feat: add Hydra-based "pyannote-audio-train" CLI

* closes #476 (CLI)
* closes #514 (support for AutoLR)
* closes #485 (hyper-parameter optimization with Hydra’s Ax sweeper)
* closes #412 (log graph to Tensorboard)
This commit is contained in:
Hervé BREDIN authored and GitHub committed 2020-11-18 10:59:39 +01:00
1 parent 298ba39b5e
commit 06d3b6116c
28 files changed
+449 -46

No files matched your search

+4
View File
@@ -0,0 +1,4 @@
recursive-include pyannote *.py
recursive-include pyannote *.yaml
global-exclude *.pyc
global-exclude __pycache__
+10 -2
View File
@@ -1,6 +1,14 @@
# pyannote-audio-v2
Rewriting pyannote.audio from scratch
## CLI
```bash
pyannote-audio-train task=vad model=debug protocol=Debug.SpeakerDiarization.Debug
```
## Contributing
The commands below will setup pre-commit hooks and packages needed for developing the `pyannote.audio` library.
@@ -12,8 +20,8 @@ pre-commit install
## Testing
Tests rely on a set of debugging files available in [`test/data`](test/data) directory.
Set `PYANNOTE_DATABASE_CONFIG` environment variable to `test/data/database.yml` before running tests:
Tests rely on a set of debugging files available in [`test/data`](test/data) directory.
Set `PYANNOTE_DATABASE_CONFIG` environment variable to `test/data/database.yml` before running tests:
```bash
PYANNOTE_DATABASE_CONFIG=tests/data/database.yml pytest
+1 -1
View File
@@ -147,7 +147,7 @@
"outputs": [],
"source": [
"from pyannote.audio.tasks.speaker_verification.task import SpeakerEmbeddingArcFace\n",
"emb = SpeakerEmbeddingArcFace(protocol, duration=2., batch_size=32, num_workers=4)\n",
"emb = SpeakerEmbeddingArcFace(protocol, duration=2., num_workers=4)\n",
"model = SimpleEmbeddingModel(task=emb)\n",
"trainer = pl.Trainer(max_epochs=1)\n",
"_ = trainer.fit(model, emb)"
+1 -1
View File
@@ -146,7 +146,7 @@
"outputs": [],
"source": [
"from pyannote.audio.tasks.speaker_verification.task import SpeakerEmbeddingArcFace\n",
"emb = SpeakerEmbeddingArcFace(protocol, duration=2., batch_size=32, num_workers=4)\n",
"emb = SpeakerEmbeddingArcFace(protocol, duration=2., num_workers=4)\n",
"from pyannote.audio.models.debug import SimpleEmbeddingModel\n",
"model = SimpleEmbeddingModel(task=emb)\n",
"trainer = pl.Trainer(max_epochs=1, default_root_dir='inference/emb')\n",
View File
Whitespace-only changes.
+111
View File
@@ -0,0 +1,111 @@
# 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 typing import Iterable
import hydra
from hydra.utils import instantiate
from omegaconf import DictConfig
from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
from pytorch_lightning.loggers import TensorBoardLogger
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.database import FileFinder, get_protocol
@hydra.main(config_path="train_config", config_name="config")
def main(cfg: DictConfig) -> None:
protocol = get_protocol(cfg.protocol, preprocessors={"audio": FileFinder()})
# TODO: configure augmentation
# TODO: configure scheduler
# TODO: configure layer freezing
def optimizer(parameters: Iterable[Parameter], lr: float = 1e-3) -> Optimizer:
return instantiate(cfg.optimizer, parameters, lr=lr)
task = instantiate(
cfg.task,
protocol,
optimizer=optimizer,
learning_rate=cfg.optimizer.lr,
)
model = instantiate(cfg.model, task=task)
monitor, mode = task.validation_monitor
model_checkpoint = ModelCheckpoint(
monitor=monitor,
mode=mode,
save_top_k=10,
period=1,
save_last=True,
save_weights_only=False,
dirpath=".",
filename=f"{{epoch}}-{{{monitor}:.3f}}",
verbose=cfg.verbose,
)
early_stopping = EarlyStopping(
monitor=monitor,
mode=mode,
min_delta=0.0,
patience=10,
strict=True,
verbose=cfg.verbose,
)
logger = TensorBoardLogger(
".",
name="",
version="",
log_graph=True,
)
trainer = instantiate(
cfg.trainer,
callbacks=[model_checkpoint, early_stopping],
logger=logger,
)
if cfg.trainer.auto_lr_find == True:
