Merge pull request #160 from alibaba-damo-academy/dev_onnx

zhifu gao

2023-02-27 19467b57f6476cc0ba5493c0dcde3d15a0c88c2c

Merge pull request #160 from alibaba-damo-academy/dev_onnx

Dev onnx

 .gitignore

@@ -7,4 +7,6 @@
init_model/
*.tar.gz
test_local/
RapidASR
RapidASR
export/*
*.pyc

 funasr/export/models/__init__.py

@@ -1,10 +1,13 @@
from funasr.models.e2e_asr_paraformer import Paraformer
from funasr.models.e2e_asr_paraformer import Paraformer, BiCifParaformer
from funasr.export.models.e2e_asr_paraformer import Paraformer as Paraformer_export
from funasr.export.models.e2e_asr_paraformer import BiCifParaformer as BiCifParaformer_export
from funasr.models.e2e_uni_asr import UniASR

def get_model(model, export_config=None):

    if isinstance(model, Paraformer):
def get_model(model, export_config=None):
    if isinstance(model, BiCifParaformer):
        return BiCifParaformer_export(model, **export_config)
    elif isinstance(model, Paraformer):
        return Paraformer_export(model, **export_config)
    else:
        raise "The model is not exist!"
        raise "Funasr does not support the given model type currently."

 funasr/export/models/decoder/transformer_decoder.py

New file
@@ -0,0 +1,143 @@
import os
from funasr.export import models

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.modules.attention import MultiHeadedAttentionSANMDecoder
from funasr.export.models.modules.multihead_att import MultiHeadedAttentionSANMDecoder as MultiHeadedAttentionSANMDecoder_export
from funasr.modules.attention import MultiHeadedAttentionCrossAtt, MultiHeadedAttention
from funasr.export.models.modules.multihead_att import MultiHeadedAttentionCrossAtt as MultiHeadedAttentionCrossAtt_export
from funasr.export.models.modules.multihead_att import OnnxMultiHeadedAttention
from funasr.modules.positionwise_feed_forward import PositionwiseFeedForwardDecoderSANM
from funasr.export.models.modules.feedforward import PositionwiseFeedForwardDecoderSANM as PositionwiseFeedForwardDecoderSANM_export
from funasr.export.models.modules.decoder_layer import DecoderLayer as DecoderLayer_export


class ParaformerDecoderSAN(nn.Module):
    def __init__(self, model,
                 max_seq_len=512,
                 model_name='decoder',
                 onnx: bool = True,):
        super().__init__()
        # self.embed = model.embed #Embedding(model.embed, max_seq_len)
        self.model = model
        if onnx:
            self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        else:
            self.make_pad_mask = sequence_mask(max_seq_len, flip=False)

        for i, d in enumerate(self.model.decoders):
            if isinstance(d.feed_forward, PositionwiseFeedForwardDecoderSANM):
                d.feed_forward = PositionwiseFeedForwardDecoderSANM_export(d.feed_forward)
            if isinstance(d.self_attn, MultiHeadedAttentionSANMDecoder):
                d.self_attn = MultiHeadedAttentionSANMDecoder_export(d.self_attn)
            # if isinstance(d.src_attn, MultiHeadedAttentionCrossAtt):
            #     d.src_attn = MultiHeadedAttentionCrossAtt_export(d.src_attn)
            if isinstance(d.src_attn, MultiHeadedAttention):
                d.src_attn = OnnxMultiHeadedAttention(d.src_attn)
            self.model.decoders[i] = DecoderLayer_export(d)
        
        self.output_layer = model.output_layer
        self.after_norm = model.after_norm
        self.model_name = model_name
        

    def prepare_mask(self, mask):
        mask_3d_btd = mask[:, :, None]
        if len(mask.shape) == 2:
            mask_4d_bhlt = 1 - mask[:, None, None, :]
        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(
        self,
        hs_pad: torch.Tensor,
        hlens: torch.Tensor,
        ys_in_pad: torch.Tensor,
        ys_in_lens: torch.Tensor,
    ):

