提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 funasr/models/paraformer/cif_predictor.py

@@ -10,11 +10,22 @@
from funasr.register import tables

from funasr.train_utils.device_funcs import to_device

from funasr.models.transformer.utils.nets_utils import make_pad_mask

from torch.cuda.amp import autocast





@tables.register("predictor_classes", "CifPredictor")

class CifPredictor(torch.nn.Module):

    def __init__(self, idim, l_order, r_order, threshold=1.0, dropout=0.1, smooth_factor=1.0, noise_threshold=0, tail_threshold=0.45):

    def __init__(

        self,

        idim,

        l_order,

        r_order,

        threshold=1.0,

        dropout=0.1,

        smooth_factor=1.0,

        noise_threshold=0,

        tail_threshold=0.45,

    ):

        super().__init__()



        self.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)

@@ -26,44 +37,55 @@
        self.noise_threshold = noise_threshold

        self.tail_threshold = tail_threshold



    def forward(self, hidden, target_label=None, mask=None, ignore_id=-1, mask_chunk_predictor=None,

                target_label_length=None):

        h = hidden

        context = h.transpose(1, 2)

        queries = self.pad(context)

        memory = self.cif_conv1d(queries)

        output = memory + context

        output = self.dropout(output)

        output = output.transpose(1, 2)

        output = torch.relu(output)

        output = self.cif_output(output)

        alphas = torch.sigmoid(output)

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

        if mask is not None:

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

            alphas = alphas * mask

        if mask_chunk_predictor is not None:

            alphas = alphas * mask_chunk_predictor

        alphas = alphas.squeeze(-1)

        mask = mask.squeeze(-1)

        if target_label_length is not None:

            target_length = target_label_length

        elif target_label is not None:

            target_length = (target_label != ignore_id).float().sum(-1)

        else:

            target_length = None

        token_num = alphas.sum(-1)

        if target_length is not None:

            alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))

        elif self.tail_threshold > 0.0:

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

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

        
        if target_length is None and self.tail_threshold > 0.0:

            token_num_int = torch.max(token_num).type(torch.int32).item()

            acoustic_embeds = acoustic_embeds[:, :token_num_int, :]

            
    def forward(

        self,

        hidden,

        target_label=None,

        mask=None,

        ignore_id=-1,

        mask_chunk_predictor=None,

        target_label_length=None,

    ):



        with autocast(False):

            h = hidden

            context = h.transpose(1, 2)

            queries = self.pad(context)

            memory = self.cif_conv1d(queries)

            output = memory + context

            output = self.dropout(output)

            output = output.transpose(1, 2)

            output = torch.relu(output)

            output = self.cif_output(output)

            alphas = torch.sigmoid(output)

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

            if mask is not None:

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

                alphas = alphas * mask

            if mask_chunk_predictor is not None:

                alphas = alphas * mask_chunk_predictor

            alphas = alphas.squeeze(-1)

            mask = mask.squeeze(-1)

            if target_label_length is not None:

                target_length = target_label_length

            elif target_label is not None:

                target_length = (target_label != ignore_id).float().sum(-1)

            else:

                target_length = None

            token_num = alphas.sum(-1)

            if target_length is not None:

                alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))

            elif self.tail_threshold > 0.0:

                hidden, alphas, token_num = self.tail_process_fn(

                    hidden, alphas, token_num, mask=mask

                )



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



            if target_length is None and self.tail_threshold > 0.0:

                token_num_int = torch.max(token_num).type(torch.int32).item()

                acoustic_embeds = acoustic_embeds[:, :token_num_int, :]



        return acoustic_embeds, token_num, alphas, cif_peak



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

@@ -89,10 +111,9 @@


        return hidden, alphas, token_num_floor





    def gen_frame_alignments(self,

                             alphas: torch.Tensor = None,

                             encoder_sequence_length: torch.Tensor = None):

    def gen_frame_alignments(

        self, alphas: torch.Tensor = None, encoder_sequence_length: torch.Tensor = None

    ):

        batch_size, maximum_length = alphas.size()

        int_type = torch.int32



@@ -115,11 +136,15 @@
        index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)

        index_div_bool_zeros = index_div.eq(0)

        index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1

        index_div_bool_zeros_count = torch.clamp(index_div_bool_zeros_count, 0, encoder_sequence_length.max())

        index_div_bool_zeros_count = torch.clamp(

            index_div_bool_zeros_count, 0, encoder_sequence_length.max()

