mirror of
https://github.com/pyannote/pyannote-audio.git
synced 2026-11-20 13:07:51 +09:00
feat: add support for batch inference (#2027)
This commit is contained in:
1 parent
5dd8d59148
commit
83a7c04623
2 files changed
+138
-55
No files matched your search
@@ -4,7 +4,7 @@
|
||||
# MIT License
|
||||
#
|
||||
# Copyright (c) 2024-2025 CNRS
|
||||
# Copyright (c) 2025 pyannoteAI
|
||||
# Copyright (c) 2025- pyannoteAI
|
||||
#
|
||||
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
# of this software and associated documentation files (the "Software"), to deal
|
||||
@@ -97,6 +97,7 @@ def parse_device(device: Device) -> torch.device:
|
||||
|
||||
|
||||
def get_diarization(prediction) -> Annotation:
|
||||
|
||||
# if result is an Annotation, assume it is speaker diarization
|
||||
if isinstance(prediction, Annotation):
|
||||
return prediction
|
||||
@@ -603,7 +604,10 @@ def benchmark(
|
||||
|
||||
# load pretrained pipeline
|
||||
pretrained_pipeline = Pipeline.from_pretrained(
|
||||
pipeline, revision=revision, 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}.")
|
||||
@@ -646,10 +650,6 @@ def benchmark(
|
||||
if num_speakers == NumSpeakers.ORACLE:
|
||||
benchmark_name += ".OracleNumSpeakers"
|
||||
|
||||
# used to store processing time and file duration
|
||||
processing_time: dict[str, float] = dict()
|
||||
playing_time: dict[str, float] = dict()
|
||||
|
||||
# used to store raw predictions in JSON format
|
||||
serialized_predictions: dict[str, dict] = dict()
|
||||
|
||||
@@ -676,19 +676,14 @@ def benchmark(
|
||||
if rttm_file.exists():
|
||||
raise FileExistsError(f"{rttm_file} already exists.")
|
||||
|
||||
# iterate over all files in the specified subset
|
||||
for file in track(files, disable=not progress):
|
||||
# gather file metadata
|
||||
uri: str = file["uri"]
|
||||
playing_time[uri] = Audio().get_duration(file)
|
||||
if hasattr(pretrained_pipeline, "apply_batch"):
|
||||
iterator = pretrained_pipeline(files, progress=progress)
|
||||
else:
|
||||
iterator = track(pretrained_pipeline(files), disable=not progress)
|
||||
|
||||
tic: float = time.time()
|
||||
|
||||
# apply pretrained pipeline to file
|
||||
prediction = pretrained_pipeline(file, **file.get("pipeline_kwargs", {}))
|
||||
|
||||
tac: float = time.time()
|
||||
processing_time[uri] = tac - tic
|
||||
tic: float = time.time()
|
||||
for file, prediction in iterator:
|
||||
uri = file["uri"]
|
||||
|
||||
# if prediction has a built-in serialize method, save serialized version
|
||||
if hasattr(prediction, "serialize"):
|
||||
@@ -731,6 +726,8 @@ def benchmark(
|
||||
if optimize:
|
||||
file["speaker_diarization"] = speaker_diarization
|
||||
|
||||
tac: float = time.time()
|
||||
|
||||
# save serialized predictions to disk (might contain more than just diarization results)
|
||||
if serialized_predictions and not per_file:
|
||||
with open(into / f"{benchmark_name}.json", "w") as f:
|
||||
@@ -738,8 +735,8 @@ def benchmark(
|
||||
|
||||
# log processing time and capacity
|
||||
processing = dict()
|
||||
total_processing_time: float = sum(processing_time.values())
|
||||
total_playing_time: float = sum(playing_time.values())
|
||||
total_processing_time = tac - tic
|
||||
total_playing_time = sum(Audio().get_duration(file) for file in files)
|
||||
processing["seconds_per_hour"] = total_processing_time / (total_playing_time / 3600)