        tgt = ys_in_pad
        tgt_mask = self.make_pad_mask(ys_in_lens)
        tgt_mask, _ = self.prepare_mask(tgt_mask)
        # tgt_mask = myutils.sequence_mask(ys_in_lens, device=tgt.device)[:, :, None]

        memory = hs_pad
        memory_mask = self.make_pad_mask(hlens)
        _, memory_mask = self.prepare_mask(memory_mask)
        # 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 = self.after_norm(x)
        x = self.output_layer(x)

        return x, ys_in_lens


    def get_dummy_inputs(self, enc_size):
        tgt = torch.LongTensor([0]).unsqueeze(0)
        memory = torch.randn(1, 100, enc_size)
        pre_acoustic_embeds = torch.randn(1, 1, enc_size)
        cache_num = len(self.model.decoders) + len(self.model.decoders2)
        cache = [
            torch.zeros((1, self.model.decoders[0].size, self.model.decoders[0].self_attn.kernel_size))
            for _ in range(cache_num)
        ]
        return (tgt, memory, pre_acoustic_embeds, cache)

    def is_optimizable(self):
        return True

    def get_input_names(self):
        cache_num = len(self.model.decoders) + len(self.model.decoders2)
        return ['tgt', 'memory', 'pre_acoustic_embeds'] \
               + ['cache_%d' % i for i in range(cache_num)]

    def get_output_names(self):
        cache_num = len(self.model.decoders) + len(self.model.decoders2)
        return ['y'] \
               + ['out_cache_%d' % i for i in range(cache_num)]

    def get_dynamic_axes(self):
        ret = {
            'tgt': {
                0: 'tgt_batch',
                1: 'tgt_length'
            },
            'memory': {
                0: 'memory_batch',
                1: 'memory_length'
            },
            'pre_acoustic_embeds': {
                0: 'acoustic_embeds_batch',
                1: 'acoustic_embeds_length',
            }
        }
        cache_num = len(self.model.decoders) + len(self.model.decoders2)
        ret.update({
            'cache_%d' % d: {
                0: 'cache_%d_batch' % d,
                2: 'cache_%d_length' % d
            }
            for d in range(cache_num)
        })
        return ret

    def get_model_config(self, path):
        return {
            "dec_type": "XformerDecoder",
            "model_path": os.path.join(path, f'{self.model_name}.onnx'),
            "n_layers": len(self.model.decoders) + len(self.model.decoders2),
            "odim": self.model.decoders[0].size
        }

 funasr/export/models/e2e_asr_paraformer.py

@@ -1,17 +1,21 @@
import logging


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
from funasr.models.encoder.conformer_encoder import ConformerEncoder
from funasr.export.models.encoder.sanm_encoder import SANMEncoder as SANMEncoder_export
from funasr.models.predictor.cif import CifPredictorV2
from funasr.export.models.encoder.conformer_encoder import ConformerEncoder as ConformerEncoder_export
from funasr.models.predictor.cif import CifPredictorV2, CifPredictorV3
from funasr.export.models.predictor.cif import CifPredictorV2 as CifPredictorV2_export
from funasr.export.models.predictor.cif import CifPredictorV3 as CifPredictorV3_export
from funasr.models.decoder.sanm_decoder import ParaformerSANMDecoder
from funasr.models.decoder.transformer_decoder import ParaformerDecoderSAN
from funasr.export.models.decoder.sanm_decoder import ParaformerSANMDecoder as ParaformerSANMDecoder_export
from funasr.export.models.decoder.transformer_decoder import ParaformerDecoderSAN as ParaformerDecoderSAN_export


class Paraformer(nn.Module):
    """
@@ -34,10 +38,14 @@
            onnx = kwargs["onnx"]
        if isinstance(model.encoder, SANMEncoder):
            self.encoder = SANMEncoder_export(model.encoder, onnx=onnx)
        elif isinstance(model.encoder, ConformerEncoder):
            self.encoder = ConformerEncoder_export(model.encoder, onnx=onnx)
        if isinstance(model.predictor, CifPredictorV2):
            self.predictor = CifPredictorV2_export(model.predictor)
        if isinstance(model.decoder, ParaformerSANMDecoder):
            self.decoder = ParaformerSANMDecoder_export(model.decoder, onnx=onnx)
        elif isinstance(model.decoder, ParaformerDecoderSAN):
            self.decoder = ParaformerDecoderSAN_export(model.decoder, onnx=onnx)
        