        )

        token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)

        index_div_bool_zeros_count *= token_num_mask



        index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(1, 1, maximum_length)

        index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(

            1, 1, maximum_length

        )

        ones = torch.ones_like(index_div_bool_zeros_count_tile)

        zeros = torch.zeros_like(index_div_bool_zeros_count_tile)

        ones = torch.cumsum(ones, dim=2)

@@ -130,30 +155,37 @@
        index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)

        index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)

        index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)

        predictor_mask = (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max())).type(

            int_type).to(encoder_sequence_length.device)

        predictor_mask = (

            (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max()))

            .type(int_type)

            .to(encoder_sequence_length.device)

        )

        index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask



        predictor_alignments = index_div_bool_zeros_count_tile_out

        predictor_alignments_length = predictor_alignments.sum(-1).type(encoder_sequence_length.dtype)

        predictor_alignments_length = predictor_alignments.sum(-1).type(

            encoder_sequence_length.dtype

        )

        return predictor_alignments.detach(), predictor_alignments_length.detach()





@tables.register("predictor_classes", "CifPredictorV2")

class CifPredictorV2(torch.nn.Module):

    def __init__(self,

                 idim,

                 l_order,

                 r_order,

                 threshold=1.0,

                 dropout=0.1,

                 smooth_factor=1.0,

                 noise_threshold=0,

                 tail_threshold=0.0,

                 tf2torch_tensor_name_prefix_torch="predictor",

                 tf2torch_tensor_name_prefix_tf="seq2seq/cif",

                 tail_mask=True,

                 ):

        super(CifPredictorV2, self).__init__()

    def __init__(

        self,

        idim,

        l_order,

        r_order,

        threshold=1.0,

        dropout=0.1,

        smooth_factor=1.0,

        noise_threshold=0,

        tail_threshold=0.0,

        tf2torch_tensor_name_prefix_torch="predictor",

        tf2torch_tensor_name_prefix_tf="seq2seq/cif",

        tail_mask=True,

    ):

        super().__init__()



        self.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)

        self.cif_conv1d = torch.nn.Conv1d(idim, idim, l_order + r_order + 1)

@@ -167,43 +199,56 @@
        self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf

        self.tail_mask = tail_mask



    def forward(self, hidden, target_label=None, mask=None, ignore_id=-1, mask_chunk_predictor=None,

                target_label_length=None):

        h = hidden

        context = h.transpose(1, 2)

        queries = self.pad(context)

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

        output = output.transpose(1, 2)

    def forward(

        self,

        hidden,

        target_label=None,

        mask=None,

        ignore_id=-1,

        mask_chunk_predictor=None,

        target_label_length=None,

    ):



        output = self.cif_output(output)

        alphas = torch.sigmoid(output)

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

        if mask is not None:

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

            alphas = alphas * mask

        if mask_chunk_predictor is not None:

            alphas = alphas * mask_chunk_predictor

        alphas = alphas.squeeze(-1)

        mask = mask.squeeze(-1)

        if target_label_length is not None:

            target_length = target_label_length.squeeze(-1)

        elif target_label is not None:

            target_length = (target_label != ignore_id).float().sum(-1)

        else:

            target_length = None

        token_num = alphas.sum(-1)

        if target_length is not None:

            alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))

        elif self.tail_threshold > 0.0:

            if self.tail_mask:

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

        with autocast(False):

            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)

            if mask is not None:

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

                alphas = alphas * mask

            if mask_chunk_predictor is not None:

                alphas = alphas * mask_chunk_predictor

            alphas = alphas.squeeze(-1)

            mask = mask.squeeze(-1)

            if target_label_length is not None:

                target_length = target_label_length.squeeze(-1)

            elif target_label is not None:

                target_length = (target_label != ignore_id).float().sum(-1)

            else:

                hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, token_num, mask=None)

                target_length = None

            token_num = alphas.sum(-1)

            if target_length is not None:

                alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))

            elif self.tail_threshold > 0.0:

                if self.tail_mask:

                    hidden, alphas, token_num = self.tail_process_fn(

                        hidden, alphas, token_num, mask=mask

                    )

                else:

                    hidden, alphas, token_num = self.tail_process_fn(

                        hidden, alphas, token_num, mask=None

                    )



