From 6dcd960fda8be389af355ede4ecc583b036029d4 Mon Sep 17 00:00:00 2001
From: 雾聪 <wucong.lyb@alibaba-inc.com>
Date: 星期二, 05 三月 2024 16:45:25 +0800
Subject: [PATCH] update cutsplit step
---
funasr/models/seaco_paraformer/model.py | 85 ++++++++++++++++++------------------------
1 files changed, 36 insertions(+), 49 deletions(-)
diff --git a/funasr/models/seaco_paraformer/model.py b/funasr/models/seaco_paraformer/model.py
index 1867bbf..a8b1f1f 100644
--- a/funasr/models/seaco_paraformer/model.py
+++ b/funasr/models/seaco_paraformer/model.py
@@ -19,20 +19,18 @@
from funasr.register import tables
from funasr.utils import postprocess_utils
-from funasr.metrics.compute_acc import th_accuracy
from funasr.models.paraformer.model import Paraformer
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.paraformer.search import Hypothesis
-from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.train_utils.device_funcs import force_gatherable
from funasr.models.bicif_paraformer.model import BiCifParaformer
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
-from funasr.utils.timestamp_tools import ts_prediction_lfr6_standard
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
+from funasr.utils.timestamp_tools import ts_prediction_lfr6_standard
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
else:
@@ -66,7 +64,6 @@
# bias encoder
if self.bias_encoder_type == 'lstm':
- logging.warning("enable bias encoder sampling and contextual training")
self.bias_encoder = torch.nn.LSTM(self.inner_dim,
self.inner_dim,
2,
@@ -77,9 +74,8 @@
self.lstm_proj = torch.nn.Linear(self.inner_dim*2, self.inner_dim)
else:
self.lstm_proj = None
- self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
+ # self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
elif self.bias_encoder_type == 'mean':
- logging.warning("enable bias encoder sampling and contextual training")
self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
else:
logging.error("Unsupport bias encoder type: {}".format(self.bias_encoder_type))
@@ -132,9 +128,8 @@
hotword_pad = kwargs.get("hotword_pad")
hotword_lengths = kwargs.get("hotword_lengths")
dha_pad = kwargs.get("dha_pad")
-
+
batch_size = speech.shape[0]
- self.step_cur += 1
# for data-parallel
text = text[:, : text_lengths.max()]
speech = speech[:, :speech_lengths.max()]
@@ -214,25 +209,24 @@
nfilter=50,
seaco_weight=1.0):
# decoder forward
+
decoder_out, decoder_hidden, _ = self.decoder(encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, return_hidden=True, return_both=True)
+
decoder_pred = torch.log_softmax(decoder_out, dim=-1)
if hw_list is not None:
hw_lengths = [len(i) for i in hw_list]
hw_list_ = [torch.Tensor(i).long() for i in hw_list]
hw_list_pad = pad_list(hw_list_, 0).to(encoder_out.device)
selected = self._hotword_representation(hw_list_pad, torch.Tensor(hw_lengths).int().to(encoder_out.device))
+
contextual_info = selected.squeeze(0).repeat(encoder_out.shape[0], 1, 1).to(encoder_out.device)
num_hot_word = contextual_info.shape[1]
_contextual_length = torch.Tensor([num_hot_word]).int().repeat(encoder_out.shape[0]).to(encoder_out.device)
-
+
# ASF Core
if nfilter > 0 and nfilter < num_hot_word:
- for dec in self.seaco_decoder.decoders:
- dec.reserve_attn = True
- # cif_attended, _ = self.decoder2(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
- dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
- # cif_filter = torch.topk(self.decoder2.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1], min(nfilter, num_hot_word-1))[1].tolist()
- hotword_scores = self.seaco_decoder.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1]
+ hotword_scores = self.seaco_decoder.forward_asf6(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
+ hotword_scores = hotword_scores[0].sum(0).sum(0)
# hotword_scores /= torch.sqrt(torch.tensor(hw_lengths)[:-1].float()).to(hotword_scores.device)
dec_filter = torch.topk(hotword_scores, min(nfilter, num_hot_word-1))[1].tolist()
add_filter = dec_filter
@@ -243,15 +237,12 @@
contextual_info = selected.squeeze(0).repeat(encoder_out.shape[0], 1, 1).to(encoder_out.device)
num_hot_word = contextual_info.shape[1]
_contextual_length = torch.Tensor([num_hot_word]).int().repeat(encoder_out.shape[0]).to(encoder_out.device)