        self.feats_dim = feats_dim
        self.model_name = model_name
@@ -99,4 +107,113 @@
                0: 'batch_size',
                1: 'logits_length'
            },
        }


class BiCifParaformer(nn.Module):
    """
    Author: Speech Lab, Alibaba Group, China
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
    https://arxiv.org/abs/2206.08317
    """

    def __init__(
            self,
            model,
            max_seq_len=512,
            feats_dim=560,
            model_name='model',
            **kwargs,
    ):
        super().__init__()
        onnx = False
        if "onnx" in kwargs:
            onnx = kwargs["onnx"]
        if isinstance(model.encoder, SANMEncoder):
            self.encoder = SANMEncoder_export(model.encoder, onnx=onnx)
        elif isinstance(model.encoder, ConformerEncoder):
            self.encoder = ConformerEncoder_export(model.encoder, onnx=onnx)
        else:
            logging.warning("Unsupported encoder type to export.")
        if isinstance(model.predictor, CifPredictorV3):
            self.predictor = CifPredictorV3_export(model.predictor)
        else:
            logging.warning("Wrong predictor type to export.")
        if isinstance(model.decoder, ParaformerSANMDecoder):
            self.decoder = ParaformerSANMDecoder_export(model.decoder, onnx=onnx)
        elif isinstance(model.decoder, ParaformerDecoderSAN):
            self.decoder = ParaformerDecoderSAN_export(model.decoder, onnx=onnx)
        else:
            logging.warning("Unsupported decoder type to export.")
        
        self.feats_dim = feats_dim
        self.model_name = model_name

        if onnx:
            self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        else:
            self.make_pad_mask = sequence_mask(max_seq_len, flip=False)
        
    def forward(
            self,
            speech: torch.Tensor,
            speech_lengths: torch.Tensor,
    ):
        # a. To device
        batch = {"speech": speech, "speech_lengths": speech_lengths}
        # batch = to_device(batch, device=self.device)
    
        enc, enc_len = self.encoder(**batch)
        mask = self.make_pad_mask(enc_len)[:, None, :]
        pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = self.predictor(enc, mask)
        pre_token_length = pre_token_length.round().type(torch.int32)

        decoder_out, _ = self.decoder(enc, enc_len, pre_acoustic_embeds, pre_token_length)
        decoder_out = torch.log_softmax(decoder_out, dim=-1)
        
        # get predicted timestamps
        us_alphas, us_cif_peak = self.predictor.get_upsample_timestmap(enc, mask, pre_token_length)

        return decoder_out, pre_token_length, us_alphas, us_cif_peak

    def get_dummy_inputs(self):
        speech = torch.randn(2, 30, self.feats_dim)
        speech_lengths = torch.tensor([6, 30], dtype=torch.int32)
        return (speech, speech_lengths)

    def get_dummy_inputs_txt(self, txt_file: str = "/mnt/workspace/data_fbank/0207/12345.wav.fea.txt"):
        import numpy as np
        fbank = np.loadtxt(txt_file)
        fbank_lengths = np.array([fbank.shape[0], ], dtype=np.int32)
        speech = torch.from_numpy(fbank[None, :, :].astype(np.float32))
        speech_lengths = torch.from_numpy(fbank_lengths.astype(np.int32))
        return (speech, speech_lengths)

    def get_input_names(self):
        return ['speech', 'speech_lengths']

    def get_output_names(self):
        return ['logits', 'token_num', 'us_alphas', 'us_cif_peak']

    def get_dynamic_axes(self):
        return {
            'speech': {
                0: 'batch_size',
                1: 'feats_length'
            },
            'speech_lengths': {
                0: 'batch_size',
            },
            'logits': {
                0: 'batch_size',
                1: 'logits_length'
            },
            'us_alphas': {
                0: 'batch_size',
                1: 'alphas_length'
            },
            'us_cif_peak': {
                0: 'batch_size',
                1: 'alphas_length'
            },
        }

