From 9b4e9cc8a0311e5243d69b73ed073e7ea441982e Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 27 三月 2024 16:05:29 +0800
Subject: [PATCH] train update
---
funasr/models/contextual_paraformer/model.py | 32 ++++++++++++++------------------
1 files changed, 14 insertions(+), 18 deletions(-)
diff --git a/funasr/models/contextual_paraformer/model.py b/funasr/models/contextual_paraformer/model.py
index 1c0805a..b9fd3c4 100644
--- a/funasr/models/contextual_paraformer/model.py
+++ b/funasr/models/contextual_paraformer/model.py
@@ -17,9 +17,6 @@
from distutils.version import LooseVersion
from funasr.register import tables
-from funasr.losses.label_smoothing_loss import (
- LabelSmoothingLoss, # noqa: H301
-)
from funasr.utils import postprocess_utils
from funasr.metrics.compute_acc import th_accuracy
from funasr.models.paraformer.model import Paraformer
@@ -29,7 +26,7 @@
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
-import pdb
+
if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
from torch.cuda.amp import autocast
@@ -81,7 +78,6 @@
self.attn_loss = torch.nn.L1Loss()
self.crit_attn_smooth = crit_attn_smooth
-
def forward(
self,
speech: torch.Tensor,
@@ -98,16 +94,14 @@
text: (Batch, Length)
text_lengths: (Batch,)
"""
- if len(text_lengths.size()) > 1:
- text_lengths = text_lengths[:, 0]
- if len(speech_lengths.size()) > 1:
- speech_lengths = speech_lengths[:, 0]
+ text_lengths = text_lengths.squeeze()
+ speech_lengths = speech_lengths.squeeze()
batch_size = speech.shape[0]
hotword_pad = kwargs.get("hotword_pad")
hotword_lengths = kwargs.get("hotword_lengths")
- dha_pad = kwargs.get("dha_pad")
+ # dha_pad = kwargs.get("dha_pad")
# 1. Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
@@ -156,7 +150,6 @@
loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
return loss, stats, weight
-
def _calc_att_clas_loss(
self,
encoder_out: torch.Tensor,
@@ -189,13 +182,10 @@
# 0. sampler
decoder_out_1st = None
if self.sampling_ratio > 0.0:
- if self.step_cur < 2:
- logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
+
sematic_embeds, decoder_out_1st = self.sampler(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens,
pre_acoustic_embeds, contextual_info)
else:
- if self.step_cur < 2:
- logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
sematic_embeds = pre_acoustic_embeds
# 1. Forward decoder
@@ -234,7 +224,6 @@
return loss_att, acc_att, cer_att, wer_att, loss_pre, loss_ideal
-
def sampler(self, encoder_out, encoder_out_lens, ys_pad, ys_pad_lens, pre_acoustic_embeds, contextual_info):
tgt_mask = (~make_pad_mask(ys_pad_lens, maxlen=ys_pad_lens.max())[:, :, None]).to(ys_pad.device)
ys_pad = ys_pad * tgt_mask[:, :, 0]
@@ -266,7 +255,6 @@
sematic_embeds = pre_acoustic_embeds.masked_fill(~input_mask_expand_dim, 0) + ys_pad_embed.masked_fill(
input_mask_expand_dim, 0)
return sematic_embeds * tgt_mask, decoder_out * tgt_mask
-
def cal_decoder_with_predictor(self, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, hw_list=None,
clas_scale=1.0):
@@ -417,7 +405,6 @@
return results, meta_data
-
def generate_hotwords_list(self, hotword_list_or_file, tokenizer=None, frontend=None):
def load_seg_dict(seg_dict_file):
seg_dict = {}
@@ -519,3 +506,12 @@
hotword_list = None
return hotword_list
+ def export(
+ self,
+ **kwargs,
+ ):
+ if 'max_seq_len' not in kwargs:
+ kwargs['max_seq_len'] = 512
+ from .export_meta import export_rebuild_model
+ models = export_rebuild_model(model=self, **kwargs)
+ return models
--
Gitblit v1.9.1