#  HACK: these two lines below should be removed once
#  the corresponding bug is fixed in pytorch-lighting.
#  https://github.com/pyannote/pyannote-audio/issues/514
task.setup(stage="fit")
model.setup(stage="fit")
trainer.tune(model, task)
trainer.fit(model, task)
best_monitor = float(early_stopping.best_score)
if mode == "min":
return best_monitor
else:
return -best_monitor
if __name__ == "__main__":
main()
Whitespace-only changes.
@@ -0,0 +1,10 @@
protocol: ???
verbose: False
defaults:
- task: vad
- model: debug
- optimizer: adam
- trainer: default
- hydra: train
@@ -0,0 +1,80 @@
# @package _group_
run:
# dir: train/${now:%Y-%m-%d}/${now:%H-%M-%S}
dir: ${protocol}/${task._target_}/${now:%Y-%m-%d}/${now:%H-%M-%S}
sweep:
dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}/${protocol}/${task._target_}
subdir: ${hydra.job.num}
output_subdir: ""
help:
app_name: pyannote-audio-train
# Help header, customize to describe your app to your users
header: == ${hydra.help.app_name} ==
footer: |-
Powered by Hydra (https://hydra.cc)
Use --hydra-help to view Hydra specific help
template: |-
${hydra.help.header}
pyannote-audio-train protocol={protocol_name} task={task} model={model}
{task} can be any of the following:
* vad (default) = voice activity detection
* scd = speaker change detection
* osd = overlapped speech detection
* xseg = multi-task segmentation
{model} can be any of the following:
* debug (default) = simple segmentation model for debugging purposes
{optimizer} can be any of the following
* adam (default) = Adam optimizer
{trainer} can be any of the following
* fast_dev_run for debugging
* default (default) for training the model
Options
=======
Here, we describe the most common options: use "--cfg job" option to get a complete list.
* task.duration: audio chunk duration (in seconds)
* task.batch_size: number of audio chunks per batch
* task.num_workers: number of workers used for generating training chunks
* optimizer.lr: learning rate
* trainer.auto_lr_find: use pytorch-lightning AutoLR
Hyper-parameter optimization
============================
Because it is powered by Hydra (https://hydra.cc), one can run grid search using the --multirun option.
For instance, the following command will run the same job three times, with three different learning rates:
pyannote-audio-train --multirun protocol={protocol_name} task={task} optimizer.lr=1e-3,1e-2,1e-1
Even better, one can use Ax (https://ax.dev) sweeper to optimize learning rate directly:
pyannote-audio-train --multirun hydra/sweeper=ax protocol={protocol_name} task={task} optimizer.lr="interval(1e-3, 1e-1)"
See https://hydra.cc/docs/plugins/ax_sweeper for more details.
User-defined task or model
==========================
1. define your_package.YourTask (or your_package.YourModel) class
2. create file /path/to/your_config/task/your_task.yaml (or /path/to/your_config/model/your_model.yaml)
# @package _group_
_target_: your_package.YourTask # or YourModel
param1: value1
param2: value2
3. call pyannote-audio-train --config-dir /path/to/your_config task=your_task task.param1=modified_value1 model=your_model ...