|
||||
processing["times_faster_than_realtime"] = (
|
||||
total_playing_time / total_processing_time
|
||||
|
||||
@@ -22,13 +22,14 @@
|
||||
# SOFTWARE.
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import warnings
|
||||
from collections import OrderedDict
|
||||
from collections.abc import Iterator
|
||||
from functools import partial
|
||||
from pathlib import Path
|
||||
from typing import Callable, Dict, List, Optional
|
||||
from typing import Any, Callable, Dict, List, Mapping, Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
@@ -90,7 +91,9 @@ def expand_subfolders(
|
||||
revision = parent_revision
|
||||
|
||||
if parent_subfolder:
|
||||
subfolder = f"{parent_subfolder.rstrip('/')}/{subfolder.lstrip('/')}"
|
||||
subfolder = (
|
||||
f"{parent_subfolder.rstrip('/')}/{subfolder.lstrip('/')}"
|
||||
)
|
||||
|
||||
config[key] = {
|
||||
"checkpoint": model_id,
|
||||
@@ -122,7 +125,9 @@ def expand_subfolders(
|
||||
revision = parent_revision
|
||||
|
||||
if parent_subfolder:
|
||||
subfolder = f"{parent_subfolder.rstrip('/')}/{subfolder.lstrip('/')}"
|
||||
subfolder = (
|
||||
f"{parent_subfolder.rstrip('/')}/{subfolder.lstrip('/')}"
|
||||
)
|
||||
|
||||
config[idx] = {
|
||||
"checkpoint": model_id,
|
||||
@@ -435,45 +440,25 @@ class Pipeline(_Pipeline):
|
||||
"""
|
||||
raise NotImplementedError()
|
||||
|
||||
def __call__(self, file: AudioFile, preload: bool = False, **kwargs):
|
||||
"""Validate file, (optionally) load it in memory, then process it
|
||||
def prepare_one(self, file: AudioFile, preload: bool = False) -> Mapping:
|
||||
"""Prepare one file for being processed by the pipeline
|
||||
|
||||
1. Validate file
|
||||
2. Add pipeline preprocessors if any
|
||||
3. Preload in memory if requested
|
||||
|
||||
Parameters
|
||||
----------
|
||||
file : AudioFile
|
||||
File to process
|
||||
preload : bool, optional
|
||||
Whether to preload waveform before applying the pipeline.
|
||||
kwargs : keyword arguments, optional
|
||||
Additional keyword arguments passed to `self.apply(...)`
|
||||
preload : bool
|
||||
Whether to preload in memory
|
||||
|
||||
Returns
|
||||
-------
|
||||
output : Any
|
||||
Whatever `self.apply(...)` returns
|
||||
file : Mapping
|
||||
Prepared file
|
||||
"""
|
||||
fix_reproducibility(getattr(self, "device", torch.device("cpu")))
|
||||
|
||||
if not self.instantiated:
|
||||
# instantiate with default parameters when available
|
||||
try:
|
||||
default_parameters = self.default_parameters()
|
||||
except NotImplementedError:
|
||||
raise RuntimeError(
|
||||
"A pipeline must be instantiated with `pipeline.instantiate(parameters)` before it can be applied."
|
||||
)
|
||||
|
||||
try:
|
||||
self.instantiate(default_parameters)
|
||||
except ValueError:
|
||||
raise RuntimeError(
|
||||
"A pipeline must be instantiated with `pipeline.instantiate(parameters)` before it can be applied. "
|
||||
"Tried to use parameters provided by `pipeline.default_parameters()` but those are not compatible. "
|
||||
)
|
||||
|
||||
warnings.warn(
|
||||
f"The pipeline has been automatically instantiated with {default_parameters}."
