Merge pull request #2269 from modelscope/dev_sx2

Shi Xian

2024-12-05 22b928dd3ff37ccee57ab2b5c2e4fcda4d33d24d

 funasr/models/sense_voice/model.py

@@ -1,6 +1,3 @@
import logging
from dataclasses import dataclass
from typing import Dict
from typing import Iterable, Optional
import types
import time
@@ -10,45 +7,653 @@
from torch import Tensor
from torch import nn
from torch.cuda.amp import autocast
from funasr.metrics.compute_acc import compute_accuracy
from funasr.metrics.compute_acc import compute_accuracy, th_accuracy
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
from funasr.train_utils.device_funcs import force_gatherable
from . import whisper_lib as whisper

from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.ctc.ctc import CTC

from funasr.register import tables


@tables.register("model_classes", "SenseVoice")
class SenseVoice(nn.Module):
from funasr.models.paraformer.search import Hypothesis
from funasr.models.sense_voice.utils.ctc_alignment import ctc_forced_align


class SinusoidalPositionEncoder(torch.nn.Module):
    """ """

    def __int__(self, d_model=80, dropout_rate=0.1):
        pass

    def encode(
        self, positions: torch.Tensor = None, depth: int = None, dtype: torch.dtype = torch.float32
    ):
        batch_size = positions.size(0)
        positions = positions.type(dtype)
        device = positions.device
        log_timescale_increment = torch.log(torch.tensor([10000], dtype=dtype, device=device)) / (
            depth / 2 - 1
        )
        inv_timescales = torch.exp(
            torch.arange(depth / 2, device=device).type(dtype) * (-log_timescale_increment)
        )
        inv_timescales = torch.reshape(inv_timescales, [batch_size, -1])
        scaled_time = torch.reshape(positions, [1, -1, 1]) * torch.reshape(
            inv_timescales, [1, 1, -1]
        )
        encoding = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=2)
        return encoding.type(dtype)

    def forward(self, x):
        batch_size, timesteps, input_dim = x.size()
        positions = torch.arange(1, timesteps + 1, device=x.device)[None, :]
        position_encoding = self.encode(positions, input_dim, x.dtype).to(x.device)

        return x + position_encoding


class PositionwiseFeedForward(torch.nn.Module):
    """Positionwise feed forward layer.

    Args:
        idim (int): Input dimenstion.
        hidden_units (int): The number of hidden units.
        dropout_rate (float): Dropout rate.

    """

    def __init__(self, idim, hidden_units, dropout_rate, activation=torch.nn.ReLU()):
        """Construct an PositionwiseFeedForward object."""
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = torch.nn.Linear(idim, hidden_units)
        self.w_2 = torch.nn.Linear(hidden_units, idim)
        self.dropout = torch.nn.Dropout(dropout_rate)
        self.activation = activation

    def forward(self, x):
        """Forward function."""
        return self.w_2(self.dropout(self.activation(self.w_1(x))))


class MultiHeadedAttentionSANM(nn.Module):
    """Multi-Head Attention layer.

    Args:
        n_head (int): The number of heads.
        n_feat (int): The number of features.
        dropout_rate (float): Dropout rate.

    """

    def __init__(
        self,
        n_head,
        in_feat,
        n_feat,
        dropout_rate,
        kernel_size,
        sanm_shfit=0,
        lora_list=None,
        lora_rank=8,
        lora_alpha=16,
        lora_dropout=0.1,
    ):
        """Construct an MultiHeadedAttention object."""
        super().__init__()
        assert n_feat % n_head == 0
        # We assume d_v always equals d_k
        self.d_k = n_feat // n_head
        self.h = n_head
        # self.linear_q = nn.Linear(n_feat, n_feat)
        # self.linear_k = nn.Linear(n_feat, n_feat)
        # self.linear_v = nn.Linear(n_feat, n_feat)

        self.linear_out = nn.Linear(n_feat, n_feat)
        self.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
        self.attn = None
        self.dropout = nn.Dropout(p=dropout_rate)

        self.fsmn_block = nn.Conv1d(
            n_feat, n_feat, kernel_size, stride=1, padding=0, groups=n_feat, bias=False
        )
        # padding
        left_padding = (kernel_size - 1) // 2
        if sanm_shfit > 0:
            left_padding = left_padding + sanm_shfit
        right_padding = kernel_size - 1 - left_padding
        self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)

    def forward_fsmn(self, inputs, mask, mask_shfit_chunk=None):
        b, t, d = inputs.size()
        if mask is not None:
            mask = torch.reshape(mask, (b, -1, 1))
            if mask_shfit_chunk is not None:
                mask = mask * mask_shfit_chunk
            inputs = inputs * mask

        x = inputs.transpose(1, 2)
        x = self.pad_fn(x)
        x = self.fsmn_block(x)
        x = x.transpose(1, 2)
        x += inputs
        x = self.dropout(x)
        if mask is not None:
            x = x * mask
        return x

    def forward_qkv(self, x):
        """Transform query, key and value.

