mirror of
https://github.com/pyannote/pyannote-audio.git
synced 2026-11-20 13:07:51 +09:00
Merge branch 'release/4.0.3'
This commit is contained in:
commit
6328b97b94
5 files changed
+112
-47
No files matched your search
@@ -1,5 +1,12 @@
|
||||
# CHANGELOG
|
||||
|
||||
## Version 4.0.3 (2025-12-07)
|
||||
|
||||
- feat(cli): add `--revision` option to most CLI commands
|
||||
- feat(util): add `Calibration.safe_transform` method (supports NaNs as well as any shape)
|
||||
- fix(model): fix `Model.from_pretrained` to support `lightning` 2.6+
|
||||
- setup: update `pyannote-database` dependency to `6.1+`
|
||||
|
||||
## Version 4.0.2 (2025-11-19)
|
||||
|
||||
- BREAKING(util): make `Binarize.__call__` return `string` tracks (instead of `int`) [@benniekiss](https://github.com/benniekiss/)
|
||||
|
||||
+1
-1
@@ -16,7 +16,7 @@ dependencies = [
|
||||
"opentelemetry-sdk>=1.34.0",
|
||||
"opentelemetry-exporter-otlp>=1.34.0",
|
||||
"pyannote-core>=6.0.1",
|
||||
"pyannote-database>=6.0.0",
|
||||
"pyannote-database>=6.1.1",
|
||||
"pyannote-metrics>=4.0.0",
|
||||
"pyannote-pipeline>=4.0.0",
|
||||
"pytorch-metric-learning>=2.8.1",
|
||||
|
||||
@@ -40,7 +40,7 @@ import pyannote.database
|
||||
import torch
|
||||
import typer
|
||||
import yaml
|
||||
from pyannote.audio import Audio, Pipeline, Model
|
||||
from pyannote.audio import Audio, Model, Pipeline
|
||||
from pyannote.core import Annotation
|
||||
from pyannote.metrics.base import BaseMetric
|
||||
from pyannote.metrics.diarization import DiarizationErrorRate, JaccardErrorRate
|
||||
@@ -285,16 +285,20 @@ def download(
|
||||
help="Pretrained pipeline (e.g. pyannote/speaker-diarization-community-1)"
|
||||
),
|
||||
],
|
||||
token: Annotated[
|
||||
str,
|
||||
typer.Argument(
|
||||
help="Huggingface token to be used for downloading from Huggingface hub."
|
||||
),
|
||||
],
|
||||
cache: Annotated[
|
||||
Path,
|
||||
revision: Annotated[
|
||||
Optional[str],
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface hub are stored.",
|
||||
help="Pretrained pipeline revision.",
|
||||
),
|
||||
] = None,
|
||||
token: Annotated[
|
||||
Optional[str],
|
||||
typer.Argument(help="Huggingface token."),
|
||||
] = None,
|
||||
cache: Annotated[
|
||||
Optional[Path],
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface are stored.",
|
||||
exists=True,
|
||||
dir_okay=True,
|
||||
file_okay=False,
|
||||
@@ -309,7 +313,7 @@ def download(
|
||||
|
||||
# load pretrained pipeline
|
||||
pretrained_pipeline = Pipeline.from_pretrained(
|
||||
pipeline, token=token, cache_dir=cache
|
||||
pipeline, revision=revision, token=token, cache_dir=cache
|
||||
)
|
||||
if pretrained_pipeline is None:
|
||||
print(f"Could not load pretrained pipeline from {pipeline}.")
|
||||
@@ -335,7 +339,7 @@ def apply(
|
||||
),
|
||||
],
|
||||
into: Annotated[
|
||||
Path,
|
||||
Optional[Path],
|
||||
typer.Option(
|
||||
help="Path to file or directory where results are saved.",
|
||||
exists=False,
|
||||
@@ -345,13 +349,20 @@ def apply(
|
||||
resolve_path=True,
|
||||
),
|
||||
] = None,
|
||||
device: Annotated[
|
||||
Device, typer.Option(help="Accelerator to use (CPU, CUDA, MPS)")
|
||||
] = Device.AUTO,
|
||||
cache: Annotated[
|
||||
Path,
|
||||
revision: Annotated[
|
||||
Optional[str],
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface hub are stored.",
|
||||
help="Pretrained pipeline revision.",
|
||||
),
|
||||
] = None,
|
||||
token: Annotated[
|
||||
Optional[str],
|
||||
typer.Argument(help="Huggingface token."),
|
||||
] = None,
|
||||
cache: Annotated[
|
||||
Optional[Path],
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface are stored.",
|
||||
exists=True,
|
||||
dir_okay=True,
|
||||
file_okay=False,
|
||||
@@ -359,13 +370,18 @@ def apply(
|
||||
resolve_path=True,
|
||||
),
|
||||
] = None,
|
||||
device: Annotated[
|
||||
Device, typer.Option(help="Accelerator to use (CPU, CUDA, MPS)")
|
||||
] = Device.AUTO,
|
||||
):
|
||||
"""
|
||||
Apply a pretrained PIPELINE to an AUDIO file or directory
|
||||
"""
|
||||
|
||||
# load pretrained pipeline
|
||||
pretrained_pipeline = Pipeline.from_pretrained(pipeline, cache_dir=cache)
|
||||
pretrained_pipeline = Pipeline.from_pretrained(
|
||||
pipeline, revision=revision, token=token, cache_dir=cache
|
||||
)
|
||||
if pretrained_pipeline is None:
|
||||
print(f"Could not load pretrained pipeline from {pipeline}.")
