speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Optimize memory (#612)

Browse Source 
Co-authored-by: 凌匀 <ailsa.zly@alibaba-inc.com>
This commit is contained in:
lingyunfly 2023-06-09 16:19:35 +08:00 committed by GitHub
parent 72cdfe65e3
commit 6bdae5723e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 14 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
funasr/models/e2e_vad.py
View File

@ -252,8 +252,8 @@ class E2EVadModel(nn.Module):
self.data_buf = None
self.data_buf_all = None
self.waveform = None
self.ResetDetection()
self.frontend = frontend
self.last_drop_frames = 0
def AllResetDetection(self):
self.data_buf_start_frame = 0
@ -282,7 +282,8 @@ class E2EVadModel(nn.Module):
self.data_buf = None
self.data_buf_all = None
self.waveform = None
self.ResetDetection()
self.last_drop_frames = 0
self.windows_detector.Reset()
def ResetDetection(self):
self.continous_silence_frame_count = 0
@ -295,6 +296,14 @@ class E2EVadModel(nn.Module):
self.sil_frame = 0
self.frame_probs = []
assert self.output_data_buf[-1].contain_seg_end_point == True
drop_frames = int(self.output_data_buf[-1].end_ms / self.vad_opts.frame_in_ms)
real_drop_frames = drop_frames - self.last_drop_frames
self.last_drop_frames = drop_frames
self.data_buf_all = self.data_buf_all[real_drop_frames * int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
self.decibel = self.decibel[real_drop_frames:]
self.scores = self.scores[:, real_drop_frames:, :]
def ComputeDecibel(self) -> None:
frame_sample_length = int(self.vad_opts.frame_length_ms * self.vad_opts.sample_rate / 1000)
frame_shift_length = int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000)
@ -322,7 +331,7 @@ class E2EVadModel(nn.Module):
while self.data_buf_start_frame < frame_idx:
if len(self.data_buf) >= int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):
self.data_buf_start_frame += 1
self.data_buf = self.data_buf_all[self.data_buf_start_frame * int(
self.data_buf = self.data_buf_all[(self.data_buf_start_frame - self.last_drop_frames) * int(
self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
def PopDataToOutputBuf(self, start_frm: int, frm_cnt: int, first_frm_is_start_point: bool,
@ -543,7 +552,7 @@ class E2EVadModel(nn.Module):
return 0
for i in range(self.vad_opts.nn_eval_block_size - 1, -1, -1):
frame_state = FrameState.kFrameStateInvalid
frame_state = self.GetFrameState(self.frm_cnt - 1 - i)
frame_state = self.GetFrameState(self.frm_cnt - 1 - i - self.last_drop_frames)
self.DetectOneFrame(frame_state, self.frm_cnt - 1 - i, False)
return 0
@ -553,7 +562,7 @@ class E2EVadModel(nn.Module):
return 0
for i in range(self.vad_opts.nn_eval_block_size - 1, -1, -1):
frame_state = FrameState.kFrameStateInvalid
frame_state = self.GetFrameState(self.frm_cnt - 1 - i)
frame_state = self.GetFrameState(self.frm_cnt - 1 - i - self.last_drop_frames)
if i != 0:
self.DetectOneFrame(frame_state, self.frm_cnt - 1 - i, False)
else: