From bbda5496ffae1d9ab052e8736a8c0b080ea017f5 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 21 三月 2024 16:28:00 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR merge
---
funasr/models/contextual_paraformer/model.py | 62 ++++++++++++++++---------------
1 files changed, 32 insertions(+), 30 deletions(-)
diff --git a/funasr/models/contextual_paraformer/model.py b/funasr/models/contextual_paraformer/model.py
index 49868a8..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
@@ -63,7 +60,6 @@
crit_attn_smooth = kwargs.get("crit_attn_smooth", 0.0)
bias_encoder_dropout_rate = kwargs.get("bias_encoder_dropout_rate", 0.0)
-
if bias_encoder_type == 'lstm':
self.bias_encoder = torch.nn.LSTM(inner_dim, inner_dim, 1, batch_first=True, dropout=bias_encoder_dropout_rate)
self.bias_embed = torch.nn.Embedding(self.vocab_size, inner_dim)
@@ -82,7 +78,6 @@
self.attn_loss = torch.nn.L1Loss()
self.crit_attn_smooth = crit_attn_smooth
-
def forward(
self,
speech: torch.Tensor,
@@ -99,21 +94,18 @@
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)
-
loss_ctc, cer_ctc = None, None
stats = dict()
@@ -128,12 +120,11 @@
stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
stats["cer_ctc"] = cer_ctc
-
# 2b. Attention decoder branch
loss_att, acc_att, cer_att, wer_att, loss_pre, loss_ideal = self._calc_att_clas_loss(
encoder_out, encoder_out_lens, text, text_lengths, hotword_pad, hotword_lengths
)
-
+
# 3. CTC-Att loss definition
if self.ctc_weight == 0.0:
loss = loss_att + loss_pre * self.predictor_weight
@@ -159,7 +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,
@@ -171,22 +161,24 @@
):
encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
encoder_out.device)
+
if self.predictor_bias == 1:
_, ys_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
ys_pad_lens = ys_pad_lens + self.predictor_bias
+
pre_acoustic_embeds, pre_token_length, _, _ = self.predictor(encoder_out, ys_pad, encoder_out_mask,
ignore_id=self.ignore_id)
-
# -1. bias encoder
if self.use_decoder_embedding:
hw_embed = self.decoder.embed(hotword_pad)
else:
hw_embed = self.bias_embed(hotword_pad)
+
hw_embed, (_, _) = self.bias_encoder(hw_embed)
_ind = np.arange(0, hotword_pad.shape[0]).tolist()
selected = hw_embed[_ind, [i - 1 for i in hotword_lengths.detach().cpu().tolist()]]
contextual_info = selected.squeeze(0).repeat(ys_pad.shape[0], 1, 1).to(ys_pad.device)
-
+
# 0. sampler
decoder_out_1st = None
if self.sampling_ratio > 0.0:
@@ -195,7 +187,7 @@
pre_acoustic_embeds, contextual_info)
else:
sematic_embeds = pre_acoustic_embeds
-
+
# 1. Forward decoder
decoder_outs = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=contextual_info
@@ -211,7 +203,7 @@
loss_ideal = None
'''
loss_ideal = None
-
+
if decoder_out_1st is None:
decoder_out_1st = decoder_out
# 2. Compute attention loss
@@ -231,7 +223,6 @@
cer_att, wer_att = self.error_calculator(ys_hat.cpu(), ys_pad.cpu())
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)
@@ -265,7 +256,6 @@
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):
if hw_list is None:
@@ -288,10 +278,11 @@
enforce_sorted=False)
_, (h_n, _) = self.bias_encoder(hw_embed)
hw_embed = h_n.repeat(encoder_out.shape[0], 1, 1)
-
+
decoder_outs = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=hw_embed, clas_scale=clas_scale
)
+
decoder_out = decoder_outs[0]
decoder_out = torch.log_softmax(decoder_out, dim=-1)
return decoder_out, ys_pad_lens
@@ -305,38 +296,42 @@
**kwargs,
):
# init beamsearch
+
is_use_ctc = kwargs.get("decoding_ctc_weight", 0.0) > 0.00001 and self.ctc != None
is_use_lm = kwargs.get("lm_weight", 0.0) > 0.00001 and kwargs.get("lm_file", None) is not None
if self.beam_search is None and (is_use_lm or is_use_ctc):
logging.info("enable beam_search")
self.init_beam_search(**kwargs)
self.nbest = kwargs.get("nbest", 1)
-
+
meta_data = {}
# extract fbank feats
time1 = time.perf_counter()
+
audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000))
+
time2 = time.perf_counter()
meta_data["load_data"] = f"{time2 - time1:0.3f}"
+
speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
frontend=frontend)
time3 = time.perf_counter()
meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
meta_data[
"batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
-
+
speech = speech.to(device=kwargs["device"])
speech_lengths = speech_lengths.to(device=kwargs["device"])
# hotword
self.hotword_list = self.generate_hotwords_list(kwargs.get("hotword", None), tokenizer=tokenizer, frontend=frontend)
-
+
# Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
if isinstance(encoder_out, tuple):
encoder_out = encoder_out[0]
-
+
# predictor
predictor_outs = self.calc_predictor(encoder_out, encoder_out_lens)
pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = predictor_outs[0], predictor_outs[1], \
@@ -344,8 +339,7 @@
pre_token_length = pre_token_length.round().long()
if torch.max(pre_token_length) < 1:
return []
-
-
+
decoder_outs = self.cal_decoder_with_predictor(encoder_out, encoder_out_lens,
pre_acoustic_embeds,
pre_token_length,
@@ -410,7 +404,6 @@
results.append(result_i)
return results, meta_data
-
def generate_hotwords_list(self, hotword_list_or_file, tokenizer=None, frontend=None):
def load_seg_dict(seg_dict_file):
@@ -513,3 +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