Merge branch 'release/4.0.7'

This commit is contained in:
Hervé BREDIN committed 2026-06-30 16:27:43 +02:00
commit b749285c5c
4 files changed
+60 -31

No files matched your search

+3 -2
View File
@@ -1,8 +1,9 @@
# CHANGELOG
## next
## Version 4.0.7 (2026-06-30)
- improve(pipeline): add support for `subfolder` in `Pipeline.from_pretrained`
- fix(pipeline): add missing support for `subfolder` in `Pipeline.from_pretrained`
- feat(pipeline): add basic support for (sequential) batch inference
## Version 4.0.5 (2026-06-22)
+29
View File
@@ -43,6 +43,7 @@ try:
import torchcodec
from torchcodec import AudioSamples
from torchcodec.decoders import AudioDecoder, AudioStreamMetadata
TORCHCODEC_AVAILABLE = True
except Exception as e:
warnings.warn(
"\ntorchcodec is not installed correctly so built-in audio decoding will fail. Solutions are:\n"
@@ -50,6 +51,9 @@ except Exception as e:
"* fix torchcodec installation. Error message was:\n\n"
f"{e}"
)
TORCHCODEC_AVAILABLE = False
AudioDecoder = None
AudioStreamMetadata = None
AudioFile = str | Path | IOBase | Mapping
@@ -81,8 +85,19 @@ def get_audio_metadata(file: AudioFile) -> "AudioStreamMetadata":
-------
metadata : AudioStreamMetadata
Audio file metadata
Raises
------
RuntimeError if torchcodec is not available.
"""
if not TORCHCODEC_AVAILABLE:
raise RuntimeError(
"torchcodec is not available. Cannot read audio metadata. "
"Please install torchcodec or provide audio as a waveform dictionary: "
"{'waveform': (channel, time) torch.Tensor, 'sample_rate': int}"
)
metadata = AudioDecoder(file["audio"]).metadata
# rewind if needed
@@ -316,6 +331,13 @@ class Audio:
return self.downmix_and_resample(waveform, sample_rate, channel=channel)
if not TORCHCODEC_AVAILABLE:
raise RuntimeError(
"torchcodec is not available. Cannot read audio file. "
"Please install torchcodec or provide audio as a waveform dictionary: "
"{'waveform': (channel, time) torch.Tensor, 'sample_rate': int}"
)
decoder = AudioDecoder(file["audio"])
samples: AudioSamples = decoder.get_all_samples()
@@ -391,6 +413,13 @@ class Audio:
data = F.pad(data, (pad_start, pad_end))
return self.downmix_and_resample(data, sample_rate, channel=channel)
if not TORCHCODEC_AVAILABLE:
raise RuntimeError(
"torchcodec is not available. Cannot crop audio file. "
"Please install torchcodec or provide audio as a waveform dictionary: "
"{'waveform': (channel, time) torch.Tensor, 'sample_rate': int}"
)
decoder = AudioDecoder(file["audio"])
metadata: AudioStreamMetadata = decoder.metadata
+27 -28
View File
@@ -486,12 +486,33 @@ class Pipeline(_Pipeline):
return file
def _apply_batch(
self,
files: list,
**kwargs,
) -> Iterator[tuple[Any, Any]]:
"""Process a list of prepared files, yielding (file, prediction) pairs"""
# if pipeline natively supports batch processing, use it
if hasattr(self, "apply_batch"):
# TODO: warn user that pipeline_kwargs might not be taken into account.
for f, prediction in self.apply_batch(files, **kwargs):
track_pipeline_apply(self, f, **kwargs)
yield f, prediction
# otherwise process files sequentially
else:
for f in files:
prediction = self.apply(f, **f.get("pipeline_kwargs", {}), **kwargs)
track_pipeline_apply(self, f, **kwargs)
yield f, prediction
def __call__(
self,
file: AudioFile | list[AudioFile],
preload: bool = False,
**kwargs,
) -> Any | Iterator[tuple[str, Any]]:
) -> Any | Iterator[tuple[Any, Any]]:
"""Validate file(s), (optionally) load it/them in memory, then process it/them
Parameters
@@ -511,13 +532,12 @@ class Pipeline(_Pipeline):
Returns (for single file)
-------------------------
output : Any
Whatever `self.apply(...)` returns for a single file, or a dict
mapping each URI to its output for batch input.
Whatever `self.apply(...)` returns for a single file.
Yields (for multiple files)
Returns (for list of files)
---------------------------
file : AudioFile
output : Any
predictions : Iterator[tuple[AudioFile, Any]]
Iterator of (file, prediction) pairs.
"""
fix_reproducibility(getattr(self, "device", torch.device("cpu")))
@@ -557,28 +577,7 @@ class Pipeline(_Pipeline):
f"Duplicate URIs: {duplicates}"
)
# if pipeline natively supports batch processing, use it
if hasattr(self, "apply_batch"):
# TODO: warn user that pipeline_kwargs might not be taken into account.
for f, prediction in self.apply_batch(files, **kwargs):
# send file duration to telemetry as well as
# requested number of speakers in case of diarization
track_pipeline_apply(self, f, **kwargs)
yield f, prediction
# otherwise process files sequentially
else:
for f in files:
prediction = self.apply(f, **f.get("pipeline_kwargs", {}), **kwargs)
# send file duration to telemetry as well as
# requested number of speakers in case of diarization
track_pipeline_apply(self, f, **kwargs)
yield f, prediction
return
return self._apply_batch(files, **kwargs)
file = self.prepare_one(file, preload=preload)
prediction = self.apply(file, **file.get("pipeline_kwargs", {}), **kwargs)
+1 -1
View File
@@ -334,7 +334,7 @@ class Peak:
alpha: float = 0.5,
min_duration: float = 1.0,
):
super(Peak, self).__init__()
super().__init__()
self.alpha = alpha
self.min_duration = min_duration