From 03d4ce829814b4a7f57235fda049351c524ba32b Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 17 三月 2023 14:06:56 +0800
Subject: [PATCH] Merge branch 'main' into dev_xw
---
funasr/models/decoder/sanm_decoder.py | 101 ++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 100 insertions(+), 1 deletions(-)
diff --git a/funasr/models/decoder/sanm_decoder.py b/funasr/models/decoder/sanm_decoder.py
index ab03f0b..3bfcffc 100644
--- a/funasr/models/decoder/sanm_decoder.py
+++ b/funasr/models/decoder/sanm_decoder.py
@@ -94,6 +94,47 @@
if self.self_attn:
if self.normalize_before:
tgt = self.norm2(tgt)
+ x, _ = self.self_attn(tgt, tgt_mask)
+ x = residual + self.dropout(x)
+
+ if self.src_attn is not None:
+ residual = x
+ if self.normalize_before:
+ x = self.norm3(x)
+
+ x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
+
+
+ return x, tgt_mask, memory, memory_mask, cache
+
+ def forward_chunk(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
+ """Compute decoded features.
+
+ Args:
+ tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
+ tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
+ memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
+ memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
+ cache (List[torch.Tensor]): List of cached tensors.
+ Each tensor shape should be (#batch, maxlen_out - 1, size).
+
+ Returns:
+ torch.Tensor: Output tensor(#batch, maxlen_out, size).
+ torch.Tensor: Mask for output tensor (#batch, maxlen_out).
+ torch.Tensor: Encoded memory (#batch, maxlen_in, size).
+ torch.Tensor: Encoded memory mask (#batch, maxlen_in).
+
+ """
+ # tgt = self.dropout(tgt)
+ residual = tgt
+ if self.normalize_before:
+ tgt = self.norm1(tgt)
+ tgt = self.feed_forward(tgt)
+
+ x = tgt
+ if self.self_attn:
+ if self.normalize_before:
+ tgt = self.norm2(tgt)
if self.training:
cache = None
x, cache = self.self_attn(tgt, tgt_mask, cache=cache)
@@ -108,7 +149,6 @@
return x, tgt_mask, memory, memory_mask, cache
-
class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
"""
@@ -947,6 +987,65 @@
)
return logp.squeeze(0), state
+ def forward_chunk(
+ self,
+ memory: torch.Tensor,
+ tgt: torch.Tensor,
+ cache: dict = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ """Forward decoder.
+
+ Args:
+ hs_pad: encoded memory, float32 (batch, maxlen_in, feat)
+ hlens: (batch)
+ ys_in_pad:
+ input token ids, int64 (batch, maxlen_out)
+ if input_layer == "embed"
+ input tensor (batch, maxlen_out, #mels) in the other cases
+ ys_in_lens: (batch)
+ Returns:
+ (tuple): tuple containing:
+
+ x: decoded token score before softmax (batch, maxlen_out, token)
+ if use_output_layer is True,
+ olens: (batch, )
+ """
+ x = tgt
+ if cache["decode_fsmn"] is None:
+ cache_layer_num = len(self.decoders)
+ if self.decoders2 is not None:
+ cache_layer_num += len(self.decoders2)
+ new_cache = [None] * cache_layer_num
+ else:
+ new_cache = cache["decode_fsmn"]
+ for i in range(self.att_layer_num):
+ decoder = self.decoders[i]
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
+ x, None, memory, None, cache=new_cache[i]
+ )
+ new_cache[i] = c_ret
+
+ if self.num_blocks - self.att_layer_num > 1:
+ for i in range(self.num_blocks - self.att_layer_num):
+ j = i + self.att_layer_num
+ decoder = self.decoders2[i]
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
+ x, None, memory, None, cache=new_cache[j]
+ )
+ new_cache[j] = c_ret
+
+ for decoder in self.decoders3:
+
+ x, tgt_mask, memory, memory_mask, _ = decoder.forward_chunk(
+ x, None, memory, None, cache=None
+ )
+ if self.normalize_before:
+ x = self.after_norm(x)
+ if self.output_layer is not None:
+ x = self.output_layer(x)
+ cache["decode_fsmn"] = new_cache
+ return x
+
def forward_one_step(
self,
tgt: torch.Tensor,
--
Gitblit v1.9.1