|
||||
)
|
||||
|
||||
file = Audio.validate_file(file)
|
||||
|
||||
@@ -499,11 +484,110 @@ class Pipeline(_Pipeline):
|
||||
# therefore we remove the `channel` key from the file
|
||||
file.pop("channel", None)
|
||||
|
||||
return file
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
file: AudioFile | list[AudioFile],
|
||||
preload: bool = False,
|
||||
**kwargs,
|
||||
) -> Any | Iterator[tuple[str, Any]]:
|
||||
"""Validate file(s), (optionally) load it/them in memory, then process it/them
|
||||
|
||||
Parameters
|
||||
----------
|
||||
file : AudioFile or list[AudioFile]
|
||||
File to process, or a list of AudioFiles for batch processing.
|
||||
All files in the list must have distinct URIs.
|
||||
Files are processed in sequence unless the pipeline defines an
|
||||
`apply_batch` method, which is called instead.
|
||||
preload : bool, optional
|
||||
Whether to preload waveform before applying the pipeline.
|
||||
Ignored for batch input when `apply_batch` is available.
|
||||
kwargs : keyword arguments, optional
|
||||
Additional keyword arguments passed to `self.apply(...)` or
|
||||
`self.apply_batch(...)`
|
||||
|
||||
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.
|
||||
|
||||
Yields (for multiple files)
|
||||
---------------------------
|
||||
file : AudioFile
|
||||
output : Any
|
||||
|
||||
"""
|
||||
fix_reproducibility(getattr(self, "device", torch.device("cpu")))
|
||||
|
||||
if not self.instantiated:
|
||||
# instantiate with default parameters when available
|
||||
try:
|
||||
default_parameters = self.default_parameters()
|
||||
except NotImplementedError:
|
||||
raise RuntimeError(
|
||||
"A pipeline must be instantiated with `pipeline.instantiate(parameters)` before it can be applied."
|
||||
)
|
||||
|
||||
try:
|
||||
self.instantiate(default_parameters)
|
||||
except ValueError:
|
||||
raise RuntimeError(
|
||||
"A pipeline must be instantiated with `pipeline.instantiate(parameters)` before it can be applied. "
|
||||
"Tried to use parameters provided by `pipeline.default_parameters()` but those are not compatible. "
|
||||
)
|
||||
|
||||
warnings.warn(
|
||||
f"The pipeline has been automatically instantiated with {default_parameters}."
|
||||
)
|
||||
|
||||
if isinstance(file, list):
|
||||
# prepare every file
|
||||
files = [self.prepare_one(f, preload=preload) for f in file]
|
||||
|
||||
# check for duplicate uris (and complain if any)
|
||||
uris = [f["uri"] for f in files]
|
||||
if len(uris) != len(set(uris)):
|
||||
seen: set[str] = set()
|
||||
duplicates = [u for u in uris if u in seen or seen.add(u)]
|
||||
raise ValueError(
|
||||
f"All files in a batch must have distinct URIs. "
|
||||
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
|
||||
|
||||
file = self.prepare_one(file, preload=preload)
|
||||
prediction = self.apply(file, **file.get("pipeline_kwargs", {}), **kwargs)
|
||||
|
||||
# send file duration to telemetry as well as
|
||||
# requested number of speakers in case of diarization
|
||||
track_pipeline_apply(self, file, **kwargs)
|
||||
|
||||
return self.apply(file, **kwargs)
|
||||
return prediction
|
||||
|
||||
def to(self, device: torch.device) -> Pipeline:
|
||||
"""Send pipeline to `device`"""
|
||||
@@ -535,5 +619,7 @@ class Pipeline(_Pipeline):
|
||||
return self.to(torch.device("cuda", device))
|
||||
else:
|
||||
if device.type != "cuda":
|
||||
raise ValueError("Expected CUDA device. Use `Pipeline.to(device)` for other devices.")
|
||||
raise ValueError(
|
||||
"Expected CUDA device. Use `Pipeline.to(device)` for other devices."
|
||||
)
|
||||
return self.to(device)
|
||||
Reference in new issue
Block a user