 
     if url is None:
-        url = "https://github.com/tky823/Audyn/releases/download/v0.0.1.dev4/primary_labels.txt"
+        url = "https://github.com/tky823/Audyn/releases/download/v0.0.1.dev6/birdclef2024-primary-labels.txt"  # noqa: E501
 
     path = os.path.join(root, filename)
 

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/utils/data/birdclef/birdclef2024/composer.py

@@ -247,12 +247,13 @@ class BirdCLEF2024AudioComposer(Composer):
 
         if sample_rate != target_sample_rate:
             audio = aF.resample(audio, sample_rate, target_sample_rate)
+            sample_rate = target_sample_rate
             sample[sample_rate_key] = torch.full(
                 (), fill_value=sample_rate, dtype=sample_rate_dtype
             )
 
         if duration is not None:
-            length = int(target_sample_rate * duration)
+            length = int(sample_rate * duration)
             padding = length - audio.size(-1)
 
             if padding > 0:
@@ -284,6 +285,7 @@ class BirdCLEF2024AudioComposer(Composer):
         output = {
             waveform_key: audio,
             melspectrogram_key: melspectrogram,
+            sample_rate_key: sample[sample_rate_key],
             filename_key: sample[filename_key],
         }
 

{audyn-0.0.1.dev6 → audyn-0.0.1.dev7}/audyn/utils/data/composer.py

@@ -1,5 +1,5 @@
 import re
-from typing import Any, Callable, Dict, Iterable
+from typing import Any, Callable, Dict, Iterable, List
 
 import torch
 import torch.nn as nn
@@ -9,6 +9,7 @@ from .webdataset import decode_audio, supported_audio_extensions
 __all__ = [
     "Composer",
     "AudioFeatureExtractionComposer",
+    "SequentialComposer",
 ]
 
 
@@ -108,3 +109,27 @@ class AudioFeatureExtractionComposer(Composer):
         sample[feature_key] = feature
 
         return sample
+
+
+class SequentialComposer(Composer):
+    """Module to apply multiple composers."""
+
+    def __init__(
+        self,
+        *composers,
+        decode_audio_as_waveform: bool = True,
+        decode_audio_as_monoral: bool = True,
+    ) -> None:
+        super().__init__(
+            decode_audio_as_waveform=decode_audio_as_waveform,
+            decode_audio_as_monoral=decode_audio_as_monoral,
+        )
+
+        self.composers: List[Composer] = list(composers)
+
+    def process(self, sample: Dict[str, Any]) -> Dict[str, Any]:
+        """Process to edit each sample."""
+        for composer in self.composers:
+            sample = composer.process(sample)
+
+        return sample