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

        if target_length is None and self.tail_threshold > 0.0:

            token_num_int = torch.max(token_num).type(torch.int32).item()

            acoustic_embeds = acoustic_embeds[:, :token_num_int, :]

            acoustic_embeds, cif_peak = cif_v1(hidden, alphas, self.threshold)

            if target_length is None and self.tail_threshold > 0.0:

                token_num_int = torch.max(token_num).type(torch.int32).item()

                acoustic_embeds = acoustic_embeds[:, :token_num_int, :]



        return acoustic_embeds, token_num, alphas, cif_peak



@@ -228,9 +273,9 @@
        cache_hiddens = []



        if cache is not None and "chunk_size" in cache:

            alphas[:, :cache["chunk_size"][0]] = 0.0

            alphas[:, : cache["chunk_size"][0]] = 0.0

            if not is_final:

                alphas[:, sum(cache["chunk_size"][:2]):] = 0.0

                alphas[:, sum(cache["chunk_size"][:2]) :] = 0.0

        if cache is not None and "cif_alphas" in cache and "cif_hidden" in cache:

            cache["cif_hidden"] = to_device(cache["cif_hidden"], device=hidden.device)

            cache["cif_alphas"] = to_device(cache["cif_alphas"], device=alphas.device)

@@ -280,10 +325,12 @@


        max_token_len = max(token_length)

        if max_token_len == 0:

             return hidden, torch.stack(token_length, 0), None, None

            return hidden, torch.stack(token_length, 0), None, None

        list_ls = []

        for b in range(batch_size):

            pad_frames = torch.zeros((max_token_len - token_length[b], hidden_size), device=alphas.device)

            pad_frames = torch.zeros(

                (max_token_len - token_length[b], hidden_size), device=alphas.device

            )

            if token_length[b] == 0:

                list_ls.append(pad_frames)

            else:

@@ -295,7 +342,6 @@
        cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)

        cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)

        return torch.stack(list_ls, 0), torch.stack(token_length, 0), None, None





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

        b, t, d = hidden.size()

@@ -323,9 +369,9 @@


        return hidden, alphas, token_num_floor



    def gen_frame_alignments(self,

                             alphas: torch.Tensor = None,

                             encoder_sequence_length: torch.Tensor = None):

    def gen_frame_alignments(

        self, alphas: torch.Tensor = None, encoder_sequence_length: torch.Tensor = None

    ):

        batch_size, maximum_length = alphas.size()

        int_type = torch.int32



@@ -348,11 +394,15 @@
        index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)

        index_div_bool_zeros = index_div.eq(0)

        index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1

        index_div_bool_zeros_count = torch.clamp(index_div_bool_zeros_count, 0, encoder_sequence_length.max())

        index_div_bool_zeros_count = torch.clamp(

            index_div_bool_zeros_count, 0, encoder_sequence_length.max()

        )

        token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)

        index_div_bool_zeros_count *= token_num_mask



        index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(1, 1, maximum_length)

        index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(

            1, 1, maximum_length

        )

        ones = torch.ones_like(index_div_bool_zeros_count_tile)

        zeros = torch.zeros_like(index_div_bool_zeros_count_tile)

        ones = torch.cumsum(ones, dim=2)

@@ -363,69 +413,197 @@
        index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)

        index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)

        index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)

        predictor_mask = (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max())).type(

            int_type).to(encoder_sequence_length.device)

        predictor_mask = (

            (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max()))

            .type(int_type)

            .to(encoder_sequence_length.device)

        )

        index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask



        predictor_alignments = index_div_bool_zeros_count_tile_out

        predictor_alignments_length = predictor_alignments.sum(-1).type(encoder_sequence_length.dtype)

        predictor_alignments_length = predictor_alignments.sum(-1).type(

            encoder_sequence_length.dtype

        )

        return predictor_alignments.detach(), predictor_alignments_length.detach()



    def gen_tf2torch_map_dict(self):

    
        tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch

        tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf

        map_dict_local = {

            ## predictor

            "{}.cif_conv1d.weight".format(tensor_name_prefix_torch):

                {"name": "{}/conv1d/kernel".format(tensor_name_prefix_tf),

                 "squeeze": None,

                 "transpose": (2, 1, 0),

                 },  # (256,256,3),(3,256,256)

            "{}.cif_conv1d.bias".format(tensor_name_prefix_torch):

                {"name": "{}/conv1d/bias".format(tensor_name_prefix_tf),

                 "squeeze": None,

                 "transpose": None,

                 },  # (256,),(256,)

