From e9d2cfc3a134b00f4e98271fbee3838d1ccecbcc Mon Sep 17 00:00:00 2001
From: VirtuosoQ <2416050435@qq.com>
Date: 星期五, 26 四月 2024 14:59:30 +0800
Subject: [PATCH] FunASR java http client
---
funasr/models/fsmn_vad_streaming/encoder.py | 90 ++++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 88 insertions(+), 2 deletions(-)
diff --git a/funasr/models/fsmn_vad_streaming/encoder.py b/funasr/models/fsmn_vad_streaming/encoder.py
index ae91852..bc51a6f 100755
--- a/funasr/models/fsmn_vad_streaming/encoder.py
+++ b/funasr/models/fsmn_vad_streaming/encoder.py
@@ -1,5 +1,6 @@
from typing import Tuple, Dict
import copy
+import os
import numpy as np
import torch
@@ -134,6 +135,25 @@
x3 = self.affine(x2)
x4 = self.relu(x3)
return x4
+class BasicBlock_export(nn.Module):
+ def __init__(self,
+ model,
+ ):
+ super(BasicBlock_export, self).__init__()
+ self.linear = model.linear
+ self.fsmn_block = model.fsmn_block
+ self.affine = model.affine
+ self.relu = model.relu
+
+ def forward(self, input: torch.Tensor, in_cache: torch.Tensor):
+ x = self.linear(input) # B T D
+ # cache_layer_name = 'cache_layer_{}'.format(self.stack_layer)
+ # if cache_layer_name not in in_cache:
+ # in_cache[cache_layer_name] = torch.zeros(x1.shape[0], x1.shape[-1], (self.lorder - 1) * self.lstride, 1)
+ x, out_cache = self.fsmn_block(x, in_cache)
+ x = self.affine(x)
+ x = self.relu(x)
+ return x, out_cache
class FsmnStack(nn.Sequential):
@@ -174,7 +194,7 @@
output_affine_dim: int,
output_dim: int
):
- super(FSMN, self).__init__()
+ super().__init__()
self.input_dim = input_dim
self.input_affine_dim = input_affine_dim
@@ -192,7 +212,7 @@
self.out_linear1 = AffineTransform(linear_dim, output_affine_dim)
self.out_linear2 = AffineTransform(output_affine_dim, output_dim)
self.softmax = nn.Softmax(dim=-1)
-
+
def fuse_modules(self):
pass
@@ -219,6 +239,72 @@
return x7
+@tables.register("encoder_classes", "FSMNExport")
+class FSMNExport(nn.Module):
+ def __init__(
+ self, model, **kwargs,
+ ):
+ super().__init__()
+
+ # self.input_dim = input_dim
+ # self.input_affine_dim = input_affine_dim
+ # self.fsmn_layers = fsmn_layers
+ # self.linear_dim = linear_dim
+ # self.proj_dim = proj_dim
+ # self.output_affine_dim = output_affine_dim
+ # self.output_dim = output_dim
+ #
+ # self.in_linear1 = AffineTransform(input_dim, input_affine_dim)
+ # self.in_linear2 = AffineTransform(input_affine_dim, linear_dim)
+ # self.relu = RectifiedLinear(linear_dim, linear_dim)
+ # self.fsmn = FsmnStack(*[BasicBlock(linear_dim, proj_dim, lorder, rorder, lstride, rstride, i) for i in
+ # range(fsmn_layers)])
+ # self.out_linear1 = AffineTransform(linear_dim, output_affine_dim)
+ # self.out_linear2 = AffineTransform(output_affine_dim, output_dim)
+ # self.softmax = nn.Softmax(dim=-1)
+ self.in_linear1 = model.in_linear1
+ self.in_linear2 = model.in_linear2
+ self.relu = model.relu
+ # self.fsmn = model.fsmn
+ self.out_linear1 = model.out_linear1
+ self.out_linear2 = model.out_linear2
+ self.softmax = model.softmax
+ self.fsmn = model.fsmn
+ for i, d in enumerate(model.fsmn):
+ if isinstance(d, BasicBlock):
+ self.fsmn[i] = BasicBlock_export(d)
+
+ def fuse_modules(self):
+ pass
+
+ def forward(
+ self,
+ input: torch.Tensor,
+ *args,
+ ):
+ """
+ Args:
+ input (torch.Tensor): Input tensor (B, T, D)
+ in_cache: when in_cache is not None, the forward is in streaming. The type of in_cache is a dict, egs,
+ {'cache_layer_1': torch.Tensor(B, T1, D)}, T1 is equal to self.lorder. It is {} for the 1st frame
+ """
+
+ x = self.in_linear1(input)
+ x = self.in_linear2(x)
+ x = self.relu(x)
+ # x4 = self.fsmn(x3, in_cache) # self.in_cache will update automatically in self.fsmn
+ out_caches = list()
+ for i, d in enumerate(self.fsmn):
+ in_cache = args[i]
+ x, out_cache = d(x, in_cache)
+ out_caches.append(out_cache)
+ x = self.out_linear1(x)
+ x = self.out_linear2(x)
+ x = self.softmax(x)
+
+ return x, out_caches
+
+
'''
one deep fsmn layer
dimproj: projection dimension, input and output dimension of memory blocks
--
Gitblit v1.9.1