        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).

        Returns:
            torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
            torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
            torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).

        """
        b, t, d = x.size()
        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 = torch.reshape(q, (b, t, self.h, self.d_k)).transpose(
            1, 2
        )  # (batch, head, time1, d_k)
        k_h = torch.reshape(k, (b, t, self.h, self.d_k)).transpose(
            1, 2
        )  # (batch, head, time2, d_k)
        v_h = torch.reshape(v, (b, t, self.h, self.d_k)).transpose(
            1, 2
        )  # (batch, head, time2, d_k)

        return q_h, k_h, v_h, v

    def forward_attention(self, value, scores, mask, mask_att_chunk_encoder=None):
        """Compute attention context vector.

        Args:
            value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
            scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
            mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).

        Returns:
            torch.Tensor: Transformed value (#batch, time1, d_model)
                weighted by the attention score (#batch, time1, time2).

        """
        n_batch = value.size(0)
        if mask is not None:
            if mask_att_chunk_encoder is not None:
                mask = mask * mask_att_chunk_encoder

            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)

            min_value = -float(
                "inf"
            )  # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
            scores = scores.masked_fill(mask, min_value)
            attn = torch.softmax(scores, dim=-1).masked_fill(
                mask, 0.0
            )  # (batch, head, time1, time2)
        else:
            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)

        p_attn = self.dropout(attn)
        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
        x = (
            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
        )  # (batch, time1, d_model)

        return self.linear_out(x)  # (batch, time1, d_model)

    def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
        """Compute scaled dot product attention.

        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).
            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
                (#batch, time1, time2).

        Returns:
            torch.Tensor: Output tensor (#batch, time1, d_model).

        """
        q_h, k_h, v_h, v = self.forward_qkv(x)
        fsmn_memory = self.forward_fsmn(v, mask, mask_shfit_chunk)
        q_h = q_h * self.d_k ** (-0.5)
        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
        att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
        return att_outs + fsmn_memory

    def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
        """Compute scaled dot product attention.

        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).
            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
                (#batch, time1, time2).

        Returns:
            torch.Tensor: Output tensor (#batch, time1, d_model).

        """
        q_h, k_h, v_h, v = self.forward_qkv(x)
        if chunk_size is not None and look_back > 0 or look_back == -1:
            if cache is not None:
                k_h_stride = k_h[:, :, : -(chunk_size[2]), :]
                v_h_stride = v_h[:, :, : -(chunk_size[2]), :]
                k_h = torch.cat((cache["k"], k_h), dim=2)
                v_h = torch.cat((cache["v"], v_h), dim=2)

                cache["k"] = torch.cat((cache["k"], k_h_stride), dim=2)
                cache["v"] = torch.cat((cache["v"], v_h_stride), dim=2)
                if look_back != -1:
                    cache["k"] = cache["k"][:, :, -(look_back * chunk_size[1]) :, :]
                    cache["v"] = cache["v"][:, :, -(look_back * chunk_size[1]) :, :]
            else:
                cache_tmp = {
                    "k": k_h[:, :, : -(chunk_size[2]), :],
                    "v": v_h[:, :, : -(chunk_size[2]), :],
                }
                cache = cache_tmp
        fsmn_memory = self.forward_fsmn(v, None)
        q_h = q_h * self.d_k ** (-0.5)
        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
        att_outs = self.forward_attention(v_h, scores, None)
        return att_outs + fsmn_memory, cache


class LayerNorm(nn.LayerNorm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def forward(self, input):
        output = F.layer_norm(
            input.float(),
            self.normalized_shape,
            self.weight.float() if self.weight is not None else None,
            self.bias.float() if self.bias is not None else None,
            self.eps,
        )
        return output.type_as(input)


def sequence_mask(lengths, maxlen=None, dtype=torch.float32, device=None):
    if maxlen is None:
        maxlen = lengths.max()
    row_vector = torch.arange(0, maxlen, 1).to(lengths.device)
    matrix = torch.unsqueeze(lengths, dim=-1)
    mask = row_vector < matrix
    mask = mask.detach()

    return mask.type(dtype).to(device) if device is not None else mask.type(dtype)


class EncoderLayerSANM(nn.Module):
    def __init__(
        self,
        in_size,
        size,
        self_attn,
        feed_forward,
        dropout_rate,
        normalize_before=True,
        concat_after=False,
        stochastic_depth_rate=0.0,
    ):
        """Construct an EncoderLayer object."""
        super(EncoderLayerSANM, self).__init__()
        self.self_attn = self_attn
        self.feed_forward = feed_forward
        self.norm1 = LayerNorm(in_size)
        self.norm2 = LayerNorm(size)
        self.dropout = nn.Dropout(dropout_rate)
        self.in_size = in_size
        self.size = size
        self.normalize_before = normalize_before
        self.concat_after = concat_after
        if self.concat_after:
            self.concat_linear = nn.Linear(size + size, size)
        self.stochastic_depth_rate = stochastic_depth_rate
        self.dropout_rate = dropout_rate

    def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
        """Compute encoded features.

