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

auto frontend

Browse Source
This commit is contained in:
游雁 2024-06-07 16:18:18 +08:00
parent 8d7f76af46
commit 162efb747f
2 changed files with 23 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
funasr/models/llm_asr/adaptor.py
View File

@ -83,25 +83,27 @@ class Transformer(nn.Module):
from funasr.models.transformer.attention import MultiHeadedAttention from funasr.models.transformer.attention import MultiHeadedAttention
from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward
self.blocks = nn.ModuleList( self.blocks = None
[ if kwargs.get("n_layer", 2) > 0:
EncoderLayer( self.blocks = nn.ModuleList(
llm_dim, [
MultiHeadedAttention( EncoderLayer(
kwargs.get("attention_heads", 8),
llm_dim, llm_dim,
kwargs.get("attention_dropout_rate", 0.0), MultiHeadedAttention(
), kwargs.get("attention_heads", 8),
PositionwiseFeedForward( llm_dim,
llm_dim, kwargs.get("attention_dropout_rate", 0.0),
llm_dim // 4, ),
PositionwiseFeedForward(
llm_dim,
llm_dim // 4,
kwargs.get("dropout_rate", 0.0),
),
kwargs.get("dropout_rate", 0.0), kwargs.get("dropout_rate", 0.0),
), )
kwargs.get("dropout_rate", 0.0), for i in range(kwargs.get("n_layer", 2))
) ]
for i in range(kwargs.get("n_layer", 2)) )
]
)
def forward(self, x, ilens=None): def forward(self, x, ilens=None):
@ -123,6 +125,7 @@ class Transformer(nn.Module):
olens = None olens = None
olens = (ilens - 1) // self.k + 1 olens = (ilens - 1) // self.k + 1
masks = (~make_pad_mask(olens)[:, None, :]).to(x.device) masks = (~make_pad_mask(olens)[:, None, :]).to(x.device)
for layer, block in enumerate(self.blocks): if self.blocks is not None:
x, masks = block(x, masks) for layer, block in enumerate(self.blocks):
x, masks = block(x, masks)
return x, olens return x, olens

2
funasr/models/llm_asr/model.py
View File

@ -481,7 +481,7 @@ class LLMASR2(nn.Module):
batch_size, token_num, dims = inputs_embeds.shape batch_size, token_num, dims = inputs_embeds.shape
fbank_mask[fbank_mask < 0] = 0 fbank_mask[fbank_mask < 0] = 0
fbank_fake_lens = fbank_mask.sum(-1) fbank_fake_lens = fbank_mask.sum(-1).to(torch.int32)
# _, l, _ = encoder_out.shape # _, l, _ = encoder_out.shape
for batch_idx in range(batch_size): for batch_idx in range(batch_size):