提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 funasr/models/contextual_paraformer/decoder.py

@@ -15,7 +15,10 @@
from funasr.models.transformer.embedding import PositionalEncoding
from funasr.models.paraformer.decoder import DecoderLayerSANM, ParaformerSANMDecoder
from funasr.models.sanm.positionwise_feed_forward import PositionwiseFeedForwardDecoderSANM
from funasr.models.sanm.attention import MultiHeadedAttentionSANMDecoder, MultiHeadedAttentionCrossAtt
from funasr.models.sanm.attention import (
    MultiHeadedAttentionSANMDecoder,
    MultiHeadedAttentionCrossAtt,
)


class ContextualDecoderLayer(torch.nn.Module):
@@ -47,7 +50,14 @@
            self.concat_linear1 = torch.nn.Linear(size + size, size)
            self.concat_linear2 = torch.nn.Linear(size + size, size)

    def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None,):
    def forward(
        self,
        tgt,
        tgt_mask,
        memory,
        memory_mask,
        cache=None,
    ):
        # tgt = self.dropout(tgt)
        if isinstance(tgt, Tuple):
            tgt, _ = tgt
@@ -97,8 +107,9 @@
        if self.src_attn is not None:
            if self.normalize_before:
                x = self.norm3(x)
            x =  self.dropout(self.src_attn(x, memory, memory_mask))
            x = self.dropout(self.src_attn(x, memory, memory_mask))
        return x, tgt_mask, memory, memory_mask, cache


@tables.register("decoder_classes", "ContextualParaformerDecoder")
class ContextualParaformerDecoder(ParaformerSANMDecoder):
@@ -107,6 +118,7 @@
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
    https://arxiv.org/abs/2006.01713
    """

    def __init__(
        self,
        vocab_size: int,
@@ -139,7 +151,7 @@
        )

        attention_dim = encoder_output_size
        if input_layer == 'none':
        if input_layer == "none":
            self.embed = None
        if input_layer == "embed":
            self.embed = torch.nn.Sequential(
@@ -194,20 +206,20 @@
            dropout_rate=dropout_rate,
            normalize_before=True,
        )
        self.bias_output = torch.nn.Conv1d(attention_dim*2, attention_dim, 1, bias=False)
        self.bias_output = torch.nn.Conv1d(attention_dim * 2, attention_dim, 1, bias=False)
        self.last_decoder = ContextualDecoderLayer(
                attention_dim,
                MultiHeadedAttentionSANMDecoder(
                    attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
                ),
                MultiHeadedAttentionCrossAtt(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            )
            attention_dim,
            MultiHeadedAttentionSANMDecoder(
                attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
            ),
            MultiHeadedAttentionCrossAtt(
                attention_heads, attention_dim, src_attention_dropout_rate
            ),
            PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
            dropout_rate,
            normalize_before,
            concat_after,
        )
        if num_blocks - att_layer_num <= 0:
            self.decoders2 = None
        else:
@@ -273,31 +285,25 @@
        memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :]

        x = tgt
        x, tgt_mask, memory, memory_mask, _ = self.decoders(
            x, tgt_mask, memory, memory_mask
        )
        _, _, x_self_attn, x_src_attn = self.last_decoder(
            x, tgt_mask, memory, memory_mask
        )
        x, tgt_mask, memory, memory_mask, _ = self.decoders(x, tgt_mask, memory, memory_mask)
        _, _, x_self_attn, x_src_attn = self.last_decoder(x, tgt_mask, memory, memory_mask)

        # contextual paraformer related
        contextual_length = torch.Tensor([contextual_info.shape[1]]).int().repeat(hs_pad.shape[0])
        contextual_mask = myutils.sequence_mask(contextual_length, device=memory.device)[:, None, :]
        cx, tgt_mask, _, _, _ = self.bias_decoder(x_self_attn, tgt_mask, contextual_info, memory_mask=contextual_mask)
        cx, tgt_mask, _, _, _ = self.bias_decoder(
            x_self_attn, tgt_mask, contextual_info, memory_mask=contextual_mask
        )

        if self.bias_output is not None:
            x = torch.cat([x_src_attn, cx*clas_scale], dim=2)
            x = torch.cat([x_src_attn, cx * clas_scale], dim=2)
            x = self.bias_output(x.transpose(1, 2)).transpose(1, 2)  # 2D -> D
            x = x_self_attn + self.dropout(x)

        if self.decoders2 is not None:
            x, tgt_mask, memory, memory_mask, _ = self.decoders2(
                x, tgt_mask, memory, memory_mask
            )
            x, tgt_mask, memory, memory_mask, _ = self.decoders2(x, tgt_mask, memory, memory_mask)

        x, tgt_mask, memory, memory_mask, _ = self.decoders3(
            x, tgt_mask, memory, memory_mask
        )
        x, tgt_mask, memory, memory_mask, _ = self.decoders3(x, tgt_mask, memory, memory_mask)
        if self.normalize_before:
            x = self.after_norm(x)
        olens = tgt_mask.sum(1)
@@ -308,20 +314,26 @@

