AI/wekws
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[fix] Fix fsmn export error (#143) (#146)

Browse Source 
* fix: Fix fsmn export error (#143)

* fix flake8

* fix flake8
This commit is contained in:
Menglong Xu 2023-09-10 09:01:15 +08:00 committed by GitHub
parent 694928d706
commit 2c3c9ce383
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 7 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
wekws/bin/export_onnx.py
View File

@ -41,6 +41,8 @@ def main():
configs = yaml.load(fin, Loader=yaml.FullLoader) configs = yaml.load(fin, Loader=yaml.FullLoader)
feature_dim = configs['model']['input_dim'] feature_dim = configs['model']['input_dim']
model = init_model(configs['model']) model = init_model(configs['model'])
is_fsmn = configs['model']['backbone']['type'] == 'fsmn'
num_layers = configs['model']['backbone']['num_layers']
if configs['training_config'].get('criterion', 'max_pooling') == 'ctc': if configs['training_config'].get('criterion', 'max_pooling') == 'ctc':
# if we use ctc_loss, the logits need to be convert into probs # if we use ctc_loss, the logits need to be convert into probs
model.forward = model.forward_softmax model.forward = model.forward_softmax
@ -54,6 +56,8 @@ def main():
model.hdim, model.hdim,
model.backbone.padding, model.backbone.padding,
dtype=torch.float) dtype=torch.float)
if is_fsmn:
cache = cache.unsqueeze(-1).expand(-1, -1, -1, num_layers)
torch.onnx.export(model, (dummy_input, cache), torch.onnx.export(model, (dummy_input, cache),
args.onnx_model, args.onnx_model,
input_names=['input', 'cache'], input_names=['input', 'cache'],

4
wekws/model/fsmn.py
View File

@ -477,6 +477,8 @@ class FSMN(nn.Module):
if in_cache is None or len(in_cache) == 0 : if in_cache is None or len(in_cache) == 0 :
in_cache = [torch.zeros(0, 0, 0, 0, dtype=torch.float) in_cache = [torch.zeros(0, 0, 0, 0, dtype=torch.float)
for _ in range(len(self.fsmn))] for _ in range(len(self.fsmn))]
else:
in_cache = [in_cache[:, :, :, i: i + 1] for i in range(in_cache.size(-1))]
input = (input, in_cache) input = (input, in_cache)
x1 = self.in_linear1(input) x1 = self.in_linear1(input)
x2 = self.in_linear2(x1) x2 = self.in_linear2(x1)
@ -490,7 +492,7 @@ class FSMN(nn.Module):
# x7 = self.softmax(x6) # x7 = self.softmax(x6)
x7, _ = x6 x7, _ = x6
# return x7, None # return x7, None
return x7, in_cache return x7, torch.cat(in_cache, dim=-1)
def to_kaldi_net(self): def to_kaldi_net(self):
re_str = '' re_str = ''