From 33d3d2084403fd34b79c835d2f2fe04f6cd8f738 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 13 九月 2023 09:33:54 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR add
---
funasr/models/e2e_tp.py | 44 ++++++++++++++++++++++++++++++--------------
1 files changed, 30 insertions(+), 14 deletions(-)
diff --git a/funasr/models/e2e_tp.py b/funasr/models/e2e_tp.py
index 8808008..567dc70 100644
--- a/funasr/models/e2e_tp.py
+++ b/funasr/models/e2e_tp.py
@@ -9,7 +9,6 @@
import torch
import numpy as np
-from typeguard import check_argument_types
from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.models.frontend.abs_frontend import AbsFrontend
@@ -17,9 +16,8 @@
from funasr.modules.add_sos_eos import add_sos_eos
from funasr.modules.nets_utils import make_pad_mask, pad_list
from funasr.torch_utils.device_funcs import force_gatherable
-from funasr.train.abs_espnet_model import AbsESPnetModel
+from funasr.models.base_model import FunASRModel
from funasr.models.predictor.cif import CifPredictorV3
-
if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
from torch.cuda.amp import autocast
@@ -30,9 +28,9 @@
yield
-class TimestampPredictor(AbsESPnetModel):
+class TimestampPredictor(FunASRModel):
"""
- Author: Speech Lab, Alibaba Group, China
+ Author: Speech Lab of DAMO Academy, Alibaba Group
"""
def __init__(
@@ -41,8 +39,8 @@
encoder: AbsEncoder,
predictor: CifPredictorV3,
predictor_bias: int = 0,
+ token_list=None,
):
- assert check_argument_types()
super().__init__()
# note that eos is the same as sos (equivalent ID)
@@ -54,7 +52,8 @@
self.predictor = predictor
self.predictor_bias = predictor_bias
self.criterion_pre = mae_loss()
-
+ self.token_list = token_list
+
def forward(
self,
speech: torch.Tensor,
@@ -63,7 +62,6 @@
text_lengths: torch.Tensor,
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
"""Frontend + Encoder + Decoder + Calc loss
-
Args:
speech: (Batch, Length, ...)
speech_lengths: (Batch, )
@@ -111,7 +109,6 @@
self, speech: torch.Tensor, speech_lengths: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Frontend + Encoder. Note that this method is used by asr_inference.py
-
Args:
speech: (Batch, Length, ...)
speech_lengths: (Batch, )
@@ -126,7 +123,7 @@
encoder_out, encoder_out_lens, _ = self.encoder(feats, feats_lengths)
return encoder_out, encoder_out_lens
-
+
def _extract_feats(
self, speech: torch.Tensor, speech_lengths: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
@@ -148,7 +145,26 @@
def calc_predictor_timestamp(self, encoder_out, encoder_out_lens, token_num):
encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
encoder_out.device)
- ds_alphas, ds_cif_peak, us_alphas, us_cif_peak = self.predictor.get_upsample_timestamp(encoder_out,
- encoder_out_mask,
- token_num)
- return ds_alphas, ds_cif_peak, us_alphas, us_cif_peak
+ ds_alphas, ds_cif_peak, us_alphas, us_peaks = self.predictor.get_upsample_timestamp(encoder_out,
+ encoder_out_mask,
+ token_num)
+ return ds_alphas, ds_cif_peak, us_alphas, us_peaks
+
+ def collect_feats(
+ self,
+ speech: torch.Tensor,
+ speech_lengths: torch.Tensor,
+ text: torch.Tensor,
+ text_lengths: torch.Tensor,
+ ) -> Dict[str, torch.Tensor]:
+ if self.extract_feats_in_collect_stats:
+ feats, feats_lengths = self._extract_feats(speech, speech_lengths)
+ else:
+ # Generate dummy stats if extract_feats_in_collect_stats is False
+ logging.warning(
+ "Generating dummy stats for feats and feats_lengths, "
+ "because encoder_conf.extract_feats_in_collect_stats is "
+ f"{self.extract_feats_in_collect_stats}"
+ )
+ feats, feats_lengths = speech, speech_lengths
+ return {"feats": feats, "feats_lengths": feats_lengths}
--
Gitblit v1.9.1