Update beam and pandas dependencies, and add instructions to download…

… and prepare the BIRB evaluation data

PiperOrigin-RevId: 542269331

README.md

@@ -12,8 +12,8 @@ You might need the following dependencies.
 # Install Poetry for package management
 curl -sSL https://install.python-poetry.org | python3 -
 
-# Install dependencies for librosa (required for testing only)
-sudo apt-get install libsndfile1
+# Install dependencies for librosa
+sudo apt-get install libsndfile1 ffmpeg
 
 # Install all dependencies specified in the poetry configs.
 poetry install
@@ -26,4 +26,18 @@ dependencies, in which you can run the Chirp codebase. To run the tests, try
 poetry run python -m unittest discover -s chirp/tests -p "*test.py"
 ```
 
+## BIRB data preparation
+
+### Evaluation data
+
+After [installing](#installation) the `chirp` package, run the following command from the repository's root directory:
+
+```bash
+poetry run tfds build -i chirp.data.bird_taxonomy,chirp.data.soundscapes \
+    soundscapes/{ssw,hawaii,coffee_farms,sierras_kahl,high_sierras,peru}_full_length \
+    bird_taxonomy/{downstream_full_length,class_representatives_slice_peaked}
+```
+
+The process should take 36 to 48 hours to complete and use around 256 GiB of disk space.
+
 *This is not an officially supported Google product.*

chirp/data/bird_taxonomy/bird_taxonomy.py

@@ -50,6 +50,10 @@
 be retrieved from the 'filename' feature: 'XC{xeno_canto_id}.mp3'.
 """
 
+# The maximum audio sequence length to consider if a localization function is
+# provided. This is 5 * 60 seconds = 5 minutes.
+_MAX_LOCALIZATION_LENGTH_S = 5 * 60
+
 LocalizationFn = Callable[[Any, int, float], jnp.ndarray]
 
 
@@ -335,6 +339,12 @@ def _info(self) -> tfds.core.DatasetInfo:
     )
 
   def _split_generators(self, dl_manager: tfds.download.DownloadManager):
+    Increase the file handle resource soft limit to the hard limit. The
+    dataset is large enough that it causes TFDS to hit the soft limit.
+    import resource
+    _low, _high = resource.getrlimit(resource.RLIMIT_NOFILE)
+    resource.setrlimit(resource.RLIMIT_NOFILE, (_high, high))
+
     # No checksum is found for the new taxonomy_info. dl_manager may raise
     # an error when removing the line below.
     dl_manager._force_checksums_validation = (
@@ -435,6 +445,9 @@ def _process_example(row):
           # Resampling can introduce artifacts that push the signal outside the
           # [-1, 1) interval.
           audio = np.clip(audio, -1.0, 1.0 - (1.0 / float(1 << 15)))
+      # Skip empty audio files.
+      if audio.shape[0] == 0 or np.max(np.abs(audio)) == 0.0:
+        return None
       # The scrubbed foreground annotations are replaced by ''. When this is the
       # case, we translate this annotation into []  rather than [''].
       foreground_label = (
@@ -463,19 +476,24 @@ def _process_example(row):
           'sound_type': source['sound_type'],
       }
 
-    pipeline = beam.Create(source_info.iterrows()) | beam.Map(_process_example)
-
     if self.builder_config.localization_fn:
 
-      def _localize_intervals(args):
+      def localize_intervals_fn(args):
         key, example = args
         sample_rate_hz = self.builder_config.sample_rate_hz
         interval_length_s = self.builder_config.interval_length_s
         target_length = int(sample_rate_hz * interval_length_s)
 
-        audio = audio_utils.pad_to_length_if_shorter(
-            example['audio'], target_length
-        )
+        audio = example['audio']
+
+        # We limit audio sequence length to _MAX_LOCALIZATION_LENGTH_S when
+        # localizing intervals because the localization function can result in
+        # very large memory consumption for long audio sequences.