${hydra.help.footer}
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.models.PyanNet
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.models.debug.SimpleSegmentationModel
@@ -0,0 +1,7 @@
# @package _group_
_target_: torch.optim.Adam
lr: 1e-3
betas: [0.9, 0.999]
eps: 1e-08
weight_decay: 0
amsgrad: False
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.tasks.OverlappedSpeechDetection
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.tasks.SpeakerChangeDetection
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.tasks.VoiceActivityDetection
@@ -0,0 +1,2 @@
# @package _group_
_target_: pyannote.audio.tasks.MultiTaskSegmentation
@@ -0,0 +1,45 @@
# @package _group_
_target_: pytorch_lightning.Trainer
#accelerator: None
accumulate_grad_batches: 1
amp_backend: 'native'
amp_level: 'O2'
auto_lr_find: False
auto_scale_batch_size: False
auto_select_gpus: True
benchmark: True
check_val_every_n_epoch: 1
checkpoint_callback: True
deterministic: False
fast_dev_run: False
flush_logs_every_n_steps: 100
#gpus: None
gradient_clip_val: 0
limit_test_batches: 1.0
limit_train_batches: 1.0
limit_val_batches: 1.0
log_every_n_steps: 50
#log_gpu_memory: None
max_epochs: 1000
#max_steps: None
min_epochs: 1
#min_steps: None
num_nodes: 1
num_processes: 1
num_sanity_val_steps: 2
overfit_batches: 0.0
precision: 32
prepare_data_per_node: True
process_position: 0
#profiler: None
progress_bar_refresh_rate: 1
reload_dataloaders_every_epoch: False
replace_sampler_ddp: True
sync_batchnorm: False
terminate_on_nan: False
#tpu_cores: None
track_grad_norm: -1
#truncated_bptt_steps: None
val_check_interval: 1.0
#weights_save_path: None
weights_summary: 'top'
@@ -0,0 +1,3 @@
# @package _group_
_target_: pytorch_lightning.Trainer
fast_dev_run: True
+9
View File
@@ -136,6 +136,11 @@ class Model(pl.LightningModule):
# (e.g. the final classification and activation layers)
pass
#  used by Tensorboard logger to log model graph
@cached_property
def example_input_array(self) -> torch.Tensor:
return self.task.example_input_array
def helper_introspect(
self,
specifications: TaskSpecification,
@@ -299,6 +304,10 @@ class Model(pl.LightningModule):
# so that its dataloader knows how to generate targets
self.task.model_introspection = self.hparams.model_introspection
# this is needed to support pytorch-lightning auto_lr_find feature
# as it expects to find a "learning_rate" entry in model.hparams
self.hparams.learning_rate = self.task.learning_rate
def on_save_checkpoint(self, checkpoint):
#  put everything pyannote.audio-specific under pyannote.audio
+53 -9
View File
@@ -28,12 +28,13 @@ import warnings
from dataclasses import dataclass
from enum import Enum
from functools import cached_property
from typing import TYPE_CHECKING, List, Optional, Text
from typing import TYPE_CHECKING, Callable, Iterable, List, Optional, Text
import pytorch_lightning as pl
import torch
import torch.nn.functional as F
import torch.optim
from torch.nn import Parameter
from torch.optim import Adam, Optimizer
from torch.utils.data import DataLoader, IterableDataset
from pyannote.database import Protocol
@@ -86,21 +87,26 @@ class TrainDataset(IterableDataset):
def __init__(self, task: Task):
super().__init__()
self.task = task
def __iter__(self):
return self.task.train__iter__()
def __len__(self):
return self.task.train__len__()
class ValDataset(IterableDataset):
def __init__(self, task: Task):
super().__init__()
self.task = task
def __iter__(self):
return self.task.val__iter__()
def __len__(self):
return self.task.val__len__()
class Task(pl.LightningDataModule):
"""Base task class
@@ -122,13 +128,18 @@ class Task(pl.LightningDataModule):
duration : float, optional
Chunks duration. Defaults to variable duration (None).
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
Attributes
----------
@@ -142,9 +153,11 @@ class Task(pl.LightningDataModule):
self,
protocol: Protocol,
duration: float = None,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__()
@@ -172,6 +185,11 @@ class Task(pl.LightningDataModule):
self.pin_memory = pin_memory
if optimizer is None:
optimizer = Adam
self.optimizer = optimizer
self.learning_rate = learning_rate
def prepare_data(self):
"""Use this to download and prepare data
@@ -384,9 +402,35 @@ class Task(pl.LightningDataModule):
model.log("val_loss", loss)
return {"loss": loss}
def parameters(self, model: Model) -> Iterable[Parameter]:
return model.parameters()
# default configure_optimizers provided for convenience
# can obviously be overriden for each task
def configure_optimizers(self, model: Model):
# for tasks such as SpeakerEmbedding,
# other parameters should be added here
return torch.optim.Adam(model.parameters(), lr=1e-3)
# this is needed to support pytorch-lightning auto_lr_find feature
# as it modifies model.hparams.learning_rate and not task.learning_rate.
# in case one does not use auto_lr_find, Model.setup() takes care of
# setting model.hparams.learning_rate to task.learning_rate so we are safe.
lr = model.hparams.learning_rate
return self.optimizer(self.parameters(model), lr=lr)
@property
def validation_monitor(self):
"""Quantity (and direction) to monitor
Useful for model checkpointing or early stopping.