        Args:
            x_input (torch.Tensor): Input tensor (#batch, time, size).
            mask (torch.Tensor): Mask tensor for the input (#batch, time).
            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).

        Returns:
            torch.Tensor: Output tensor (#batch, time, size).
            torch.Tensor: Mask tensor (#batch, time).

        """
        skip_layer = False
        # with stochastic depth, residual connection `x + f(x)` becomes
        # `x <- x + 1 / (1 - p) * f(x)` at training time.
        stoch_layer_coeff = 1.0
        if self.training and self.stochastic_depth_rate > 0:
            skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
            stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)

        if skip_layer:
            if cache is not None:
                x = torch.cat([cache, x], dim=1)
            return x, mask

        residual = x
        if self.normalize_before:
            x = self.norm1(x)

        if self.concat_after:
            x_concat = torch.cat(
                (
                    x,
                    self.self_attn(
                        x,
                        mask,
                        mask_shfit_chunk=mask_shfit_chunk,
                        mask_att_chunk_encoder=mask_att_chunk_encoder,
                    ),
                ),
                dim=-1,
            )
            if self.in_size == self.size:
                x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
            else:
                x = stoch_layer_coeff * self.concat_linear(x_concat)
        else:
            if self.in_size == self.size:
                x = residual + stoch_layer_coeff * self.dropout(
                    self.self_attn(
                        x,
                        mask,
                        mask_shfit_chunk=mask_shfit_chunk,
                        mask_att_chunk_encoder=mask_att_chunk_encoder,
                    )
                )
            else:
                x = stoch_layer_coeff * self.dropout(
                    self.self_attn(
                        x,
                        mask,
                        mask_shfit_chunk=mask_shfit_chunk,
                        mask_att_chunk_encoder=mask_att_chunk_encoder,
                    )
                )
        if not self.normalize_before:
            x = self.norm1(x)

        residual = x
        if self.normalize_before:
            x = self.norm2(x)
        x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
        if not self.normalize_before:
            x = self.norm2(x)

        return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder

    def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
        """Compute encoded features.

        Args:
            x_input (torch.Tensor): Input tensor (#batch, time, size).
            mask (torch.Tensor): Mask tensor for the input (#batch, time).
            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).

        Returns:
            torch.Tensor: Output tensor (#batch, time, size).
            torch.Tensor: Mask tensor (#batch, time).

        """

        residual = x
        if self.normalize_before:
            x = self.norm1(x)

        if self.in_size == self.size:
            attn, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)
            x = residual + attn
        else:
            x, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)

        if not self.normalize_before:
            x = self.norm1(x)

        residual = x
        if self.normalize_before:
            x = self.norm2(x)
        x = residual + self.feed_forward(x)
        if not self.normalize_before:
            x = self.norm2(x)

        return x, cache


@tables.register("encoder_classes", "SenseVoiceEncoderSmall")
class SenseVoiceEncoderSmall(nn.Module):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    SCAMA: Streaming chunk-aware multihead attention for online end-to-end speech recognition
    https://arxiv.org/abs/2006.01713
    """

    def __init__(
        self,
        input_size: int,
        output_size: int = 256,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        tp_blocks: int = 0,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        attention_dropout_rate: float = 0.0,
        stochastic_depth_rate: float = 0.0,
        input_layer: Optional[str] = "conv2d",
        pos_enc_class=SinusoidalPositionEncoder,
        normalize_before: bool = True,
        concat_after: bool = False,
        positionwise_layer_type: str = "linear",
        positionwise_conv_kernel_size: int = 1,
        padding_idx: int = -1,
        kernel_size: int = 11,
        sanm_shfit: int = 0,
        selfattention_layer_type: str = "sanm",
        **kwargs,
    ):
        super().__init__()
        self._output_size = output_size

        self.embed = SinusoidalPositionEncoder()

        self.normalize_before = normalize_before

        positionwise_layer = PositionwiseFeedForward
        positionwise_layer_args = (
            output_size,
            linear_units,
            dropout_rate,
        )

        encoder_selfattn_layer = MultiHeadedAttentionSANM
        encoder_selfattn_layer_args0 = (
            attention_heads,
            input_size,
            output_size,
            attention_dropout_rate,
            kernel_size,
            sanm_shfit,
        )
        encoder_selfattn_layer_args = (
            attention_heads,
            output_size,
            output_size,
            attention_dropout_rate,
            kernel_size,
            sanm_shfit,
        )