- for dec in self.seaco_decoder.decoders:
- dec.attn_mat = []
- dec.reserve_attn = False
# SeACo Core
cif_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
merged = self._merge(cif_attended, dec_attended)
-
+
dha_output = self.hotword_output_layer(merged) # remove the last token in loss calculation
dha_pred = torch.log_softmax(dha_output, dim=-1)
def _merge_res(dec_output, dha_output):
@@ -265,8 +256,8 @@
# logits = dec_output * dha_mask + dha_output[:,:,:-1] * (1-dha_mask)
logits = dec_output * dha_mask + dha_output[:,:,:] * (1-dha_mask)
return logits
+
merged_pred = _merge_res(decoder_pred, dha_pred)
- # import pdb; pdb.set_trace()
return merged_pred
else:
return decoder_pred
@@ -276,6 +267,8 @@
hotword_lengths):
if self.bias_encoder_type != 'lstm':
logging.error("Unsupported bias encoder type")
+
+ '''
hw_embed = self.decoder.embed(hotword_pad)
hw_embed, (_, _) = self.bias_encoder(hw_embed)
if self.lstm_proj is not None:
@@ -283,26 +276,20 @@
_ind = np.arange(0, hw_embed.shape[0]).tolist()
selected = hw_embed[_ind, [i-1 for i in hotword_lengths.detach().cpu().tolist()]]
return selected
+ '''
- '''
- def calc_predictor(self, encoder_out, encoder_out_lens):
- encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
- encoder_out.device)
- pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index, pre_token_length2 = self.predictor(encoder_out,
- None,
- encoder_out_mask,
- ignore_id=self.ignore_id)
- return pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index
-
-
- 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_peaks = self.predictor.get_upsample_timestamp(encoder_out,
- encoder_out_mask,
- token_num)
- return ds_alphas, ds_cif_peak, us_alphas, us_peaks
- '''
+ # hw_embed = self.sac_embedding(hotword_pad)
+ hw_embed = self.decoder.embed(hotword_pad)
+ hw_embed = torch.nn.utils.rnn.pack_padded_sequence(hw_embed, hotword_lengths.cpu().type(torch.int64), batch_first=True, enforce_sorted=False)
+ packed_rnn_output, _ = self.bias_encoder(hw_embed)
+ rnn_output = torch.nn.utils.rnn.pad_packed_sequence(packed_rnn_output, batch_first=True)[0]
+ if self.lstm_proj is not None:
+ hw_hidden = self.lstm_proj(rnn_output)
+ else:
+ hw_hidden = rnn_output
+ _ind = np.arange(0, hw_hidden.shape[0]).tolist()
+ selected = hw_hidden[_ind, [i-1 for i in hotword_lengths.detach().cpu().tolist()]]
+ return selected
def inference(self,
data_in,
@@ -320,7 +307,6 @@
logging.info("enable beam_search")
self.init_beam_search(**kwargs)
self.nbest = kwargs.get("nbest", 1)
-
meta_data = {}
# extract fbank feats
@@ -337,7 +323,7 @@
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)
@@ -346,6 +332,7 @@
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, _, _ = predictor_outs[0], predictor_outs[1], \
@@ -354,15 +341,14 @@
if torch.max(pre_token_length) < 1:
return []
-
decoder_out = self._seaco_decode_with_ASF(encoder_out, encoder_out_lens,
pre_acoustic_embeds,
pre_token_length,
hw_list=self.hotword_list)
+
# decoder_out, _ = decoder_outs[0], decoder_outs[1]
_, _, us_alphas, us_peaks = self.calc_predictor_timestamp(encoder_out, encoder_out_lens,
pre_token_length)
-
results = []
b, n, d = decoder_out.size()
for i in range(b):
@@ -387,9 +373,11 @@
nbest_hyps = [Hypothesis(yseq=yseq, score=score)]
for nbest_idx, hyp in enumerate(nbest_hyps):
ibest_writer = None
- if ibest_writer is None and kwargs.get("output_dir") is not None:
- writer = DatadirWriter(kwargs.get("output_dir"))
- ibest_writer = writer[f"{nbest_idx + 1}best_recog"]
+ if kwargs.get("output_dir") is not None:
+ if not hasattr(self, "writer"):
+ self.writer = DatadirWriter(kwargs.get("output_dir"))
+ ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]
+
# remove sos/eos and get results
last_pos = -1
if isinstance(hyp.yseq, list):
@@ -415,12 +403,11 @@
token, timestamp)
result_i = {"key": key[i], "text": text_postprocessed,
- "timestamp": time_stamp_postprocessed,
+ "timestamp": time_stamp_postprocessed
}
if ibest_writer is not None:
ibest_writer["token"][key[i]] = " ".join(token)
- # ibest_writer["text"][key[i]] = text
ibest_writer["timestamp"][key[i]] = time_stamp_postprocessed
ibest_writer["text"][key[i]] = text_postprocessed
else:
--
Gitblit v1.9.1