Files
pyannote-audio/src/pyannote/audio/tasks/embedding/arcface.py
T
ea18bc70d8 Align audio CI and releases; require Python 3.11 (#2081)
* feat(cli): add --subfolder option to benchmark command (#2063)
* fix(utils): fix `SpeakerDiarizationMixin.optimal_mapping` when `reference` is a mapping
* fix(task): fix `balance` option crashing training with `KeyError: 'metadata'` (#2067)
* fix(io): fix `Audio.crop` raising when an in-memory waveform is cropped up to its very end (#2068)
* fix(model): make Model.default_activation return identity for unsupported tasks
* fix: compute sample duration from preloaded waveform
* fix(pipeline): omit unset runtime options from model references (#2080)

Co-authored-by: abeeranajam31 <128603901+abeeranajam31@users.noreply.github.com>
Co-authored-by: antoinelaurent <antoinelaurent@me.com>
Co-authored-by: Kayvan Zahiri <123409429+Kayvan-Zahiri@users.noreply.github.com>
Co-authored-by: Mohammad Hijjawi <51020382+MohammadHijjawi97@users.noreply.github.com>
2026-10-22 15:07:01 +02:00

124 lines
4.5 KiB
Python
Raw Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# MIT License
#
# Copyright (c) 2020- CNRS
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from __future__ import annotations
from collections.abc import Sequence
import pytorch_metric_learning.losses
from pyannote.database import Protocol
from torch_audiomentations.core.transforms_interface import BaseWaveformTransform
from torchmetrics import Metric
from pyannote.audio.core.task import Task
from .mixins import SupervisedRepresentationLearningTaskMixin
class SupervisedRepresentationLearningWithArcFace(
SupervisedRepresentationLearningTaskMixin,
Task,
):
"""Supervised representation learning with ArcFace loss
Representation learning is the task of ...
Parameters
----------
protocol : Protocol
pyannote.database protocol
duration : float, optional
Chunks duration in seconds. Defaults to two seconds (2.).
min_duration : float, optional
Sample training chunks duration uniformly between `min_duration`
and `duration`. Defaults to `duration` (i.e. fixed length chunks).
num_classes_per_batch : int, optional
Number of classes per batch. Defaults to 32.
num_chunks_per_class : int, optional
Number of chunks per class. Defaults to 1.
margin : float, optional
Margin. Defaults to 28.6.
scale : float, optional
Scale. Defaults to 64.
num_workers : int, optional
Number of workers used for generating training samples.
Defaults to multiprocessing.cpu_count() // 2.
pin_memory : bool, optional
If True, data loaders will copy tensors into CUDA pinned
memory before returning them. See pytorch documentation
for more details. Defaults to False.
augmentation : BaseWaveformTransform, optional
torch_audiomentations waveform transform, used by dataloader
during training.
metric : optional
Validation metric(s). Can be anything supported by torchmetrics.MetricCollection.
Defaults to AUROC (area under the ROC curve).
"""
#  TODO: add a ".metric" property that tells how speaker embedding trained with this approach
#  should be compared. could be a string like "cosine" or "euclidean" or a pdist/cdist-like
#  callable. this ".metric" property should be propagated all the way to Inference (via the model).
def __init__(
self,
protocol: Protocol,
min_duration: float | None = None,
duration: float = 2.0,
num_classes_per_batch: int = 32,
num_chunks_per_class: int = 1,
margin: float = 28.6,
scale: float = 64.0,
num_workers: int | None = None,
pin_memory: bool = False,
augmentation: BaseWaveformTransform | None = None,
metric: Metric | Sequence[Metric] | dict[str, Metric] = None,
):
self.num_chunks_per_class = num_chunks_per_class
self.num_classes_per_batch = num_classes_per_batch
self.margin = margin
self.scale = scale
super().__init__(
protocol,
duration=duration,
min_duration=min_duration,
batch_size=self.batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
augmentation=augmentation,
metric=metric,
)
def setup_loss_func(self):
_, embedding_size = self.model(self.model.example_input_array).shape
self.model.loss_func = pytorch_metric_learning.losses.ArcFaceLoss(
len(self.specifications.classes),
embedding_size,
margin=self.margin,
scale=self.scale,
)