        self.encoders0 = nn.ModuleList(
            [
                EncoderLayerSANM(
                    input_size,
                    output_size,
                    encoder_selfattn_layer(*encoder_selfattn_layer_args0),
                    positionwise_layer(*positionwise_layer_args),
                    dropout_rate,
                )
                for i in range(1)
            ]
        )
        self.encoders = nn.ModuleList(
            [
                EncoderLayerSANM(
                    output_size,
                    output_size,
                    encoder_selfattn_layer(*encoder_selfattn_layer_args),
                    positionwise_layer(*positionwise_layer_args),
                    dropout_rate,
                )
                for i in range(num_blocks - 1)
            ]
        )

        self.tp_encoders = nn.ModuleList(
            [
                EncoderLayerSANM(
                    output_size,
                    output_size,
                    encoder_selfattn_layer(*encoder_selfattn_layer_args),
                    positionwise_layer(*positionwise_layer_args),
                    dropout_rate,
                )
                for i in range(tp_blocks)
            ]
        )

        self.after_norm = LayerNorm(output_size)

        self.tp_norm = LayerNorm(output_size)

    def output_size(self) -> int:
        return self._output_size

    def forward(
        self,
        xs_pad: torch.Tensor,
        ilens: torch.Tensor,
    ):
        """Embed positions in tensor."""
        maxlen = xs_pad.shape[1]
        masks = sequence_mask(ilens, maxlen = maxlen, device=ilens.device)[:, None, :]

        xs_pad *= self.output_size() ** 0.5

        xs_pad = self.embed(xs_pad)

        # forward encoder1
        for layer_idx, encoder_layer in enumerate(self.encoders0):
            encoder_outs = encoder_layer(xs_pad, masks)
            xs_pad, masks = encoder_outs[0], encoder_outs[1]

        for layer_idx, encoder_layer in enumerate(self.encoders):
            encoder_outs = encoder_layer(xs_pad, masks)
            xs_pad, masks = encoder_outs[0], encoder_outs[1]

        xs_pad = self.after_norm(xs_pad)

        # forward encoder2
        olens = masks.squeeze(1).sum(1).int()

        for layer_idx, encoder_layer in enumerate(self.tp_encoders):
            encoder_outs = encoder_layer(xs_pad, masks)
            xs_pad, masks = encoder_outs[0], encoder_outs[1]

        xs_pad = self.tp_norm(xs_pad)
        return xs_pad, olens


@tables.register("model_classes", "SenseVoiceSmall")
class SenseVoiceSmall(nn.Module):
    """CTC-attention hybrid Encoder-Decoder model"""

    def __init__(
        self,
        specaug: str = None,
        specaug_conf: dict = None,
        normalize: str = None,
        normalize_conf: dict = None,
        encoder: str = None,
        encoder_conf: dict = None,
        ctc_conf: dict = None,
        input_size: int = 80,
        vocab_size: int = -1,
        ignore_id: int = -1,
        blank_id: int = 0,
        sos: int = 1,
        eos: int = 2,
        length_normalized_loss: bool = False,
        **kwargs,
    ):

        super().__init__()

        dims = kwargs.get("dims", {})
        dims = whisper.model.ModelDimensions(**dims)
        model = whisper.model.Whisper(dims=dims)
        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**specaug_conf)
        if normalize is not None:
            normalize_class = tables.normalize_classes.get(normalize)
            normalize = normalize_class(**normalize_conf)
        encoder_class = tables.encoder_classes.get(encoder)
        encoder = encoder_class(input_size=input_size, **encoder_conf)
        encoder_output_size = encoder.output_size()

        # encoder
        model.encoder.downsample_rate = kwargs.get("downsample_rate", 4)
        model.encoder.use_padmask = kwargs.get("use_padmask", True)
        from .encoder import sense_voice_encode_forward
        if ctc_conf is None:
            ctc_conf = {}
        ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)

        model.encoder.forward = types.MethodType(sense_voice_encode_forward, model.encoder)
        self.blank_id = blank_id
        self.sos = sos if sos is not None else vocab_size - 1
        self.eos = eos if eos is not None else vocab_size - 1
        self.vocab_size = vocab_size
        self.ignore_id = ignore_id
        self.specaug = specaug
        self.normalize = normalize
        self.encoder = encoder
        self.error_calculator = None

        # decoder
        model.decoder.use_padmask = kwargs.get("use_padmask", True)
        from .decoder import sense_voice_decode_forward
        self.ctc = ctc

        model.decoder.forward = types.MethodType(sense_voice_decode_forward, model.decoder)
        self.length_normalized_loss = length_normalized_loss
        self.encoder_output_size = encoder_output_size

        self.model = model
        self.lid_dict = {"auto": 0, "zh": 3, "en": 4, "yue": 7, "ja": 11, "ko": 12, "nospeech": 13}
        self.lid_int_dict = {24884: 3, 24885: 4, 24888: 7, 24892: 11, 24896: 12, 24992: 13}
        self.textnorm_dict = {"withitn": 14, "woitn": 15}
        self.textnorm_int_dict = {25016: 14, 25017: 15}
        self.embed = torch.nn.Embedding(
            7 + len(self.lid_dict) + len(self.textnorm_dict), input_size
        )
        self.emo_dict = {
            "unk": 25009,
            "happy": 25001,
            "sad": 25002,
            "angry": 25003,
            "neutral": 25004,
        }

