From 4eedd7ad45120598f10b84673401823719dda237 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 07 六月 2024 03:05:01 +0800
Subject: [PATCH] auto frontend
---
funasr/datasets/openai_datasets/datasets.py | 41 ++++++++++++++++++++++++++++-------------
1 files changed, 28 insertions(+), 13 deletions(-)
diff --git a/funasr/datasets/openai_datasets/datasets.py b/funasr/datasets/openai_datasets/datasets.py
index 9bd0698..73e9ec0 100644
--- a/funasr/datasets/openai_datasets/datasets.py
+++ b/funasr/datasets/openai_datasets/datasets.py
@@ -51,7 +51,7 @@
self.batch_size = kwargs.get("batch_size")
self.batch_type = kwargs.get("batch_type")
self.prompt_ids_len = 0
- self.retry = kwargs.get("retry", 5)
+ self.retry = kwargs.get("retry", 10)
self.permute = False
from funasr.frontends.whisper_frontend import WhisperFrontend
@@ -60,6 +60,8 @@
self.permute = True
self.pattern = re.compile(r"(<\|startofspeech\|>.*?<\|endofspeech\|>)")
+ # self.kwargs = kwargs
+ self.max_token_length = kwargs.get("max_token_length", 1024)
def get_source_len(self, index):
item = self.index_ds[index]
@@ -77,7 +79,9 @@
# pdb.set_trace()
output = None
+
for idx in range(self.retry):
+ badcase_flag = False
if idx == 0:
index_cur = index
else:
@@ -112,9 +116,14 @@
"<|endofspeech|>", ""
)
if sub_str.startswith("!"):
-
- data_src = load_audio_text_image_video(sub_str[1:], fs=self.fs)
-
+ try:
+ data_src = load_audio_text_image_video(sub_str[1:], fs=self.fs)
+ except Exception as e:
+ logging.error(
+ f"Loading wav failed! {str(e)}, {traceback.format_exc()}"
+ )
+ badcase_flag = True
+ continue
speech, speech_lengths = extract_fbank(
data_src,
data_type=self.data_type,
@@ -134,6 +143,8 @@
source_ids += sub_token
fbank_mask_i += [1] * len(sub_token)
+ if badcase_flag:
+ continue
source_mask = [-100] * len(source_ids)
target_out = f"{target_out}<|im_end|>"
target_ids = self.tokenizer.encode(target_out)
@@ -142,6 +153,10 @@
fbank_mask += fbank_mask_i
fbank_beg.append(fbank_beg_i)
+ if len(input_ids) > self.max_token_length:
+ badcase_flag = True
+ if badcase_flag:
+ continue
input_ids = torch.tensor(input_ids, dtype=torch.int64)
attention_mask = torch.tensor([len(input_ids)], dtype=torch.int32)
labels = torch.tensor(labels, dtype=torch.int64)
@@ -193,9 +208,9 @@
return outputs
def _filter_badcase(self, outputs, i=0):
- b, t, _ = outputs["speech"].shape
+ b, t = outputs["input_ids"].shape
- if b * t > self.batch_size * 1.25:
+ if b * t > self.batch_size * 2:
beg = torch.randint(0, 2, ()).item()
if b < 2:
beg = 0
@@ -204,12 +219,12 @@
)
for key, data_list in outputs.items():
outputs[key] = outputs[key][beg : beg + b : 2]
-
- speech_lengths_max = outputs["speech_lengths"].max().item()
- outputs["speech"] = outputs["speech"][:, :speech_lengths_max, :]
- text_lengths_max = outputs["text_lengths"].max().item()
- outputs["text"] = outputs["text"][:, :text_lengths_max]
- target_mask_lengths_max = outputs["target_mask_lengths"].max().item()
- outputs["target_mask"] = outputs["target_mask"][:, :target_mask_lengths_max]
+ #
+ # speech_lengths_max = outputs["speech_lengths"].max().item()
+ # outputs["speech"] = outputs["speech"][:, :speech_lengths_max, :]
+ # text_lengths_max = outputs["text_lengths"].max().item()
+ # outputs["text"] = outputs["text"][:, :text_lengths_max]
+ # target_mask_lengths_max = outputs["target_mask_lengths"].max().item()
+ # outputs["target_mask"] = outputs["target_mask"][:, :target_mask_lengths_max]
return outputs
--
Gitblit v1.9.1