            "{}.cif_output.weight".format(tensor_name_prefix_torch):

                {"name": "{}/conv1d_1/kernel".format(tensor_name_prefix_tf),

                 "squeeze": 0,

                 "transpose": (1, 0),

                 },  # (1,256),(1,256,1)

            "{}.cif_output.bias".format(tensor_name_prefix_torch):

                {"name": "{}/conv1d_1/bias".format(tensor_name_prefix_tf),

                 "squeeze": None,

                 "transpose": None,

                 },  # (1,),(1,)

        }

        return map_dict_local



    def convert_tf2torch(self,

                         var_dict_tf,

                         var_dict_torch,

                         ):

        map_dict = self.gen_tf2torch_map_dict()

        var_dict_torch_update = dict()

        for name in sorted(var_dict_torch.keys(), reverse=False):

            names = name.split('.')

            if names[0] == self.tf2torch_tensor_name_prefix_torch:

                name_tf = map_dict[name]["name"]

                data_tf = var_dict_tf[name_tf]

                if map_dict[name]["squeeze"] is not None:

                    data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])

                if map_dict[name]["transpose"] is not None:

                    data_tf = np.transpose(data_tf, map_dict[name]["transpose"])

                data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")

                assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,

                                                                                                var_dict_torch[

                                                                                                    name].size(),

                                                                                                data_tf.size())

                var_dict_torch_update[name] = data_tf

                logging.info(

                    "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_tf,

                                                                                  var_dict_tf[name_tf].shape))

    
        return var_dict_torch_update

@tables.register("predictor_classes", "CifPredictorV2Export")

class CifPredictorV2Export(torch.nn.Module):

    def __init__(self, model, **kwargs):

        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



    def forward(

        self,

        hidden: torch.Tensor,

        mask: torch.Tensor,

    ):

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

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

        mask = mask.squeeze(-1)

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

        acoustic_embeds, cif_peak = cif_v1_export(hidden, alphas, self.threshold)



        return acoustic_embeds, token_num, alphas, cif_peak



    def forward_cnn(

        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)



        return alphas, token_num



    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_v1_export(hidden, alphas, threshold: float):

    device = hidden.device

    dtype = hidden.dtype

    batch_size, len_time, hidden_size = hidden.size()

    threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)



    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)



    # prefix_sum = torch.cumsum(alphas, dim=1)

    prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(

        torch.float32

    )  # cumsum precision degradation cause wrong result in extreme

    prefix_sum_floor = torch.floor(prefix_sum)

    dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)

    dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)



    dislocation_prefix_sum_floor[:, 0] = 0

    dislocation_diff = prefix_sum_floor - dislocation_prefix_sum_floor



    fire_idxs = dislocation_diff > 0

    fires[fire_idxs] = 1

    fires = fires + prefix_sum - prefix_sum_floor



    # prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).repeat((1, 1, hidden_size)) * hidden, dim=1)

    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)



    batch_len = fire_idxs.sum(1)

    batch_idxs = torch.cumsum(batch_len, dim=0)

    shift_batch_idxs = torch.roll(batch_idxs, 1, dims=0)

    shift_batch_idxs[0] = 0

    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    # remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]

    remain_frames = remains[fire_idxs].unsqueeze(-1).repeat((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0



    frames = frames - shift_frames + shift_remain_frames - remain_frames



    # max_label_len = batch_len.max()

    max_label_len = alphas.sum(dim=-1)

    max_label_len = torch.floor(max_label_len).max().to(dtype=torch.int64)



    # frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames

    return frame_fires, fires





@torch.jit.script

def cif_export(hidden, alphas, threshold: float):

    batch_size, len_time, hidden_size = hidden.size()

    threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)



    # loop varss

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

    frame = torch.zeros([batch_size, hidden_size], dtype=hidden.dtype, device=hidden.device)

    # intermediate vars along time

    list_fires = []

    list_frames = []



    for t in range(len_time):

        alpha = alphas[:, t]

        distribution_completion = (

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

        )



        integrate += alpha

        list_fires.append(integrate)



        fire_place = integrate >= threshold

        integrate = torch.where(

            fire_place,

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

            integrate,

        )

        cur = torch.where(fire_place, distribution_completion, alpha)

        remainds = alpha - cur



        frame += cur[:, None] * hidden[:, t, :]

        list_frames.append(frame)

        frame = torch.where(

            fire_place[:, None].repeat(1, hidden_size), remainds[:, None] * hidden[:, t, :], frame

