From 9ba0dbd98bf69c830dfcfde8f109a400cb65e4e5 Mon Sep 17 00:00:00 2001
From: 雾聪 <wucong.lyb@alibaba-inc.com>
Date: 星期五, 29 三月 2024 17:24:59 +0800
Subject: [PATCH] fix func Forward
---
funasr/models/contextual_paraformer/model.py | 65 ++++++++++++--------------------
1 files changed, 24 insertions(+), 41 deletions(-)
diff --git a/funasr/models/contextual_paraformer/model.py b/funasr/models/contextual_paraformer/model.py
index 655ca6f..18cab60 100644
--- a/funasr/models/contextual_paraformer/model.py
+++ b/funasr/models/contextual_paraformer/model.py
@@ -102,17 +102,16 @@
text_lengths = text_lengths[:, 0]
if len(speech_lengths.size()) > 1:
speech_lengths = speech_lengths[:, 0]
- pdb.set_trace()
+
batch_size = speech.shape[0]
hotword_pad = kwargs.get("hotword_pad")
hotword_lengths = kwargs.get("hotword_lengths")
- dha_pad = kwargs.get("dha_pad")
- pdb.set_trace()
+ # dha_pad = kwargs.get("dha_pad")
+
# 1. Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
- pdb.set_trace()
loss_ctc, cer_ctc = None, None
stats = dict()
@@ -127,12 +126,11 @@
stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
stats["cer_ctc"] = cer_ctc
- pdb.set_trace()
# 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
)
- pdb.set_trace()
+
# 3. CTC-Att loss definition
if self.ctc_weight == 0.0:
loss = loss_att + loss_pre * self.predictor_weight
@@ -170,38 +168,33 @@
):
encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
encoder_out.device)
- pdb.set_trace()
+
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
- pdb.set_trace()
+
pre_acoustic_embeds, pre_token_length, _, _ = self.predictor(encoder_out, ys_pad, encoder_out_mask,
ignore_id=self.ignore_id)
- pdb.set_trace()
# -1. bias encoder
if self.use_decoder_embedding:
hw_embed = self.decoder.embed(hotword_pad)
else:
hw_embed = self.bias_embed(hotword_pad)
- pdb.set_trace()
+
hw_embed, (_, _) = self.bias_encoder(hw_embed)
- pdb.set_trace()
_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)
- pdb.set_trace()
+
# 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
- pdb.set_trace()
+
# 1. Forward decoder
decoder_outs = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=contextual_info
@@ -217,7 +210,7 @@
loss_ideal = None
'''
loss_ideal = None
- pdb.set_trace()
+
if decoder_out_1st is None:
decoder_out_1st = decoder_out
# 2. Compute attention loss
@@ -294,10 +287,11 @@
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
@@ -311,75 +305,62 @@
**kwargs,
):
# init beamsearch
- pdb.set_trace()
+
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)
- pdb.set_trace()
+
meta_data = {}
# extract fbank feats
time1 = time.perf_counter()
- pdb.set_trace()
+
audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000))
- pdb.set_trace()
+
time2 = time.perf_counter()
meta_data["load_data"] = f"{time2 - time1:0.3f}"
- pdb.set_trace()
+
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
-
- pdb.set_trace()
+
speech = speech.to(device=kwargs["device"])
speech_lengths = speech_lengths.to(device=kwargs["device"])
# hotword
- pdb.set_trace()
self.hotword_list = self.generate_hotwords_list(kwargs.get("hotword", None), tokenizer=tokenizer, frontend=frontend)
- pdb.set_trace()
-
# Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
if isinstance(encoder_out, tuple):
encoder_out = encoder_out[0]
- pdb.set_trace()
-
# 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], \
predictor_outs[2], predictor_outs[3]
- pdb.set_trace()
pre_token_length = pre_token_length.round().long()
if torch.max(pre_token_length) < 1:
return []
-
- pdb.set_trace()
decoder_outs = self.cal_decoder_with_predictor(encoder_out, encoder_out_lens,
pre_acoustic_embeds,
pre_token_length,
hw_list=self.hotword_list,
clas_scale=kwargs.get("clas_scale", 1.0))
- pdb.set_trace()
decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]
- pdb.set_trace()
results = []
b, n, d = decoder_out.size()
- pdb.set_trace()
for i in range(b):
x = encoder_out[i, :encoder_out_lens[i], :]
am_scores = decoder_out[i, :pre_token_length[i], :]
- pdb.set_trace()
if self.beam_search is not None:
nbest_hyps = self.beam_search(
x=x, am_scores=am_scores, maxlenratio=kwargs.get("maxlenratio", 0.0),
@@ -399,9 +380,11 @@
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):
--
Gitblit v1.9.1