|
||||
raise typer.exit(code=1)
|
||||
@@ -375,7 +391,6 @@ def apply(
|
||||
pretrained_pipeline.to(torch_device)
|
||||
|
||||
if audio.is_dir():
|
||||
|
||||
if into is None or not into.is_dir():
|
||||
typer.echo("When AUDIO is a directory, INTO must also be a directory.")
|
||||
raise typer.exit(code=1)
|
||||
@@ -385,7 +400,6 @@ def apply(
|
||||
jsons: list[Path | None] = [into / (path.stem + ".json") for path in inputs]
|
||||
|
||||
else:
|
||||
|
||||
if not (into is None or into.is_file()):
|
||||
typer.echo("When AUDIO is a file, INTO must also be a file.")
|
||||
raise typer.exit(code=1)
|
||||
@@ -395,7 +409,6 @@ def apply(
|
||||
jsons: list[Path | None] = [into.with_suffix(".json") if into else None]
|
||||
|
||||
for current_input, current_rttm, current_json in zip(inputs, rttms, jsons):
|
||||
|
||||
prediction = pretrained_pipeline(current_input)
|
||||
|
||||
speaker_diarization = get_diarization(prediction)
|
||||
@@ -522,6 +535,27 @@ def benchmark(
|
||||
case_sensitive=False,
|
||||
),
|
||||
] = Subset.test,
|
||||
revision: Annotated[
|
||||
Optional[str],
|
||||
typer.Option(
|
||||
help="Pretrained pipeline revision.",
|
||||
),
|
||||
] = None,
|
||||
token: Annotated[
|
||||
Optional[str],
|
||||
typer.Argument(help="Huggingface token."),
|
||||
] = None,
|
||||
cache: Annotated[
|
||||
Optional[Path],
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface are stored.",
|
||||
exists=True,
|
||||
dir_okay=True,
|
||||
file_okay=False,
|
||||
writable=True,
|
||||
resolve_path=True,
|
||||
),
|
||||
] = None,
|
||||
device: Annotated[
|
||||
Device, typer.Option(help="Accelerator to use (CPU, CUDA, MPS)")
|
||||
] = Device.AUTO,
|
||||
@@ -538,17 +572,6 @@ def benchmark(
|
||||
num_speakers: Annotated[
|
||||
NumSpeakers, typer.Option(help="Number of speakers (oracle or auto)")
|
||||
] = NumSpeakers.AUTO,
|
||||
cache: Annotated[
|
||||
Path,
|
||||
typer.Option(
|
||||
help="Path to the folder where files downloaded from Huggingface hub are stored.",
|
||||
exists=True,
|
||||
dir_okay=True,
|
||||
file_okay=False,
|
||||
writable=True,
|
||||
resolve_path=True,
|
||||
),
|
||||
] = None,
|
||||
optimize: Annotated[
|
||||
bool,
|
||||
typer.Option(
|
||||
@@ -562,10 +585,7 @@ def benchmark(
|
||||
),
|
||||
] = False,
|
||||
per_file: Annotated[
|
||||
bool,
|
||||
typer.Option(
|
||||
help="Save one RTTM/JSON file per processed audio file."
|
||||
)
|
||||
bool, typer.Option(help="Save one RTTM/JSON file per processed audio file.")
|
||||
] = False,
|
||||
):
|
||||
"""
|
||||
@@ -578,7 +598,9 @@ def benchmark(
|
||||
"""
|
||||
|
||||
# load pretrained pipeline
|
||||
pretrained_pipeline = Pipeline.from_pretrained(pipeline, cache_dir=cache)
|
||||
pretrained_pipeline = Pipeline.from_pretrained(
|
||||
pipeline, revision=revision, token=token, cache_dir=cache,
|
||||
)
|
||||
if pretrained_pipeline is None:
|
||||
print(f"Could not load pretrained pipeline from {pipeline}.")