        )



    fires = torch.stack(list_fires, 1)

    frames = torch.stack(list_frames, 1)



    fire_idxs = fires >= threshold

    frame_fires = torch.zeros_like(hidden)

    max_label_len = frames[0, fire_idxs[0]].size(0)

    for b in range(batch_size):

        frame_fire = frames[b, fire_idxs[b]]

        frame_len = frame_fire.size(0)

        frame_fires[b, :frame_len, :] = frame_fire



        if frame_len >= max_label_len:

            max_label_len = frame_len

    frame_fires = frame_fires[:, :max_label_len, :]

    return frame_fires, fires





class mae_loss(torch.nn.Module):

@@ -433,7 +611,7 @@
    def __init__(self, normalize_length=False):

        super(mae_loss, self).__init__()

        self.normalize_length = normalize_length

        self.criterion = torch.nn.L1Loss(reduction='sum')

        self.criterion = torch.nn.L1Loss(reduction="sum")



    def forward(self, token_length, pre_token_length):

        loss_token_normalizer = token_length.size(0)

@@ -462,19 +640,17 @@
        list_fires.append(integrate)



        fire_place = integrate >= threshold

        integrate = torch.where(fire_place,

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

                                integrate)

        cur = torch.where(fire_place,

                          distribution_completion,

                          alpha)

        integrate = torch.where(

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

        )

        cur = torch.where(fire_place, distribution_completion, alpha)

        remainds = alpha - cur



        frame += cur[:, None] * hidden[:, t, :]

        list_frames.append(frame)

        frame = torch.where(fire_place[:, None].repeat(1, hidden_size),

                            remainds[:, None] * hidden[:, t, :],

                            frame)

        frame = torch.where(

            fire_place[:, None].repeat(1, hidden_size), remainds[:, None] * hidden[:, t, :], frame

        )



    fires = torch.stack(list_fires, 1)

    frames = torch.stack(list_frames, 1)

@@ -487,6 +663,76 @@
        pad_l = torch.zeros([max_label_len - l.size(0), hidden_size], device=hidden.device)

        list_ls.append(torch.cat([l, pad_l], 0))

    return torch.stack(list_ls, 0), fires





def cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=False):

    batch_size, len_time = alphas.size()

    device = alphas.device

    dtype = alphas.dtype



    threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)



    fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)



    # prefix_sum = torch.cumsum(alphas, dim=1)

    prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(

        torch.float32

    )  # cumsum precision degradation cause wrong result in extreme

    prefix_sum_floor = torch.floor(prefix_sum)

    dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)

    dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)



    dislocation_prefix_sum_floor[:, 0] = 0

    dislocation_diff = prefix_sum_floor - dislocation_prefix_sum_floor



    fire_idxs = dislocation_diff > 0

    fires[fire_idxs] = 1

    fires = fires + prefix_sum - prefix_sum_floor

    if return_fire_idxs:

        return fires, fire_idxs

    return fires





def cif_v1(hidden, alphas, threshold):

    fires, fire_idxs = cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=True)



    device = hidden.device

    dtype = hidden.dtype

    batch_size, len_time, hidden_size = hidden.size()

    # frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    # prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)

    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).repeat((1, 1, hidden_size)) * hidden, dim=1)



    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)



    batch_len = fire_idxs.sum(1)

    batch_idxs = torch.cumsum(batch_len, dim=0)

    shift_batch_idxs = torch.roll(batch_idxs, 1, dims=0)

    shift_batch_idxs[0] = 0

    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    # remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]

    remain_frames = remains[fire_idxs].unsqueeze(-1).repeat((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0



    frames = frames - shift_frames + shift_remain_frames - remain_frames



    # max_label_len = batch_len.max()

    max_label_len = (

        torch.round(alphas.sum(-1)).int().max()

    )  # torch.round to calculate the max length



    # frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames

    return frame_fires, fires





def cif_wo_hidden(alphas, threshold):

@@ -504,10 +750,11 @@
        list_fires.append(integrate)



        fire_place = integrate >= threshold

        integrate = torch.where(fire_place,

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

                                integrate)

        integrate = torch.where(

            fire_place,

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

            integrate,

        )



    fires = torch.stack(list_fires, 1)

    return fires