 funasr/export/models/encoder/conformer_encoder.py

New file
@@ -0,0 +1,106 @@
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.modules.attention import MultiHeadedAttentionSANM
from funasr.export.models.modules.multihead_att import MultiHeadedAttentionSANM as MultiHeadedAttentionSANM_export
from funasr.export.models.modules.encoder_layer import EncoderLayerSANM as EncoderLayerSANM_export
from funasr.export.models.modules.encoder_layer import EncoderLayerConformer as EncoderLayerConformer_export
from funasr.modules.positionwise_feed_forward import PositionwiseFeedForward
from funasr.export.models.modules.feedforward import PositionwiseFeedForward as PositionwiseFeedForward_export
from funasr.export.models.encoder.sanm_encoder import SANMEncoder
from funasr.modules.attention import RelPositionMultiHeadedAttention
# from funasr.export.models.modules.multihead_att import RelPositionMultiHeadedAttention as RelPositionMultiHeadedAttention_export
from funasr.export.models.modules.multihead_att import OnnxRelPosMultiHeadedAttention as RelPositionMultiHeadedAttention_export


class ConformerEncoder(nn.Module):
    def __init__(
        self,
        model,
        max_seq_len=512,
        feats_dim=560,
        model_name='encoder',
        onnx: bool = True,
    ):
        super().__init__()
        self.embed = model.embed
        self.model = model
        self.feats_dim = feats_dim
        self._output_size = model._output_size

        if onnx:
            self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        else:
            self.make_pad_mask = sequence_mask(max_seq_len, flip=False)

        for i, d in enumerate(self.model.encoders):
            if isinstance(d.self_attn, MultiHeadedAttentionSANM):
                d.self_attn = MultiHeadedAttentionSANM_export(d.self_attn)
            if isinstance(d.self_attn, RelPositionMultiHeadedAttention):
                d.self_attn = RelPositionMultiHeadedAttention_export(d.self_attn)
            if isinstance(d.feed_forward, PositionwiseFeedForward):
                d.feed_forward = PositionwiseFeedForward_export(d.feed_forward)
            self.model.encoders[i] = EncoderLayerConformer_export(d)
        
        self.model_name = model_name
        self.num_heads = model.encoders[0].self_attn.h
        self.hidden_size = model.encoders[0].self_attn.linear_out.out_features

    
    def prepare_mask(self, mask):
        if len(mask.shape) == 2:
            mask = 1 - mask[:, None, None, :]
        elif len(mask.shape) == 3:
            mask = 1 - mask[:, None, :]
        
        return mask * -10000.0

    def forward(self,
                speech: torch.Tensor,
                speech_lengths: torch.Tensor,
                ):
        speech = speech * self._output_size ** 0.5
        mask = self.make_pad_mask(speech_lengths)
        mask = self.prepare_mask(mask)
        if self.embed is None:
            xs_pad = speech
        else:
            xs_pad = self.embed(speech)

        encoder_outs = self.model.encoders(xs_pad, mask)
        xs_pad, masks = encoder_outs[0], encoder_outs[1]

        if isinstance(xs_pad, tuple):
            xs_pad = xs_pad[0]
        xs_pad = self.model.after_norm(xs_pad)

        return xs_pad, speech_lengths

    def get_output_size(self):
        return self.model.encoders[0].size

    def get_dummy_inputs(self):
        feats = torch.randn(1, 100, self.feats_dim)
        return (feats)

    def get_input_names(self):
        return ['feats']

    def get_output_names(self):
        return ['encoder_out', 'encoder_out_lens', 'predictor_weight']

    def get_dynamic_axes(self):
        return {
            'feats': {
                1: 'feats_length'
            },
            'encoder_out': {
                1: 'enc_out_length'
            },
            'predictor_weight':{
                1: 'pre_out_length'
            }