Returns
-------
monitor : str
Name of quantity to monitor.
mode : {'min', 'max}
Minimize
See also
--------
pytorch_lightning.callbacks.ModelCheckpoint
pytorch_lightning.callbacks.EarlyStopping
"""
return "val_loss", "min"
@@ -22,9 +22,11 @@
import random
from typing import Mapping
from typing import Callable, Iterable, Mapping
import numpy as np
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.io import Audio
from pyannote.audio.core.task import Task
@@ -76,13 +78,18 @@ class MultiTaskSegmentation(SegmentationTaskMixin, Task):
Additional osd-specific parameters. Has no effect when `osd` is False.
See OverlappedSpeechDetection docstring for details and default value.
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
def __init__(
@@ -95,9 +102,11 @@ class MultiTaskSegmentation(SegmentationTaskMixin, Task):
scd_params: Mapping = None,
osd: bool = False,
osd_params: Mapping = None,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__(
@@ -106,6 +115,8 @@ class MultiTaskSegmentation(SegmentationTaskMixin, Task):
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
self.vad = vad
@@ -22,8 +22,11 @@
import random
from typing import Callable, Iterable
import numpy as np
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.io import Audio
from pyannote.audio.core.task import Problem, Scale, Task, TaskSpecification
@@ -65,13 +68,18 @@ class OverlappedSpeechDetection(SegmentationTaskMixin, Task):
(i.e. share the same file[domain] value). Default behavior is to not contrain
data augmentation with regards to domain.
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
def __init__(
@@ -82,9 +90,11 @@ class OverlappedSpeechDetection(SegmentationTaskMixin, Task):
snr_min: float = 0.0,
snr_max: float = 10.0,
domain: str = None,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__(
@@ -93,6 +103,8 @@ class OverlappedSpeechDetection(SegmentationTaskMixin, Task):
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
self.specifications = TaskSpecification(
@@ -20,8 +20,12 @@
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from typing import Callable, Iterable
import numpy as np
import scipy.signal
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.task import Problem, Scale, Task, TaskSpecification
from pyannote.audio.tasks.segmentation.mixins import SegmentationTaskMixin
@@ -50,13 +54,18 @@ class SpeakerChangeDetection(SegmentationTaskMixin, Task):
Mark frames less than `collar` frames away from actual change point as positive.
Defaults to 1.
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
def __init__(
@@ -64,9 +73,11 @@ class SpeakerChangeDetection(SegmentationTaskMixin, Task):
protocol: Protocol,
duration: float = 2.0,
collar: int = 1,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__(
@@ -75,6 +86,8 @@ class SpeakerChangeDetection(SegmentationTaskMixin, Task):
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
self.specifications = TaskSpecification(
@@ -20,7 +20,10 @@
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from typing import List, Text
from typing import Callable, Iterable, List, Text
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.task import Problem, Scale, Task, TaskSpecification
from pyannote.audio.tasks.segmentation.mixins import SegmentationTaskMixin
@@ -44,22 +47,29 @@ class SpeakerTracking(SegmentationTaskMixin, Task):
duration : float, optional
Chunks duration. Defaults to 2s.
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
def __init__(
self,
protocol: Protocol,
duration: float = 2.0,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__(
@@ -68,6 +78,8 @@ class SpeakerTracking(SegmentationTaskMixin, Task):
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
# for speaker tracking, task specification depends
@@ -20,8 +20,11 @@
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from typing import Callable, Iterable
import numpy as np
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.task import Problem, Scale, Task, TaskSpecification
from pyannote.audio.tasks.segmentation.mixins import SegmentationTaskMixin
@@ -44,23 +47,29 @@ class VoiceActivityDetection(SegmentationTaskMixin, Task):
duration : float, optional
Chunks duration. Defaults to 2s.
batch_size : int, optional
Number of training samples per batch.