        self.encoder_output_size = self.model.dims.n_audio_state

        self.activation_checkpoint = kwargs.get("activation_checkpoint", False)
        self.ignore_id = kwargs.get("ignore_id", -1)
        self.vocab_size = kwargs.get("vocab_size", -1)
        self.length_normalized_loss = kwargs.get("length_normalized_loss", True)
        self.criterion_att = LabelSmoothingLoss(
            size=self.vocab_size,
            padding_idx=self.ignore_id,
@@ -56,11 +661,13 @@
            normalize_length=self.length_normalized_loss,
        )

        specaug = kwargs.get("specaug", None)
        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**kwargs.get("specaug_conf", {}))
        self.specaug = specaug
    @staticmethod
    def from_pretrained(model: str = None, **kwargs):
        from funasr import AutoModel

        model, kwargs = AutoModel.build_model(model=model, trust_remote_code=True, **kwargs)

        return model, kwargs

    def forward(
        self,
@@ -70,8 +677,13 @@
        text_lengths: torch.Tensor,
        **kwargs,
    ):
        target_mask = kwargs.get("target_mask", None)

        """Encoder + Decoder + Calc loss
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                text: (Batch, Length)
                text_lengths: (Batch,)
        """
        # import pdb;
        # pdb.set_trace()
        if len(text_lengths.size()) > 1:
@@ -81,23 +693,25 @@

        batch_size = speech.shape[0]

        if self.activation_checkpoint:
            from torch.utils.checkpoint import checkpoint
        # 1. Encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, text)

            encoder_out, encoder_out_lens = checkpoint(
                self.encode, speech, speech_lengths, use_reentrant=False
            )
        else:
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
        loss_ctc, cer_ctc = None, None
        loss_rich, acc_rich = None, None
        stats = dict()

        loss_att, acc_att, cer_att, wer_att = self._calc_att_loss(
            encoder_out, encoder_out_lens, text, text_lengths, target_mask=target_mask
        loss_ctc, cer_ctc = self._calc_ctc_loss(
            encoder_out[:, 4:, :], encoder_out_lens - 4, text[:, 4:], text_lengths - 4
        )
        loss = loss_att
        stats = {}
        stats["acc"] = acc_att
        stats["loss"] = torch.clone(loss.detach())
        stats["batch_size"] = batch_size

        loss_rich, acc_rich = self._calc_rich_ce_loss(encoder_out[:, :4, :], text[:, :4])

        loss = loss_ctc + loss_rich
        # Collect total loss stats
        stats["loss_ctc"] = torch.clone(loss_ctc.detach()) if loss_ctc is not None else None
        stats["loss_rich"] = torch.clone(loss_rich.detach()) if loss_rich is not None else None
        stats["loss"] = torch.clone(loss.detach()) if loss is not None else None
        stats["acc_rich"] = acc_rich

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
@@ -106,308 +720,101 @@
        return loss, stats, weight

    def encode(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        **kwargs,
    ):
        """Encoder. Note that this method is used by asr_inference.py
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                ind: int
        """
        with autocast(False):

            # Data augmentation
            if self.specaug is not None and self.training:
                speech, speech_lengths = self.specaug(speech, speech_lengths)

        # Forward encoder
        encoder_out, encoder_out_lens = self.model.encoder(speech.permute(0, 2, 1), speech_lengths)

        return encoder_out, encoder_out_lens

    def _calc_att_loss(
        self,
        encoder_out: torch.Tensor,
        encoder_out_lens: torch.Tensor,
        ys_pad: torch.Tensor,
        ys_pad_lens: torch.Tensor,
        **kwargs,
    ):
        target_mask = kwargs.get("target_mask", None)
        stats = {}

        # 1. Forward decoder
        decoder_out = self.model.decoder(
            x=ys_pad, xa=encoder_out, hlens=encoder_out_lens, ys_in_lens=ys_pad_lens
        )

        # 2. Compute attention loss
        mask = torch.ones_like(ys_pad) * (-1)
        ys_pad_mask = (ys_pad * target_mask + mask * (1 - target_mask)).to(torch.int64)
        ys_pad_mask[ys_pad_mask == 0] = -1
        loss_att = self.criterion_att(decoder_out[:, :-1, :], ys_pad_mask[:, 1:])

        with torch.no_grad():
            preds = torch.argmax(decoder_out, -1)
            acc_att = compute_accuracy(
                preds[:, :-1], ys_pad_mask[:, 1:], ignore_label=self.ignore_id
            )

