feat: add option to pass arguments to inference (#655)

This commit is contained in:
Hervé BREDIN authored and GitHub committed 2021-04-09 13:29:15 +02:00
1 parent 7f90409756
commit 37a6760738
4 files changed
+28 -42

No files matched your search

@@ -87,16 +87,16 @@ class OverlappedSpeechDetection(Pipeline):
----------
segmentation : Model, str, or dict, optional
Pretrained segmentation (or overlapped speech detection) model.
Defaults to "pyannote/Segmentation-PyanNet-DIHARD".
Defaults to "pyannote/segmentation".
See pyannote.audio.pipelines.utils.get_model for supported format.
batch_size : int, optional
Batch size. Defaults to 32.
precision : float, optional
Optimize recall at target precision.
Defaults to optimize precision/recall fscore.
recall : float, optional
Optimize precision at target recall
Defaults to optimize precision/recall fscore
inference_kwargs : dict, optional
Keywords arguments passed to Inference.
Hyper-parameters
----------------
@@ -110,15 +110,14 @@ class OverlappedSpeechDetection(Pipeline):
def __init__(
self,
segmentation: PipelineModel = "pyannote/Segmentation-PyanNet-DIHARD",
batch_size: int = 32,
segmentation: PipelineModel = "pyannote/segmentation",
precision: Optional[float] = None,
recall: Optional[float] = None,
**inference_kwargs,
):
super().__init__()
self.segmentation = segmentation
self.batch_size = batch_size
# load model and send it to GPU (when available and not already on GPU)
model = get_model(segmentation)
@@ -126,10 +125,7 @@ class OverlappedSpeechDetection(Pipeline):
(segmentation_device,) = get_devices(needs=1)
model.to(segmentation_device)
self.segmentation_inference_ = Inference(
model,
batch_size=self.batch_size,
)
self.segmentation_inference_ = Inference(model, **inference_kwargs)
#  hyper-parameters used for hysteresis thresholding
self.onset = Uniform(0.0, 1.0)
@@ -209,7 +205,7 @@ class OverlappedSpeechDetection(Pipeline):
reference: Annotation,
hypothesis: Annotation,
uem: Timeline = None,
**kwargs
**kwargs,
) -> dict:
return super().compute_components(
to_overlap(reference), hypothesis, uem=uem, **kwargs
+9 -11
View File
@@ -50,10 +50,13 @@ class Resegmentation(Pipeline):
Parameters
----------
segmentation : Model, str, or dict, optional
Pretrained segmentation model. Defaults to "pyannote/Segmentation-PyanNet-DIHARD".
Pretrained segmentation model. Defaults to "pyannote/segmentation".
See pyannote.audio.pipelines.utils.get_model for supported format.
diarization : str, optional
File key to use as input diarization. Defaults to "diarization".
inference_kwargs : dict, optional
Keywords arguments passed to Inference.
Hyper-parameters
----------------
@@ -67,16 +70,15 @@ class Resegmentation(Pipeline):
def __init__(
self,
segmentation: PipelineModel = "pyannote/Segmentation-PyanNet-DIHARD",
batch_size: int = 32,
segmentation: PipelineModel = "pyannote/segmentation",
diarization: Text = "diarization",
**inference_kwargs,
):
super().__init__()
self.segmentation = segmentation
self.diarization = diarization
self.batch_size = batch_size
# load model and send it to GPU (when available and not already on GPU)
model = get_model(segmentation)
@@ -95,13 +97,9 @@ class Resegmentation(Pipeline):
self.seg_frames_ = model.introspection.frames
# prepare segmentation model for inference
self.seg_inference_ = Inference(
model,
window="sliding",
skip_aggregation=True,
duration=model.specifications.duration,
batch_size=32,
)
inference_kwargs["window"] = "sliding"
inference_kwargs["skip_aggregation"] = True
self.seg_inference_ = Inference(model, **inference_kwargs)
#  hyper-parameters used for hysteresis thresholding
self.onset = Uniform(0.0, 1.0)
+6 -10
View File
@@ -58,10 +58,10 @@ class Segmentation(Pipeline):
Parameters
----------
segmentation : Model, str, or dict, optional
Pretrained segmentation model. Defaults to "pyannote/Segmentation-PyanNet-DIHARD".
Pretrained segmentation model. Defaults to "pyannote/segmentation".
See pyannote.audio.pipelines.utils.get_model for supported format.
batch_size : int, optional
Batch size. Defaults to 32.
inference_kwargs : dict, optional
Keywords arguments passed to Inference.
Hyper-parameters
----------------
@@ -75,13 +75,12 @@ class Segmentation(Pipeline):
def __init__(
self,
segmentation: PipelineModel = "pyannote/Segmentation-PyanNet-DIHARD",
batch_size: int = 32,
segmentation: PipelineModel = "pyannote/segmentation",
**inference_kwargs,
):
super().__init__()
self.segmentation = segmentation
self.batch_size = batch_size
# load model and send it to GPU (when available and not already on GPU)
model = get_model(segmentation)
@@ -89,10 +88,7 @@ class Segmentation(Pipeline):
(segmentation_device,) = get_devices(needs=1)
model.to(segmentation_device)
self.segmentation_inference_ = Inference(
model,
batch_size=self.batch_size,
)
self.segmentation_inference_ = Inference(model, **inference_kwargs)
#  hyper-parameters used for hysteresis thresholding
self.onset = Uniform(0.0, 1.0)
@@ -85,13 +85,13 @@ class VoiceActivityDetection(Pipeline):
----------
segmentation : Model, str, or dict, optional
Pretrained segmentation (or voice activity detection) model.
Defaults to "hbredin/VoiceActivityDetection-PyanNet-DIHARD".
Defaults to "pyannote/segmentation".
See pyannote.audio.pipelines.utils.get_model for supported format.
batch_size : int, optional
Batch size. Defaults to 32.
fscore : bool, optional
Optimize (precision/recall) fscore. Defaults to optimizing detection
error rate.
inference_kwargs : dict, optional
Keywords arguments passed to Inference.
Hyper-parameters
----------------
@@ -105,14 +105,13 @@ class VoiceActivityDetection(Pipeline):
def __init__(
self,
segmentation: PipelineModel = "hbredin/VoiceActivityDetection-PyanNet-DIHARD",
batch_size: int = 32,
segmentation: PipelineModel = "pyannote/segmentation",
fscore: bool = False,
**inference_kwargs,
):
super().__init__()
self.segmentation = segmentation
self.batch_size = batch_size
self.fscore = fscore
# load model and send it to GPU (when available and not already on GPU)
@@ -121,10 +120,7 @@ class VoiceActivityDetection(Pipeline):
(segmentation_device,) = get_devices(needs=1)
model.to(segmentation_device)
self.segmentation_inference_ = Inference(
model,
batch_size=self.batch_size,
)
self.segmentation_inference_ = Inference(model, **inference_kwargs)
#  hyper-parameters used for hysteresis thresholding
self.onset = Uniform(0.0, 1.0)