        }

 funasr/export/models/modules/decoder_layer.py

@@ -41,3 +41,30 @@

        return x, tgt_mask, memory, memory_mask, cache


class DecoderLayer(nn.Module):
    def __init__(self, model):
        super().__init__()
        self.self_attn = model.self_attn
        self.src_attn = model.src_attn
        self.feed_forward = model.feed_forward
        self.norm1 = model.norm1
        self.norm2 = model.norm2
        self.norm3 = model.norm3
    
    def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None):
        residual = tgt
        tgt_q = tgt
        tgt_q_mask = tgt_mask
        x = residual + self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)

        residual = x
        x = self.norm2(x)
        
        x = residual + self.src_attn(x, memory, memory, memory_mask)

        residual = x
        x = self.norm3(x)
        x = residual + self.feed_forward(x)

        return x, tgt_mask, memory, memory_mask

 funasr/export/models/modules/encoder_layer.py

@@ -34,4 +34,58 @@
        return x, mask


class EncoderLayerConformer(nn.Module):
    def __init__(
        self,
        model,
    ):
        """Construct an EncoderLayer object."""
        super().__init__()
        self.self_attn = model.self_attn
        self.feed_forward = model.feed_forward
        self.feed_forward_macaron = model.feed_forward_macaron
        self.conv_module = model.conv_module
        self.norm_ff = model.norm_ff
        self.norm_mha = model.norm_mha
        self.norm_ff_macaron = model.norm_ff_macaron
        self.norm_conv = model.norm_conv
        self.norm_final = model.norm_final
        self.size = model.size

    def forward(self, x, mask):
        if isinstance(x, tuple):
            x, pos_emb = x[0], x[1]
        else:
            x, pos_emb = x, None

        if self.feed_forward_macaron is not None:
            residual = x
            x = self.norm_ff_macaron(x)
            x = residual + self.feed_forward_macaron(x)

        residual = x
        x = self.norm_mha(x)

        x_q = x

        if pos_emb is not None:
            x_att = self.self_attn(x_q, x, x, pos_emb, mask)
        else:
            x_att = self.self_attn(x_q, x, x, mask)
        x = residual + x_att

        if self.conv_module is not None:
            residual = x
            x = self.norm_conv(x)
            x = residual +  self.conv_module(x)

        residual = x
        x = self.norm_ff(x)
        x = residual + self.feed_forward(x)

        x = self.norm_final(x)

        if pos_emb is not None:
            return (x, pos_emb), mask

        return x, mask

 funasr/export/models/modules/multihead_att.py

@@ -4,6 +4,7 @@
import torch
import torch.nn as nn


class MultiHeadedAttentionSANM(nn.Module):
    def __init__(self, model):
        super().__init__()
@@ -32,7 +33,6 @@
        return x.permute(0, 2, 1, 3)

    def forward_qkv(self, x):

        q_k_v = self.linear_q_k_v(x)
        q, k, v = torch.split(q_k_v, int(self.h * self.d_k), dim=-1)
        q_h = self.transpose_for_scores(q)
@@ -41,7 +41,6 @@
        return q_h, k_h, v_h, v

    def forward_fsmn(self, inputs, mask):

        # b, t, d = inputs.size()
        # mask = torch.reshape(mask, (b, -1, 1))
        inputs = inputs * mask
@@ -52,7 +51,6 @@
        x = x + inputs
        x = x * mask
        return x


    def forward_attention(self, value, scores, mask):
        scores = scores + mask
@@ -65,6 +63,7 @@
        context_layer = context_layer.view(new_context_layer_shape)
        return self.linear_out(context_layer)  # (batch, time1, d_model)


class MultiHeadedAttentionSANMDecoder(nn.Module):
    def __init__(self, model):
        super().__init__()
@@ -74,7 +73,6 @@
        self.attn = None

    def forward(self, inputs, mask, cache=None):

        # b, t, d = inputs.size()
        # mask = torch.reshape(mask, (b, -1, 1))
        inputs = inputs * mask
@@ -91,6 +89,7 @@
        x = x + inputs
        x = x * mask
        return x, cache


class MultiHeadedAttentionCrossAtt(nn.Module):
    def __init__(self, model):
@@ -133,3 +132,104 @@
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(new_context_layer_shape)
        return self.linear_out(context_layer)  # (batch, time1, d_model)


class OnnxMultiHeadedAttention(nn.Module):
    def __init__(self, model):
        super().__init__()
        self.d_k = model.d_k
        self.h = model.h
        self.linear_q = model.linear_q
        self.linear_k = model.linear_k
        self.linear_v = model.linear_v
        self.linear_out = model.linear_out
        self.attn = None
        self.all_head_size = self.h * self.d_k
    
    def forward(self, query, key, value, mask):
        q, k, v = self.forward_qkv(query, key, value)
        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
        return self.forward_attention(v, scores, mask)

