From 6ebd2676480068c1cb27ffc1b9318b09e1662173 Mon Sep 17 00:00:00 2001
From: 九耳 <mengzhe.cmz@alibaba-inc.com>
Date: 星期三, 29 三月 2023 16:48:57 +0800
Subject: [PATCH] general punc model conversion onnx
---
funasr/export/models/target_delay_transformer.py | 160 +++++++++++++++++++++++++++++++++++++++++++++++++++++
funasr/export/export_model.py | 12 +++
2 files changed, 171 insertions(+), 1 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index 4aed49f..9afa7b1 100644
--- a/funasr/export/export_model.py
+++ b/funasr/export/export_model.py
@@ -174,7 +174,10 @@
json_file = os.path.join(model_dir, 'configuration.json')
with open(json_file, 'r') as f:
config_data = json.load(f)
- mode = config_data['model']['model_config']['mode']
+ if config_data['task'] == "punctuation":
+ mode = config_data['model']['punc_model_config']['mode']
+ else:
+ mode = config_data['model']['model_config']['mode']
if mode.startswith('paraformer'):
from funasr.tasks.asr import ASRTaskParaformer as ASRTask
config = os.path.join(model_dir, 'config.yaml')
@@ -195,6 +198,13 @@
)
self.export_config["feats_dim"] = 400
self.frontend = model.frontend
+ elif mode.startswith('punc'):
+ from funasr.tasks.punctuation import PunctuationTask as PUNCTask
+ punc_train_config = os.path.join(model_dir, 'config.yaml')
+ punc_model_file = os.path.join(model_dir, 'punc.pb')
+ model, punc_train_args = PUNCTask.build_model_from_file(
+ punc_train_config, punc_model_file, 'cpu'
+ )
self._export(model, tag_name)
diff --git a/funasr/export/models/target_delay_transformer.py b/funasr/export/models/target_delay_transformer.py
new file mode 100644
index 0000000..0a2586c
--- /dev/null
+++ b/funasr/export/models/target_delay_transformer.py
@@ -0,0 +1,160 @@
+from typing import Any
+from typing import List
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+from funasr.export.utils.torch_function import MakePadMask
+from funasr.export.utils.torch_function import sequence_mask
+#from funasr.models.encoder.sanm_encoder import SANMEncoder as Encoder
+from funasr.punctuation.sanm_encoder import SANMEncoder
+from funasr.export.models.encoder.sanm_encoder import SANMEncoder as SANMEncoder_export
+from funasr.punctuation.abs_model import AbsPunctuation
+
+
+class TargetDelayTransformer(nn.Module):
+
+ def __init__(
+ self,
+ model,
+ max_seq_len=512,
+ model_name='punc_model',
+ **kwargs,
+ ):
+ super().__init__()
+ onnx = False
+ if "onnx" in kwargs:
+ onnx = kwargs["onnx"]
+ self.embed = model.embed
+ self.decoder = model.decoder
+ self.model = model
+ self.feats_dim = self.embed.embedding_dim
+ self.num_embeddings = self.embed.num_embeddings
+ self.model_name = model_name
+ from typing import Any
+ from typing import List
+ from typing import Tuple
+
+ import torch
+ import torch.nn as nn
+
+ from funasr.export.utils.torch_function import MakePadMask
+ from funasr.export.utils.torch_function import sequence_mask
+ # from funasr.models.encoder.sanm_encoder import SANMEncoder as Encoder
+ from funasr.punctuation.sanm_encoder import SANMEncoder
+ from funasr.export.models.encoder.sanm_encoder import SANMEncoder as SANMEncoder_export
+ from funasr.punctuation.abs_model import AbsPunctuation
+
+ class TargetDelayTransformer(nn.Module):
+
+ def __init__(
+ self,
+ model,
+ max_seq_len=512,
+ model_name='punc_model',
+ **kwargs,
+ ):
+ super().__init__()
+ onnx = False
+ if "onnx" in kwargs:
+ onnx = kwargs["onnx"]
+ self.embed = model.embed
+ self.decoder = model.decoder
+ self.model = model
+ self.feats_dim = self.embed.embedding_dim
+ self.num_embeddings = self.embed.num_embeddings
+ self.model_name = model_name
+
+ if isinstance(model.encoder, SANMEncoder):
+ self.encoder = SANMEncoder_export(model.encoder, onnx=onnx)
+ else:
+ assert False, "Only support samn encode."
+
+ def forward(self, input: torch.Tensor, text_lengths: torch.Tensor) -> Tuple[torch.Tensor, None]:
+ """Compute loss value from buffer sequences.
+
+ Args:
+ input (torch.Tensor): Input ids. (batch, len)
+ hidden (torch.Tensor): Target ids. (batch, len)
+
+ """
+ x = self.embed(input)
+ # mask = self._target_mask(input)
+ h, _ = self.encoder(x, text_lengths)
+ y = self.decoder(h)
+ return y
+
+ def get_dummy_inputs(self):
+ length = 120
+ text_indexes = torch.randint(0, self.embed.num_embeddings, (2, length))
+ text_lengths = torch.tensor([length - 20, length], dtype=torch.int32)
+ return (text_indexes, text_lengths)
+
+ def get_input_names(self):
+ return ['input', 'text_lengths']
+
+ def get_output_names(self):
+ return ['logits']
+
+ def get_dynamic_axes(self):
+ return {
+ 'input': {
+ 0: 'batch_size',
+ 1: 'feats_length'
+ },
+ 'text_lengths': {
+ 0: 'batch_size',
+ },
+ 'logits': {
+ 0: 'batch_size',
+ 1: 'logits_length'
+ },
+ }
+
+ if isinstance(model.encoder, SANMEncoder):
+ self.encoder = SANMEncoder_export(model.encoder, onnx=onnx)
+ else:
+ assert False, "Only support samn encode."
+
+ def forward(self, input: torch.Tensor, text_lengths: torch.Tensor) -> Tuple[torch.Tensor, None]:
+ """Compute loss value from buffer sequences.
+
+ Args:
+ input (torch.Tensor): Input ids. (batch, len)
+ hidden (torch.Tensor): Target ids. (batch, len)
+
+ """
+ x = self.embed(input)
+ # mask = self._target_mask(input)
+ h, _ = self.encoder(x, text_lengths)
+ y = self.decoder(h)
+ return y
+
+ def get_dummy_inputs(self):
+ length = 120
+ text_indexes = torch.randint(0, self.embed.num_embeddings, (2, length))
+ text_lengths = torch.tensor([length-20, length], dtype=torch.int32)
+ return (text_indexes, text_lengths)
+
+ def get_input_names(self):
+ return ['input', 'text_lengths']
+
+ def get_output_names(self):
+ return ['logits']
+
+ def get_dynamic_axes(self):
+ return {
+ 'input': {
+ 0: 'batch_size',
+ 1: 'feats_length'
+ },
+ 'text_lengths': {
+ 0: 'batch_size',
+ },
+ 'logits': {
+ 0: 'batch_size',
+ 1: 'logits_length'
+ },
+ }
+
--
Gitblit v1.9.1