        return loss_att, acc_att, None, None

    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):
        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        if frontend is None and not hasattr(self, "frontend"):
            frontend_class = tables.frontend_classes.get("WhisperFrontend")
            frontend = frontend_class(
                n_mels=self.model.dims.n_mels, do_pad_trim=kwargs.get("do_pad_trim", True)
            )
            self.frontend = frontend
        else:
            frontend = frontend if frontend is not None else self.frontend

        meta_data = {}
        if (
            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
        ):  # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
            if speech_lengths is None:
                speech_lengths = speech.shape[1]
        else:
            # extract fbank feats
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs if hasattr(frontend, "fs") else 16000,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=tokenizer,
            )
            time2 = time.perf_counter()
            meta_data["load_data"] = f"{time2 - time1:0.3f}"
            speech, speech_lengths = extract_fbank(
                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            frame_shift = frontend.frame_shift if hasattr(frontend, "frame_shift") else 10
            lfr_n = frontend.lfr_n if hasattr(frontend, "lfr_n") else 1
            meta_data["batch_data_time"] = speech_lengths.sum().item() * frame_shift * lfr_n / 1000

        speech = speech.to(device=kwargs["device"])[0, :, :]
        speech_lengths = speech_lengths.to(device=kwargs["device"])

        DecodingOptions = kwargs.get("DecodingOptions", {})
        task = DecodingOptions.get("task", "ASR")
        if isinstance(task, str):
            task = [task]
        task = "".join([f"<|{x}|>" for x in task])
        initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")
        DecodingOptions["initial_prompt"] = initial_prompt

        language = DecodingOptions.get("language", None)
        language = None if language == "auto" else language
        DecodingOptions["language"] = language

        DecodingOptions["vocab_path"] = kwargs["tokenizer_conf"].get("vocab_path", None)

        if "without_timestamps" not in DecodingOptions:
            DecodingOptions["without_timestamps"] = True

        options = whisper.DecodingOptions(**DecodingOptions)

        result = whisper.decode(self.model, speech, options)
        text = f"{result.text}"
        results = []
        result_i = {"key": key[0], "text": text}

        results.append(result_i)

        return results, meta_data


@tables.register("model_classes", "SenseVoiceRWKV")
class SenseVoiceRWKV(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__()

        dims = kwargs.get("dims", {})
        dims = whisper.model.ModelDimensions(**dims)
        model = whisper.model.Whisper(dims=dims)

        # encoder
        model.encoder.downsample_rate = kwargs.get("downsample_rate", 4)
        model.encoder.use_padmask = kwargs.get("use_padmask", True)
        from .encoder import sense_voice_encode_forward

        model.encoder.forward = types.MethodType(sense_voice_encode_forward, model.encoder)

        # decoder
        del model.decoder
        decoder = kwargs.get("decoder", "SenseVoiceDecoder")
        decoder_class = tables.decoder_classes.get(decoder)
        decoder = decoder_class(
            n_vocab=dims.n_vocab,
            n_ctx=dims.n_text_ctx,
            n_state=dims.n_text_state,
            n_head=dims.n_text_head,
            n_layer=dims.n_text_layer,
            **kwargs.get("decoder_conf"),
        )
        model.decoder = decoder

        self.model = model

        self.encoder_output_size = self.model.dims.n_audio_state

        self.activation_checkpoint = kwargs.get("activation_checkpoint", False)
        self.ignore_id = kwargs.get("ignore_id", -1)
        self.vocab_size = kwargs.get("vocab_size", -1)
        self.length_normalized_loss = kwargs.get("length_normalized_loss", True)
        self.criterion_att = LabelSmoothingLoss(
            size=self.vocab_size,
            padding_idx=self.ignore_id,
            smoothing=kwargs.get("lsm_weight", 0.0),
            normalize_length=self.length_normalized_loss,
        )

        specaug = kwargs.get("specaug", None)
        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**kwargs.get("specaug_conf", {}))
        self.specaug = specaug

    def forward(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        **kwargs,
    ):
        target_mask = kwargs.get("target_mask", None)

        # import pdb;
        # pdb.set_trace()
        if len(text_lengths.size()) > 1:
            text_lengths = text_lengths[:, 0]
        if len(speech_lengths.size()) > 1:
            speech_lengths = speech_lengths[:, 0]

        batch_size, frames, _ = speech.shape

        if self.activation_checkpoint:
            from torch.utils.checkpoint import checkpoint

            encoder_out, encoder_out_lens = checkpoint(
                self.encode, speech, speech_lengths, use_reentrant=False
            )
        else:
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        loss_att, acc_att, cer_att, wer_att = self._calc_att_loss(
            encoder_out, encoder_out_lens, text, text_lengths, target_mask=target_mask
        )
        loss = loss_att
        stats = {}
        stats["acc"] = acc_att
        stats["loss"] = torch.clone(loss.detach())
        stats["batch_size"] = batch_size
        stats["batch_size_x_frames"] = frames * batch_size