|
||||
raise typer.exit(code=1)
|
||||
@@ -808,7 +830,9 @@ def benchmark(
|
||||
yaml.dump({"min_duration_off": best_min_duration_off}, yml)
|
||||
|
||||
if not per_file:
|
||||
optimized_rttm_file = into / f"{benchmark_name}.OptimizedMinDurationOff.rttm"
|
||||
optimized_rttm_file = (
|
||||
into / f"{benchmark_name}.OptimizedMinDurationOff.rttm"
|
||||
)
|
||||
|
||||
# make sure we don't overwrite previous results
|
||||
if optimized_rttm_file.exists():
|
||||
@@ -816,7 +840,9 @@ def benchmark(
|
||||
|
||||
for file in files:
|
||||
if per_file:
|
||||
optimized_rttm_file = rttm_dir / f"{file['uri']}.OptimizedMinDurationOff.rttm"
|
||||
optimized_rttm_file = (
|
||||
rttm_dir / f"{file['uri']}.OptimizedMinDurationOff.rttm"
|
||||
)
|
||||
|
||||
with open(optimized_rttm_file, "w" if per_file else "a") as rttm:
|
||||
file["best_speaker_diarization"].write_rttm(rttm)
|
||||
@@ -851,11 +877,11 @@ def strip(
|
||||
"""
|
||||
|
||||
keys = [
|
||||
"pytorch-lightning_version", # * pytorch-lightning needs
|
||||
"hparams_name", # those values to initialize
|
||||
"hyper_parameters", # the model architecture
|
||||
"state_dict", # * actual weights
|
||||
"pyannote.audio", # * pyannote.audio dependencies
|
||||
"pytorch-lightning_version", # * pytorch-lightning needs
|
||||
"hparams_name", # those values to initialize
|
||||
"hyper_parameters", # the model architecture
|
||||
"state_dict", # * actual weights
|
||||
"pyannote.audio", # * pyannote.audio dependencies
|
||||
]
|
||||
|
||||
old_checkpoint = torch.load(
|
||||
|
||||
@@ -40,6 +40,34 @@ class Calibration(IsotonicRegression):
|
||||
def __init__(self):
|
||||
super().__init__(y_min=0.0, y_max=1.0, increasing="auto", out_of_bounds="clip")
|
||||
|
||||
def safe_transform(
|
||||
self,
|
||||
values: np.ndarray,
|
||||
nan_value: float = 2.0,
|
||||
) -> np.ndarray:
|
||||
"""Apply calibration handling NaN values and any shape gracefully
|
||||
|
||||
Parameters
|
||||
----------
|
||||
values : np.ndarray
|
||||
Values to calibrate
|
||||
nan_value : float, optional
|
||||
Value to use in place of NaN values during calibration. Default is 2.0.
|
||||
|
||||
Returns
|
||||
-------
|
||||
calibrated_values : np.ndarray
|
||||
Calibrated values
|
||||
"""
|
||||
# temporarily replace NaN values with `nan_value` so `transform()` does not fail
|
||||
transformed = np.nan_to_num(values.reshape(-1), nan=nan_value)
|
||||
|
||||
# apply calibration
|
||||
transformed: np.ndarray = self.transform(transformed)
|
||||
|
||||
# recover original shape
|
||||
return transformed.reshape(values.shape)
|
||||
|
||||
def save(self, path: str):
|
||||
"""Save fitted calibration to disk
|
||||
|
||||
|
||||
@@ -599,7 +599,9 @@ class Model(lightning.LightningModule):
|
||||
map_location = default_map_location
|
||||
|
||||
# load checkpoint using lightning
|
||||
loaded_checkpoint = pl_load(path_to_model_checkpoint, map_location=map_location)
|
||||
loaded_checkpoint = pl_load(
|
||||
path_to_model_checkpoint, map_location=map_location, weights_only=False
|
||||
)
|
||||
|
||||
# check that the checkpoint is compatible with the current version
|
||||
versions = loaded_checkpoint["pyannote.audio"]["versions"]
|
||||
@@ -620,6 +622,7 @@ class Model(lightning.LightningModule):
|
||||
path_to_model_checkpoint,
|
||||
map_location=map_location,
|
||||
strict=strict,
|
||||
weights_only=False,
|
||||
**kwargs,
|
||||
)
|
||||
except RuntimeError as e:
|
||||
@@ -638,6 +641,7 @@ class Model(lightning.LightningModule):
|
||||
path_to_model_checkpoint,
|
||||
map_location=map_location,
|
||||
strict=False,
|
||||
weights_only=False,
|
||||
**kwargs,
|
||||
)
|
||||
return model
|
||||
|
||||
Reference in new issue
Block a user