From 9d7c6e75f0b2ab1a5fc6e3a3a795d804fdb0bc8d Mon Sep 17 00:00:00 2001
From: ShiLiang Zhang <sly.zsl@alibaba-inc.com>
Date: 星期一, 07 八月 2023 10:18:01 +0800
Subject: [PATCH] Update SDK_tutorial_online.md
---
funasr/build_utils/build_diar_model.py | 34 ++++++++++++++++++++++++++--------
1 files changed, 26 insertions(+), 8 deletions(-)
diff --git a/funasr/build_utils/build_diar_model.py b/funasr/build_utils/build_diar_model.py
index d21df85..1aa0701 100644
--- a/funasr/build_utils/build_diar_model.py
+++ b/funasr/build_utils/build_diar_model.py
@@ -3,7 +3,7 @@
import torch
from funasr.layers.global_mvn import GlobalMVN
-from funasr.layers.label_aggregation import LabelAggregate
+from funasr.layers.label_aggregation import LabelAggregate, LabelAggregateMaxPooling
from funasr.layers.utterance_mvn import UtteranceMVN
from funasr.models.e2e_diar_eend_ola import DiarEENDOLAModel
from funasr.models.e2e_diar_sond import DiarSondModel
@@ -26,6 +26,8 @@
from funasr.models.frontend.windowing import SlidingWindow
from funasr.models.specaug.specaug import SpecAug
from funasr.models.specaug.specaug import SpecAugLFR
+from funasr.models.specaug.abs_profileaug import AbsProfileAug
+from funasr.models.specaug.profileaug import ProfileAug
from funasr.modules.eend_ola.encoder import EENDOLATransformerEncoder
from funasr.modules.eend_ola.encoder_decoder_attractor import EncoderDecoderAttractor
from funasr.torch_utils.initialize import initialize
@@ -52,6 +54,15 @@
default=None,
optional=True,
)
+profileaug_choices = ClassChoices(
+ name="profileaug",
+ classes=dict(
+ profileaug=ProfileAug,
+ ),
+ type_check=AbsProfileAug,
+ default=None,
+ optional=True,
+)
normalize_choices = ClassChoices(
"normalize",
classes=dict(
@@ -64,7 +75,8 @@
label_aggregator_choices = ClassChoices(
"label_aggregator",
classes=dict(
- label_aggregator=LabelAggregate
+ label_aggregator=LabelAggregate,
+ label_aggregator_max_pool=LabelAggregateMaxPooling,
),
default=None,
optional=True,
@@ -155,6 +167,8 @@
frontend_choices,
# --specaug and --specaug_conf
specaug_choices,
+ # --profileaug and --profileaug_conf
+ profileaug_choices,
# --normalize and --normalize_conf
normalize_choices,
# --label_aggregator and --label_aggregator_conf
@@ -209,13 +223,20 @@
encoder_class = encoder_choices.get_class(args.encoder)
encoder = encoder_class(input_size=input_size, **args.encoder_conf)
- if args.model_name == "sond":
+ if args.model == "sond":
# data augmentation for spectrogram
if args.specaug is not None:
specaug_class = specaug_choices.get_class(args.specaug)
specaug = specaug_class(**args.specaug_conf)
else:
specaug = None
+
+ # Data augmentation for Profiles
+ if hasattr(args, "profileaug") and args.profileaug is not None:
+ profileaug_class = profileaug_choices.get_class(args.profileaug)
+ profileaug = profileaug_class(**args.profileaug_conf)
+ else:
+ profileaug = None
# normalization layer
if args.normalize is not None:
@@ -247,11 +268,7 @@
# decoder
decoder_class = decoder_choices.get_class(args.decoder)
- decoder = decoder_class(
- vocab_size=vocab_size,
- encoder_output_size=encoder.output_size(),
- **args.decoder_conf,
- )
+ decoder = decoder_class(**args.decoder_conf)
# logger aggregator
if getattr(args, "label_aggregator", None) is not None:
@@ -265,6 +282,7 @@
vocab_size=vocab_size,
frontend=frontend,
specaug=specaug,
+ profileaug=profileaug,
normalize=normalize,
label_aggregator=label_aggregator,
encoder=encoder,
--
Gitblit v1.9.1