+        max_length = sample_rate_hz * _MAX_LOCALIZATION_LENGTH_S
+        if audio.shape[0] > max_length:
+          audio = audio[:max_length]
+
+        audio = audio_utils.pad_to_length_if_shorter(audio, target_length)
         # Pass padded audio to avoid localization_fn having to pad again
         audio_intervals = self.builder_config.localization_fn(
             audio, sample_rate_hz, interval_length_s
@@ -499,6 +517,23 @@ def _localize_intervals(args):
           ))
         return interval_examples
 
-      pipeline = pipeline | beam.FlatMap(_localize_intervals)
-
-    return pipeline
+    else:
+      localize_intervals_fn = None
+
+    for i, key_and_example in enumerate(
+        map(_process_example, source_info.iterrows())
+    ):
+      # Since the audio files have variable length, the JAX compilation cache
+      # can use up a large amount of memory after a while.
+      if i % 100 == 0:
+        jax.clear_caches()
+
+      # Skip empty audio files.
+      if key_and_example is None:
+        continue
+
+      if localize_intervals_fn:
+        for key_and_example in localize_intervals_fn(key_and_example):
+          yield key_and_example
+      else:
+        yield key_and_example

chirp/data/filter_scrub_utils.py

@@ -314,7 +314,7 @@ def is_not_in(
 def append(df: pd.DataFrame, row: dict[str, Any]):
   if set(row.keys()) != set(df.columns):
     raise ValueError
-  new_df = df.append(row, ignore_index=True)
+  new_df = pd.concat([df, pd.DataFrame([row])], ignore_index=True)
   return new_df
 
 

chirp/data/soundscapes/soundscapes.py

@@ -604,7 +604,6 @@ def _process_group(
           beam.metrics.Metrics.counter('soundscapes', 'examples').inc()
         return valid_segments
 
-    pipeline = beam.Create(
-        enumerate(segments.groupby('filename'))
-    ) | beam.FlatMap(_process_group)
-    return pipeline
+    for group in enumerate(segments.groupby('filename')):
+      for key, example in _process_group(group):
+        yield key, example

chirp/models/class_average.py

@@ -30,8 +30,8 @@ class ClassAverage(metrics.Metric):
   form a multi-hot encoding.
   """
 
-  total: jnp.array
-  count: jnp.array
+  total: jnp.ndarray
+  count: jnp.ndarray
 
   @classmethod
   def empty(cls):

chirp/projects/bootstrap/search.py

@@ -104,8 +104,9 @@ def write_labeled_data(self, labeled_data_path: str, sample_rate: int):
       for label in labels:
         output_path = labeled_data_path / label
         output_path.mkdir(parents=True, exist_ok=True)
-        output_filepath = output_path / output_filename
-        wavfile.write(output_filepath, sample_rate, r.audio)
+        output_filepath = epath.Path(output_path / output_filename)
+        with output_filepath.open('wb') as f:
+          wavfile.write(f, sample_rate, r.audio)
         counts[label] += 1
     for label, count in counts.items():
       print(f'Wrote {count} examples for label {label}')

chirp/projects/multicluster/data_lib.py

@@ -179,14 +179,15 @@ def embed_dataset(
     exclude_classes: Sequence[str] = (),
     load_audio: bool = True,
     target_sample_rate: int = -1,
+    audio_file_pattern: str = '*.wav',
 ) -> Tuple[Sequence[str], Dict[str, np.ndarray]]:
   """Add embeddings to an eval dataset.
 
   Embed a dataset, creating an in-memory copy of all data with embeddings added.
   The base_dir should contain folders corresponding to classes, and each
-  sub-folder should contina wav files for the respective class.
+  sub-folder should contina audio files for the respective class.
 
-  Note that any wav files in the base_dir directly will be ignored.
+  Note that any audio files in the base_dir directly will be ignored.
 
   Args:
     base_dir: Directory contianing audio data.
@@ -196,6 +197,8 @@ def embed_dataset(
     load_audio: Whether to load audio into memory.
     target_sample_rate: Resample loaded audio to this sample rate. If -1, loads
       raw audio with no resampling. If -2, uses the embedding_model sample rate.
+    audio_file_pattern: The glob pattern to use for finding audio files within
+      the sub-folders.
 
   Returns:
     Ordered labels and a Dict contianing the entire embedded dataset.