@tables.register("decoder_classes", "ContextualParaformerDecoderExport")
class ContextualParaformerDecoderExport(torch.nn.Module):
    def __init__(self, model,
                 max_seq_len=512,
                 model_name='decoder',
                 onnx: bool = True,
                 **kwargs,):
    def __init__(
        self,
        model,
        max_seq_len=512,
        model_name="decoder",
        onnx: bool = True,
        **kwargs,
    ):
        super().__init__()
        from funasr.utils.torch_function import sequence_mask

        self.model = model
        self.make_pad_mask = sequence_mask(max_seq_len, flip=False)
        

        from funasr.models.sanm.attention import MultiHeadedAttentionSANMDecoderExport
        from funasr.models.sanm.attention import MultiHeadedAttentionCrossAttExport
        from funasr.models.paraformer.decoder import DecoderLayerSANMExport
        from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForwardDecoderSANMExport
        from funasr.models.transformer.positionwise_feed_forward import (
            PositionwiseFeedForwardDecoderSANMExport,
        )

        for i, d in enumerate(self.model.decoders):
            if isinstance(d.feed_forward, PositionwiseFeedForwardDecoderSANM):
@@ -344,27 +356,34 @@
            if isinstance(d.feed_forward, PositionwiseFeedForwardDecoderSANM):
                d.feed_forward = PositionwiseFeedForwardDecoderSANMExport(d.feed_forward)
            self.model.decoders3[i] = DecoderLayerSANMExport(d)
        

        self.output_layer = model.output_layer
        self.after_norm = model.after_norm
        self.model_name = model_name

        # bias decoder
        if isinstance(self.model.bias_decoder.src_attn, MultiHeadedAttentionCrossAtt):
            self.model.bias_decoder.src_attn = MultiHeadedAttentionCrossAttExport(self.model.bias_decoder.src_attn)
            self.model.bias_decoder.src_attn = MultiHeadedAttentionCrossAttExport(
                self.model.bias_decoder.src_attn
            )
        self.bias_decoder = self.model.bias_decoder
        

        # last decoder
        if isinstance(self.model.last_decoder.src_attn, MultiHeadedAttentionCrossAtt):
            self.model.last_decoder.src_attn = MultiHeadedAttentionCrossAttExport(self.model.last_decoder.src_attn)
            self.model.last_decoder.src_attn = MultiHeadedAttentionCrossAttExport(
                self.model.last_decoder.src_attn
            )
        if isinstance(self.model.last_decoder.self_attn, MultiHeadedAttentionSANMDecoder):
            self.model.last_decoder.self_attn = MultiHeadedAttentionSANMDecoderExport(self.model.last_decoder.self_attn)
            self.model.last_decoder.self_attn = MultiHeadedAttentionSANMDecoderExport(
                self.model.last_decoder.self_attn
            )
        if isinstance(self.model.last_decoder.feed_forward, PositionwiseFeedForwardDecoderSANM):
            self.model.last_decoder.feed_forward = PositionwiseFeedForwardDecoderSANMExport(self.model.last_decoder.feed_forward)
            self.model.last_decoder.feed_forward = PositionwiseFeedForwardDecoderSANMExport(
                self.model.last_decoder.feed_forward
            )
        self.last_decoder = self.model.last_decoder
        self.bias_output = self.model.bias_output
        self.dropout = self.model.dropout
        

    def prepare_mask(self, mask):
        mask_3d_btd = mask[:, :, None]
@@ -373,7 +392,7 @@
        elif len(mask.shape) == 3:
            mask_4d_bhlt = 1 - mask[:, None, :]
        mask_4d_bhlt = mask_4d_bhlt * -10000.0
    

        return mask_3d_btd, mask_4d_bhlt

    def forward(
@@ -396,22 +415,19 @@
        # memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :]

        x = tgt
        x, tgt_mask, memory, memory_mask, _ = self.model.decoders(
            x, tgt_mask, memory, memory_mask
        )
        x, tgt_mask, memory, memory_mask, _ = self.model.decoders(x, tgt_mask, memory, memory_mask)

        _, _, x_self_attn, x_src_attn = self.last_decoder(
            x, tgt_mask, memory, memory_mask
        )
        _, _, x_self_attn, x_src_attn = self.last_decoder(x, tgt_mask, memory, memory_mask)

        # contextual paraformer related
        contextual_length = torch.Tensor([bias_embed.shape[1]]).int().repeat(hs_pad.shape[0])
        # contextual_mask = myutils.sequence_mask(contextual_length, device=memory.device)[:, None, :]
        contextual_mask = self.make_pad_mask(contextual_length)
        contextual_mask, _ = self.prepare_mask(contextual_mask)
        # import pdb; pdb.set_trace()
        contextual_mask = contextual_mask.transpose(2, 1).unsqueeze(1)
        cx, tgt_mask, _, _, _ = self.bias_decoder(x_self_attn, tgt_mask, bias_embed, memory_mask=contextual_mask)
        cx, tgt_mask, _, _, _ = self.bias_decoder(
            x_self_attn, tgt_mask, bias_embed, memory_mask=contextual_mask
        )

        if self.bias_output is not None:
            x = torch.cat([x_src_attn, cx], dim=2)
@@ -422,11 +438,8 @@
            x, tgt_mask, memory, memory_mask, _ = self.model.decoders2(
                x, tgt_mask, memory, memory_mask
            )
        x, tgt_mask, memory, memory_mask, _ = self.model.decoders3(
            x, tgt_mask, memory, memory_mask
        )
        x, tgt_mask, memory, memory_mask, _ = self.model.decoders3(x, tgt_mask, memory, memory_mask)
        x = self.after_norm(x)
        x = self.output_layer(x)

        return x, ys_in_lens