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = int((text_lengths + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def encode(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        **kwargs,
    ):
        """Encoder. Note that this method is used by asr_inference.py
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                ind: int
        """
        with autocast(False):
            # Data augmentation
            if self.specaug is not None and self.training:
                speech, speech_lengths = self.specaug(speech, speech_lengths)

        # Forward encoder
        encoder_out, encoder_out_lens = self.model.encoder(speech.permute(0, 2, 1), speech_lengths)
        # Data augmentation
        if self.specaug is not None and self.training:
            speech, speech_lengths = self.specaug(speech, speech_lengths)

        # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
        if self.normalize is not None:
            speech, speech_lengths = self.normalize(speech, speech_lengths)

        lids = torch.LongTensor(
            [
                [
                    (
                        self.lid_int_dict[int(lid)]
                        if torch.rand(1) > 0.2 and int(lid) in self.lid_int_dict
                        else 0
                    )
                ]
                for lid in text[:, 0]
            ]
        ).to(speech.device)
        language_query = self.embed(lids)

        styles = torch.LongTensor(
            [[self.textnorm_int_dict[int(style)]] for style in text[:, 3]]
        ).to(speech.device)
        style_query = self.embed(styles)
        speech = torch.cat((style_query, speech), dim=1)
        speech_lengths += 1

        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
            speech.size(0), 1, 1
        )
        input_query = torch.cat((language_query, event_emo_query), dim=1)
        speech = torch.cat((input_query, speech), dim=1)
        speech_lengths += 3

        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)

        return encoder_out, encoder_out_lens

    def _calc_att_loss(
    def _calc_ctc_loss(
        self,
        encoder_out: torch.Tensor,
        encoder_out_lens: torch.Tensor,
        ys_pad: torch.Tensor,
        ys_pad_lens: torch.Tensor,
        **kwargs,
    ):
        target_mask = kwargs.get("target_mask", None)
        stats = {}
        # Calc CTC loss
        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)

        # 1. Forward decoder
        decoder_out = self.model.decoder(
            x=ys_pad, xa=encoder_out, hlens=encoder_out_lens, ys_in_lens=ys_pad_lens
        # Calc CER using CTC
        cer_ctc = None
        if not self.training and self.error_calculator is not None:
            ys_hat = self.ctc.argmax(encoder_out).data
            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
        return loss_ctc, cer_ctc

    def _calc_rich_ce_loss(
        self,
        encoder_out: torch.Tensor,
        ys_pad: torch.Tensor,
    ):
        decoder_out = self.ctc.ctc_lo(encoder_out)
        # 2. Compute attention loss
        loss_rich = self.criterion_att(decoder_out, ys_pad.contiguous())
        acc_rich = th_accuracy(
            decoder_out.view(-1, self.vocab_size),
            ys_pad.contiguous(),
            ignore_label=self.ignore_id,
        )

        # 2. Compute attention loss
        mask = torch.ones_like(ys_pad) * (-1)
        ys_pad_mask = (ys_pad * target_mask + mask * (1 - target_mask)).to(torch.int64)
        ys_pad_mask[ys_pad_mask == 0] = -1
        loss_att = self.criterion_att(decoder_out[:, :-1, :], ys_pad_mask[:, 1:])

        with torch.no_grad():
            preds = torch.argmax(decoder_out, -1)
            acc_att = compute_accuracy(
                preds[:, :-1], ys_pad_mask[:, 1:], ignore_label=self.ignore_id
            )

        return loss_att, acc_att, None, None
        return loss_rich, acc_rich

    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        key: list = ["wav_file_tmp_name"],
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):
        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        if frontend is None and not hasattr(self, "frontend"):
            frontend_class = tables.frontend_classes.get("WhisperFrontend")
            frontend = frontend_class(
                n_mels=self.model.dims.n_mels, do_pad_trim=kwargs.get("do_pad_trim", True)
            )
            self.frontend = frontend
        else:
            frontend = frontend if frontend is not None else self.frontend

        meta_data = {}
        if (
@@ -423,7 +830,7 @@
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs if hasattr(frontend, "fs") else 16000,
                fs=frontend.fs,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=tokenizer,
@@ -435,37 +842,136 @@
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            frame_shift = frontend.frame_shift if hasattr(frontend, "frame_shift") else 10
            lfr_n = frontend.lfr_n if hasattr(frontend, "lfr_n") else 1
            meta_data["batch_data_time"] = speech_lengths.sum().item() * frame_shift * lfr_n / 1000
            meta_data["batch_data_time"] = (
                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            )

        speech = speech.to(device=kwargs["device"])[0, :, :]
        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])