Number of training samples per batch. Defaults to 32.
num_workers : int, optional
Number of workers used for generating training samples.
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.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
def __init__(
self,
protocol: Protocol,
duration: float = 2.0,
batch_size: int = None,
batch_size: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
super().__init__(
@@ -69,6 +78,8 @@ class VoiceActivityDetection(SegmentationTaskMixin, Task):
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
self.specifications = TaskSpecification(
@@ -20,12 +20,16 @@
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from __future__ import annotations
import math
from itertools import chain
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, Callable, Iterable
import pytorch_metric_learning.losses
import torch.optim
from torch.nn import Parameter
from torch.optim import Optimizer
from pyannote.audio.core.task import Problem, Scale, Task, TaskSpecification
from pyannote.audio.utils.random import create_rng_for_worker
@@ -44,6 +48,11 @@ class SpeakerEmbeddingArcFace(Task):
If True, data loaders will copy tensors into CUDA pinned
memory before returning them. See pytorch documentation
for more details. Defaults to False.
optimizer : callable, optional
Callable that takes model parameters as input and returns
an Optimizer instance. Defaults to `torch.optim.Adam`.
learning_rate : float, optional
Learning rate. Defaults to 1e-3.
"""
@@ -56,17 +65,27 @@ class SpeakerEmbeddingArcFace(Task):
self,
protocol: Protocol,
duration: float = 2.0,
batch_size: int = None,
num_chunks_per_speaker: int = 1,
num_speakers_per_batch: int = 32,
num_workers: int = 1,
pin_memory: bool = False,
optimizer: Callable[[Iterable[Parameter]], Optimizer] = None,
learning_rate: float = 1e-3,
):
self.num_chunks_per_speaker = num_chunks_per_speaker
self.num_speakers_per_batch = num_speakers_per_batch
batch_size = self.num_chunks_per_speaker * self.num_speakers_per_batch
super().__init__(
protocol,
duration=duration,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=pin_memory,
optimizer=optimizer,
learning_rate=learning_rate,
)
# there is no such thing as a "class" in representation
@@ -152,8 +171,8 @@ class SpeakerEmbeddingArcFace(Task):
# speaker index in original sorted order
y = self.specifications.classes.index(speaker)
# three chunks per speaker
for _ in range(3):
# multiple chunks per speaker
for _ in range(self.num_chunks_per_speaker):
# select one file at random (with probability proportional to its speaker duration)
file, *_ = rng.choices(
@@ -199,6 +218,5 @@ class SpeakerEmbeddingArcFace(Task):
def val_dataloader(self):
return None
def configure_optimizers(self, model: "Model"):
parameters = chain(model.parameters(), self.loss_func.parameters())
return torch.optim.Adam(parameters, lr=1e-3)
def parameters(self, model: Model) -> Iterable[Parameter]:
return chain(model.parameters(), self.loss_func.parameters())
+1
View File
@@ -1,4 +1,5 @@
einops >= 0.3.0
hydra-core >= 1.0.4
librosa >= 0.8
pyannote.core >= 4.1
pyannote.database >= 4.0.3
+4 -12
View File
@@ -50,20 +50,12 @@ dev =
black>=19.10b0
[options.entry_points]
# Add here console scripts like:
# console_scripts =
# script_name = fastaudio.module:function
# For example:
# console_scripts =
# fibonacci = fastaudio.skeleton:run
# And any other entry points, for example:
console_scripts =
pyannote-audio-train=pyannote.audio.cli.train:main
# pyscaffold.cli =
# awesome = pyscaffoldext.awesome.extension:AwesomeExtension
#console_scripts =
# pyannote-audio=pyannote.audio.applications.pyannote_audio:main
# pyannote-speech-feature=pyannote.audio.applications.feature_extraction:main
#prodigy_recipes =
# pyannote.sad.manual = pyannote.audio.interactive.recipes.sad:sad_manual
# pyannote.dia.binary = pyannote.audio.interactive.recipes.dia:dia_binary