@@ -224,7 +227,18 @@ def embed_dataset(
   for label_idx, label in enumerate(labels):
     label_hot = np.zeros([len(labels)], np.int32)
     label_hot[label_idx] = 1
-    filepaths = [fp.as_posix() for fp in (base_dir / label).glob('*.wav')]
+
+    filepaths = [
+        fp.as_posix() for fp in (base_dir / label).glob(audio_file_pattern)
+    ]
+
+    if not filepaths:
+      raise ValueError(
+          'No files matching {} were found in directory {}'.format(
+              audio_file_pattern, base_dir / label
+          )
+      )
+
     audio_iterator = audio_utils.multi_load_audio_window(
         filepaths, None, target_sample_rate, -1
     )
@@ -235,7 +249,18 @@ def embed_dataset(
         audio = _pad_audio(audio, window_size)
       audio = audio.astype(np.float32)
       outputs = embedding_model.embed(audio)
-      embeds = outputs.pooled_embeddings(time_pooling, 'squeeze')
+
+      if not outputs.embeddings:
+        raise ValueError('Embedding model did not produce any embeddings!')
+
+      # If the audio was separated then the raw audio is in the first channel.
+      # Embedding shape is either [B, F, C, D] or [F, C, D] so channel is
+      # always -2.
+      channel_pooling = (
+          'squeeze' if outputs.embeddings.shape[-2] == 1 else 'first'
+      )
+
+      embeds = outputs.pooled_embeddings(time_pooling, channel_pooling)
       merged['embeddings'].append(embeds)
 
       filename = epath.Path(fp).name

chirp/projects/sfda/mca.py

@@ -44,7 +44,7 @@ def empty(cls) -> "MCA":
 
   @classmethod
   def from_model_output(
-      cls, scores: jnp.array, label: jnp.array, **_
+      cls, scores: jnp.ndarray, label: jnp.ndarray, **_
   ) -> clu_metrics.Metric:
     num_classes = label.shape[-1]
     if scores.shape[-1] != num_classes:

chirp/projects/sfda/method_utils.py

@@ -43,7 +43,7 @@
 
 
 @jax.jit
-def jax_cdist(features_a: jnp.array, features_b: jnp.array) -> jnp.array:
+def jax_cdist(features_a: jnp.ndarray, features_b: jnp.ndarray) -> jnp.ndarray:
   """A jax equivalent of scipy.spatial.distance.cdist.
 
   Computes the pairwise squared euclidean distance between each pair of features

chirp/projects/sfda/methods/nrc.py

@@ -49,9 +49,9 @@
 class NRCLoss(clu_metrics.Metric):
   """Computes NRC's loss for the standard single-label case."""
 
-  probabilities_sum: jnp.array
-  nn_loss_sum: jnp.array
-  extended_nn_loss_sum: jnp.array
+  probabilities_sum: jnp.ndarray
+  nn_loss_sum: jnp.ndarray
+  extended_nn_loss_sum: jnp.ndarray
   label_mask: jnp.ndarray | None
   n_samples: int
 
@@ -138,10 +138,10 @@ def compute(self):
 class NRCMultiLoss(clu_metrics.Metric):
   """Computes NRC's loss for the multi-label case."""
 
-  probabilities_sum: jnp.array
-  nn_loss_sum: jnp.array
-  extended_nn_loss_sum: jnp.array
-  label_mask: jnp.array
+  probabilities_sum: jnp.ndarray
+  nn_loss_sum: jnp.ndarray
+  extended_nn_loss_sum: jnp.ndarray
+  label_mask: jnp.ndarray
   n_samples: int
 
   @classmethod

chirp/projects/sfda/methods/shot.py

@@ -35,10 +35,10 @@
 class SHOTMultiLabelLoss(clu_metrics.Metric):
   """Computes the loss used in SHOT-full for the multi-label case."""
 
-  probabilities_sum: jnp.array
-  entropy_sum: jnp.array
-  pl_xent_sum: jnp.array
-  label_mask: jnp.array
+  probabilities_sum: jnp.ndarray
+  entropy_sum: jnp.ndarray
+  pl_xent_sum: jnp.ndarray
+  label_mask: jnp.ndarray
   n_samples: int
   beta: float
 
@@ -109,9 +109,9 @@ def compute(self):
 class SHOTLoss(clu_metrics.Metric):
   """Computes the loss used in SHOT-full for the single-label case."""