        DecodingOptions = kwargs.get("DecodingOptions", {})
        task = DecodingOptions.get("task", "ASR")
        if isinstance(task, str):
            task = [task]
        task = "".join([f"<|{x}|>" for x in task])
        initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")
        DecodingOptions["initial_prompt"] = initial_prompt
        language = kwargs.get("language", "auto")
        language_query = self.embed(
            torch.LongTensor([[self.lid_dict[language] if language in self.lid_dict else 0]]).to(
                speech.device
            )
        ).repeat(speech.size(0), 1, 1)

        language = DecodingOptions.get("language", None)
        language = None if language == "auto" else language
        DecodingOptions["language"] = language
        use_itn = kwargs.get("use_itn", False)
        textnorm = kwargs.get("text_norm", None)
        output_timestamp = kwargs.get("output_timestamp", False)

        DecodingOptions["vocab_path"] = kwargs["tokenizer_conf"].get("vocab_path", None)
        if textnorm is None:
            textnorm = "withitn" if use_itn else "woitn"
        textnorm_query = self.embed(
            torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)
        ).repeat(speech.size(0), 1, 1)
        speech = torch.cat((textnorm_query, speech), dim=1)
        speech_lengths += 1

        if "without_timestamps" not in DecodingOptions:
            DecodingOptions["without_timestamps"] = True
        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
            speech.size(0), 1, 1
        )
        input_query = torch.cat((language_query, event_emo_query), dim=1)
        speech = torch.cat((input_query, speech), dim=1)
        speech_lengths += 3

        options = whisper.DecodingOptions(**DecodingOptions)
        # Encoder
        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)
        if isinstance(encoder_out, tuple):
            encoder_out = encoder_out[0]

        result = whisper.decode(self.model, speech, options)
        text = f"{result.text}"
        # c. Passed the encoder result and the beam search
        ctc_logits = self.ctc.log_softmax(encoder_out)
        if kwargs.get("ban_emo_unk", False):
            ctc_logits[:, :, self.emo_dict["unk"]] = -float("inf")

        results = []
        result_i = {"key": key[0], "text": text}
        b, n, d = encoder_out.size()
        if isinstance(key[0], (list, tuple)):
            key = key[0]
        if len(key) < b:
            key = key * b
        for i in range(b):
            x = ctc_logits[i, : encoder_out_lens[i].item(), :]
            yseq = x.argmax(dim=-1)
            yseq = torch.unique_consecutive(yseq, dim=-1)

        results.append(result_i)
            ibest_writer = None
            if kwargs.get("output_dir") is not None:
                if not hasattr(self, "writer"):
                    self.writer = DatadirWriter(kwargs.get("output_dir"))
                ibest_writer = self.writer[f"1best_recog"]

            mask = yseq != self.blank_id
            token_int = yseq[mask].tolist()

            # Change integer-ids to tokens
            text = tokenizer.decode(token_int)

            # result_i = {"key": key[i], "text": text}
            # results.append(result_i)

            if ibest_writer is not None:
                ibest_writer["text"][key[i]] = text

            if output_timestamp:
                from itertools import groupby
                timestamp = []
                tokens = tokenizer.text2tokens(text)[4:]
                logits_speech = self.ctc.softmax(encoder_out)[i, 4:encoder_out_lens[i].item(), :]
                pred = logits_speech.argmax(-1).cpu()
                logits_speech[pred==self.blank_id, self.blank_id] = 0
                align = ctc_forced_align(
                    logits_speech.unsqueeze(0).float(),
                    torch.Tensor(token_int[4:]).unsqueeze(0).long().to(logits_speech.device),
                    (encoder_out_lens-4).long(),
                    torch.tensor(len(token_int)-4).unsqueeze(0).long().to(logits_speech.device),
                    ignore_id=self.ignore_id,
                )
                pred = groupby(align[0, :encoder_out_lens[0]])
                _start = 0
                token_id = 0
                ts_max = encoder_out_lens[i] - 4
                for pred_token, pred_frame in pred:
                    _end = _start + len(list(pred_frame))
                    if pred_token != 0:
                        ts_left = max((_start*60-30)/1000, 0)
                        ts_right = min((_end*60-30)/1000, (ts_max*60-30)/1000)
                        timestamp.append([tokens[token_id], ts_left, ts_right])
                        token_id += 1
                    _start = _end
                timestamp = self.post(timestamp)
                result_i = {"key": key[i], "text": text, "timestamp": timestamp}
                results.append(result_i)
            else:
                result_i = {"key": key[i], "text": text}
                results.append(result_i)
        return results, meta_data

    def post(self, timestamp):
        timestamp_new = []
        for i, t in enumerate(timestamp):
            word, start, end = t
            if word == '▁':
                continue
            if i == 0:
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start*1000), int(end*1000)])
            elif word.startswith("▁") or len(word) == 1 or not word[1].isalpha():
                word = word[1:]
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start*1000), int(end*1000)])
            else:
                # timestamp_new[-1][0] += word
                timestamp_new[-1][1] = int(end*1000)
        return timestamp_new
    def export(self, **kwargs):
        from export_meta import export_rebuild_model

        if "max_seq_len" not in kwargs:
            kwargs["max_seq_len"] = 512
        models = export_rebuild_model(model=self, **kwargs)
        return models

        return results, meta_data