    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
        new_x_shape = x.size()[:-1] + (self.h, self.d_k)
        x = x.view(new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward_qkv(self, query, key, value):
        q = self.linear_q(query)
        k = self.linear_k(key)
        v = self.linear_v(value)
        q = self.transpose_for_scores(q)
        k = self.transpose_for_scores(k)
        v = self.transpose_for_scores(v)
        return q, k, v
    
    def forward_attention(self, value, scores, mask):
        scores = scores + mask

        self.attn = torch.softmax(scores, dim=-1)
        context_layer = torch.matmul(self.attn, value)  # (batch, head, time1, d_k)
        
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(new_context_layer_shape)
        return self.linear_out(context_layer)  # (batch, time1, d_model)


class OnnxRelPosMultiHeadedAttention(OnnxMultiHeadedAttention):
    def __init__(self, model):
        super().__init__(model)
        self.linear_pos = model.linear_pos
        self.pos_bias_u = model.pos_bias_u
        self.pos_bias_v = model.pos_bias_v
    
    def forward(self, query, key, value, pos_emb, mask):
        q, k, v = self.forward_qkv(query, key, value)
        q = q.transpose(1, 2)  # (batch, time1, head, d_k)

        p = self.transpose_for_scores(self.linear_pos(pos_emb)) # (batch, head, time1, d_k)

        # (batch, head, time1, d_k)
        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
        # (batch, head, time1, d_k)
        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)

        # compute attention score
        # first compute matrix a and matrix c
        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
        # (batch, head, time1, time2)
        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))

        # compute matrix b and matrix d
        # (batch, head, time1, time1)
        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
        matrix_bd = self.rel_shift(matrix_bd)

        scores = (matrix_ac + matrix_bd) / math.sqrt(
            self.d_k
        )  # (batch, head, time1, time2)

        return self.forward_attention(v, scores, mask)

    def rel_shift(self, x):
        zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
        x_padded = torch.cat([zero_pad, x], dim=-1)

        x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
        x = x_padded[:, :, 1:].view_as(x)[
            :, :, :, : x.size(-1) // 2 + 1
        ]  # only keep the positions from 0 to time2
        return x

    def forward_attention(self, value, scores, mask):
        scores = scores + mask

        self.attn = torch.softmax(scores, dim=-1)
        context_layer = torch.matmul(self.attn, value)  # (batch, head, time1, d_k)
        
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(new_context_layer_shape)
        return self.linear_out(context_layer)  # (batch, time1, d_model)
        

 funasr/export/models/predictor/cif.py

@@ -1,9 +1,8 @@
#!/usr/bin/env python3

# -*- coding: utf-8 -*-



import torch

from torch import nn

import logging

import numpy as np





def sequence_mask(lengths, maxlen=None, dtype=torch.float32, device=None):

@@ -48,11 +47,11 @@
        alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)

        mask = mask.transpose(-1, -2).float()

        alphas = alphas * mask

		
        alphas = alphas.squeeze(-1)

		
        token_num = alphas.sum(-1)

        

        mask = mask.squeeze(-1)

        hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, mask=mask)

        acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)

        

        return acoustic_embeds, token_num, alphas, cif_peak

@@ -63,12 +62,14 @@
        

        zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)

        ones_t = torch.ones_like(zeros_t)



        mask_1 = torch.cat([mask, zeros_t], dim=1)

        mask_2 = torch.cat([ones_t, mask], dim=1)

        mask = mask_2 - mask_1

        tail_threshold = mask * tail_threshold

        alphas = torch.cat([alphas, tail_threshold], dim=1)

		
        alphas = torch.cat([alphas, zeros_t], dim=1)

        alphas = torch.add(alphas, tail_threshold)



        zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)

        hidden = torch.cat([hidden, zeros], dim=1)

        token_num = alphas.sum(dim=-1)