 
-  probabilities_sum: jnp.array
-  entropy_sum: jnp.array
-  pl_xent_sum: jnp.array
+  probabilities_sum: jnp.ndarray
+  entropy_sum: jnp.ndarray
+  pl_xent_sum: jnp.ndarray
   label_mask: jnp.ndarray | None
   n_samples: int
   beta: float
@@ -121,7 +121,7 @@ def from_model_output(
       cls,
       probabilities: jnp.ndarray,
       pseudo_label: jnp.ndarray,
-      label_mask: jnp.array,
+      label_mask: jnp.ndarray,
       beta: float,
       **_
   ) -> "SHOTLoss":

chirp/projects/sfda/metrics.py

@@ -52,7 +52,7 @@ class MarginalEntropy(clu_metrics.Metric):
   non-`averageable` metric, which is why we dedicate a separate metric for it.
   """
 
-  probability_sum: jnp.array
+  probability_sum: jnp.ndarray
   n_samples: int
   multi_label: bool
   label_mask: jnp.ndarray | None
@@ -87,7 +87,7 @@ def compute(self):
 
   @classmethod
   def empty(cls) -> "MarginalEntropy":
-    return cls(
+    return cls(  # pytype: disable=wrong-arg-types  # jnp-array
         probability_sum=0.0, n_samples=0, multi_label=False, label_mask=None
     )
 
@@ -100,9 +100,9 @@ class MarginalBinaryEntropy(clu_metrics.Metric):
   multi_label.
   """
 
-  probability_sum: jnp.array
+  probability_sum: jnp.ndarray
   n_samples: int
-  label_mask: jnp.array
+  label_mask: jnp.ndarray
   multi_label: bool
 
   @classmethod
@@ -140,6 +140,6 @@ def compute(self):
 
   @classmethod
   def empty(cls) -> "MarginalBinaryEntropy":
-    return cls(
+    return cls(  # pytype: disable=wrong-arg-types  # jnp-array
         probability_sum=0.0, n_samples=0, label_mask=0.0, multi_label=False
     )

chirp/tests/bird_taxonomy_test.py

@@ -76,9 +76,15 @@ def setUpClass(cls):
     subdir = epath.Path(cls.tempdir) / 'audio-data' / 'comter'
     subdir.mkdir(parents=True)
     for i in range(4):
-      tfds.core.lazy_imports.pydub.AudioSegment.silent(duration=10000).export(
-          subdir / f'XC{i:05d}.mp3', format='mp3'
-      )
+      tfds.core.lazy_imports.pydub.AudioSegment(
+          b'\0\1' * int(10_000 * 10),
+          metadata={
+              'channels': 1,
+              'sample_width': 2,
+              'frame_rate': 10_000,
+              'frame_width': 2,
+          },
+      ).export(subdir / f'XC{i:05d}.mp3', format='mp3')
 
   @classmethod
   def tearDownClass(cls):

chirp/tests/filter_scrub_test.py

@@ -335,7 +335,9 @@ def test_append_query(self):
     new_df = fsu.apply_query(self.toy_df, append_query)
     self.assertEqual(
         new_df.to_dict(),
-        self.toy_df.append(new_row, ignore_index=True).to_dict(),
+        pd.concat(
+            [self.toy_df, pd.DataFrame([new_row])], ignore_index=True
+        ).to_dict(),
     )
 
     # Append query with keys not matching the dataframe
@@ -456,7 +458,10 @@ def test_merge_concat_no_duplicates(self):
     # of .to_dict().
     self.assertTrue(
         fsu.apply_parallel(self.toy_df, query_parallel).equals(
-            self.toy_df.loc[[0]].append([scrubbed_r0, self.toy_df.loc[1]])
+            pd.concat([
+                self.toy_df.loc[[0]],
+                pd.DataFrame([scrubbed_r0, self.toy_df.loc[1]]),
+            ])
         )
     )