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

Dev gzf exp (#1664)

Browse Source 
* rwkv 5

* rwkv v4

* rwkv v4

* rwkv

* rwkv

* update

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step

* resume from step
This commit is contained in:
zhifu gao 2024-04-26 01:37:29 +08:00 committed by GitHub
parent e971e000ad
commit 8fdc372c81
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 23 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
funasr/datasets/sense_voice_datasets/datasets.py
View File

@ -99,8 +99,9 @@ class SenseVoiceDataset(torch.utils.data.Dataset):
target_mask = ( target_mask = (
[0] * (prompt_ids_len) + [1] * (target_ids_len) + [1] [0] * (prompt_ids_len) + [1] * (target_ids_len) + [1]
) # [sos, task, lid, text, eos]: [0, 0, 1, 1, 1] ) # [sos, task, lid, text, eos]: [0, 0, 1, 1, 1]
target_mask_lengths = len(target_mask)
target_mask = torch.tensor(target_mask, dtype=torch.float32) target_mask = torch.tensor(target_mask, dtype=torch.float32)
target_mask_lengths = torch.tensor([target_mask_lengths], dtype=torch.int32)
return { return {
"speech": speech[0, :, :], "speech": speech[0, :, :],
"speech_lengths": speech_lengths, "speech_lengths": speech_lengths,
@ -130,30 +131,26 @@ class SenseVoiceDataset(torch.utils.data.Dataset):
) )
if self.batch_type != "example": if self.batch_type != "example":
b, t, _ = outputs["speech"].shape for i in range(3):
if b * t > self.batch_size: outputs = self._filter_badcase(outputs)
beg = torch.randint(0, 2, ()).item()
logging.info( return outputs
f"Warning, b * t: {b * t} > {self.batch_size}, drop half data 1st, beg:{beg}"
) def _filter_badcase(self, outputs, i=0):
for key, data_list in outputs.items(): b, t, _ = outputs["speech"].shape
outputs[key] = outputs[key][beg : beg + b : 2] if b * t > self.batch_size:
beg = torch.randint(0, 2, ()).item()
logging.info(
f"Warning, b * t: {b * t} > {self.batch_size}, drop half data {i}th, beg:{beg}"
)
for key, data_list in outputs.items():
outputs[key] = outputs[key][beg : beg + b : 2]
speech_lengths_max = outputs["speech_lengths_max"].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"].max().item()
outputs["target_mask"] = outputs["target_mask"][:, :target_mask_lengths_max]
b, t, _ = outputs["speech"].shape
if b * t > self.batch_size:
beg = torch.randint(0, 2, ()).item()
logging.info(
f"Warning, b * t: {b * t} > {self.batch_size}, drop half data 2nd, beg:{beg}"
)
for key, data_list in outputs.items():
outputs[key] = outputs[key][beg : beg + b : 2]
b, t, _ = outputs["speech"].shape
if b * t > self.batch_size:
beg = torch.randint(0, 2, ()).item()
logging.info(
f"Warning, b * t: {b * t} > {self.batch_size}, drop half data 3th, beg:{beg}"
)
for key, data_list in outputs.items():
outputs[key] = outputs[key][beg : beg + b : 2]
return outputs return outputs