@@ -173,3 +174,115 @@
            max_label_len = frame_len

    frame_fires = frame_fires[:, :max_label_len, :]

    return frame_fires, fires





class CifPredictorV3(nn.Module):

    def __init__(self, model):

        super().__init__()

		
        self.pad = model.pad

        self.cif_conv1d = model.cif_conv1d

        self.cif_output = model.cif_output

        self.threshold = model.threshold

        self.smooth_factor = model.smooth_factor

        self.noise_threshold = model.noise_threshold

        self.tail_threshold = model.tail_threshold



        self.upsample_times = model.upsample_times

        self.upsample_cnn = model.upsample_cnn

        self.blstm = model.blstm

        self.cif_output2 = model.cif_output2

        self.smooth_factor2 = model.smooth_factor2

        self.noise_threshold2 = model.noise_threshold2

	
    def forward(self, hidden: torch.Tensor,

                mask: torch.Tensor,

                ):

        h = hidden

        context = h.transpose(1, 2)

        queries = self.pad(context)

        output = torch.relu(self.cif_conv1d(queries))

        output = output.transpose(1, 2)

		
        output = self.cif_output(output)

        alphas = torch.sigmoid(output)

        alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)

        mask = mask.transpose(-1, -2).float()

        alphas = alphas * mask

        alphas = alphas.squeeze(-1)

        token_num = alphas.sum(-1)

		
        mask = mask.squeeze(-1)

        hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, mask=mask)

        acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)

		
        return acoustic_embeds, token_num, alphas, cif_peak

	
    def get_upsample_timestmap(self, hidden, mask=None, token_num=None):

        h = hidden

        b = hidden.shape[0]

        context = h.transpose(1, 2)



        # generate alphas2

        _output = context

        output2 = self.upsample_cnn(_output)

        output2 = output2.transpose(1, 2)

        output2, (_, _) = self.blstm(output2)

        alphas2 = torch.sigmoid(self.cif_output2(output2))

        alphas2 = torch.nn.functional.relu(alphas2 * self.smooth_factor2 - self.noise_threshold2)

		
        mask = mask.repeat(1, self.upsample_times, 1).transpose(-1, -2).reshape(alphas2.shape[0], -1)

        mask = mask.unsqueeze(-1)

        alphas2 = alphas2 * mask

        alphas2 = alphas2.squeeze(-1)

        _token_num = alphas2.sum(-1)

        alphas2 *= (token_num / _token_num)[:, None].repeat(1, alphas2.size(1))

        # upsampled alphas and cif_peak

        us_alphas = alphas2

        us_cif_peak = cif_wo_hidden(us_alphas, self.threshold - 1e-4)

        return us_alphas, us_cif_peak



    def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):

        b, t, d = hidden.size()

        tail_threshold = self.tail_threshold

		
        zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)

        ones_t = torch.ones_like(zeros_t)



        mask_1 = torch.cat([mask, zeros_t], dim=1)

        mask_2 = torch.cat([ones_t, mask], dim=1)

        mask = mask_2 - mask_1

        tail_threshold = mask * tail_threshold

        alphas = torch.cat([alphas, zeros_t], dim=1)

        alphas = torch.add(alphas, tail_threshold)



        zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)

        hidden = torch.cat([hidden, zeros], dim=1)

        token_num = alphas.sum(dim=-1)

        token_num_floor = torch.floor(token_num)

		
        return hidden, alphas, token_num_floor





@torch.jit.script

def cif_wo_hidden(alphas, threshold: float):

    batch_size, len_time = alphas.size()



    # loop varss

    integrate = torch.zeros([batch_size], dtype=alphas.dtype, device=alphas.device)

    # intermediate vars along time

    list_fires = []



    for t in range(len_time):

        alpha = alphas[:, t]



        integrate += alpha

        list_fires.append(integrate)



        fire_place = integrate >= threshold

        integrate = torch.where(fire_place,

                                integrate - torch.ones([batch_size], device=alphas.device),

                                integrate)



    fires = torch.stack(list_fires, 1)

    return fires

 funasr/runtime/python/onnxruntime/demo.py

@@ -1,10 +1,12 @@

from rapid_paraformer import Paraformer
from rapid_paraformer import BiCifParaformer

model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model = Paraformer(model_dir, batch_size=1)
model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
# model = Paraformer(model_dir, batch_size=1)
model = BiCifParaformer(model_dir, batch_size=1)

wav_path = ['/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']
wav_path = ['/Users/shixian/code/funasr2/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch/example/asr_example.wav']

result = model(wav_path)
print(result)

 funasr/runtime/python/onnxruntime/rapid_paraformer/__init__.py

@@ -2,3 +2,4 @@
# @Author: SWHL
# @Contact: liekkaskono@163.com
from .paraformer_onnx import Paraformer
from .paraformer_onnx import BiCifParaformer

 funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py

@@ -5,6 +5,7 @@
from pathlib import Path
from typing import List, Union, Tuple

import copy
import librosa
import numpy as np

@@ -13,6 +14,7 @@
                          read_yaml)
from .utils.postprocess_utils import sentence_postprocess
from .utils.frontend import WavFrontend
from funasr.utils.timestamp_tools import time_stamp_lfr6_pl

logging = get_logger()

@@ -134,8 +136,67 @@

        # Change integer-ids to tokens
        token = self.converter.ids2tokens(token_int)
        token = token[:valid_token_num-1]
        # token = token[:valid_token_num-1]
        texts = sentence_postprocess(token)
        text = texts[0]
        # text = self.tokenizer.tokens2text(token)
        return text


class BiCifParaformer(Paraformer):
    def infer(self, feats: np.ndarray,
              feats_len: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
        am_scores, token_nums, us_alphas, us_cif_peak = self.ort_infer([feats, feats_len])
        return am_scores, token_nums, us_alphas, us_cif_peak
    def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
        waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
        waveform_nums = len(waveform_list)

        asr_res = []
        for beg_idx in range(0, waveform_nums, self.batch_size):
            res = {}
            end_idx = min(waveform_nums, beg_idx + self.batch_size)
            feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
            am_scores, valid_token_lens, us_alphas, us_cif_peak = self.infer(feats, feats_len)

            try:
                am_scores, valid_token_lens, us_alphas, us_cif_peak = self.infer(feats, feats_len)
            except ONNXRuntimeError:
                #logging.warning(traceback.format_exc())
                logging.warning("input wav is silence or noise")
                preds = ['']
            else:
                token = self.decode(am_scores, valid_token_lens)
                timestamp = time_stamp_lfr6_pl(us_alphas, us_cif_peak, copy.copy(token[0]), log=False)
                texts = sentence_postprocess(token[0], timestamp)
                # texts = sentence_postprocess(token[0])
                text = texts[0]
            res['text'] = text
            res['timestamp'] = timestamp
            asr_res.append(res)

        return asr_res

    def decode_one(self,
                   am_score: np.ndarray,
                   valid_token_num: int) -> List[str]:
        yseq = am_score.argmax(axis=-1)
        score = am_score.max(axis=-1)
        score = np.sum(score, axis=-1)

        # pad with mask tokens to ensure compatibility with sos/eos tokens
        # asr_model.sos:1  asr_model.eos:2
        yseq = np.array([1] + yseq.tolist() + [2])
        hyp = Hypothesis(yseq=yseq, score=score)

        # remove sos/eos and get results
        last_pos = -1
        token_int = hyp.yseq[1:last_pos].tolist()

        # remove blank symbol id, which is assumed to be 0
        token_int = list(filter(lambda x: x not in (0, 2), token_int))

        # Change integer-ids to tokens
        token = self.converter.ids2tokens(token_int)
        # token = token[:valid_token_num-1]
        return token

 funasr/utils/timestamp_tools.py

@@ -4,6 +4,7 @@
import numpy as np
from typing import Any, List, Tuple, Union


def time_stamp_lfr6_pl(us_alphas, us_cif_peak, char_list, begin_time=0.0, end_time=None):
    if not len(char_list):
        return []