Update FsmnVad.cc

zhuzizyf

2023-04-22 4f8bce944e273e317cb84c7046ea514b9d958b4b

 funasr/models/predictor/cif.py

@@ -200,6 +200,7 @@
        return acoustic_embeds, token_num, alphas, cif_peak



    def forward_chunk(self, hidden, cache=None):

        b, t, d = hidden.size()

        h = hidden

        context = h.transpose(1, 2)

        queries = self.pad(context)

@@ -220,6 +221,8 @@
            alphas = alphas * mask_chunk_predictor

      

        if cache is not None:

            if cache["is_final"]:

                alphas[:, cache["stride"] + cache["pad_left"] - 1] += 0.45

            if cache["cif_hidden"] is not None:

                hidden = torch.cat((cache["cif_hidden"], hidden), 1)

            if cache["cif_alphas"] is not None:

@@ -231,6 +234,7 @@
        last_fire_place = len_time - 1

        last_fire_remainds = 0.0

        pre_alphas_length = 0

        last_fire = False

 

        mask_chunk_peak_predictor = None

        if cache is not None:

@@ -241,7 +245,6 @@
                mask_chunk_peak_predictor[:, :pre_alphas_length] = 1.0

            mask_chunk_peak_predictor[:, pre_alphas_length + cache["pad_left"]:pre_alphas_length + cache["stride"] + cache["pad_left"]] = 1.0

            



        if mask_chunk_peak_predictor is not None:

            cif_peak = cif_peak * mask_chunk_peak_predictor.squeeze(-1)

        

@@ -249,10 +252,15 @@
            if cif_peak[0][len_time - 1 - i] > self.threshold or cif_peak[0][len_time - 1 - i] == self.threshold:

                last_fire_place = len_time - 1 - i

                last_fire_remainds = cif_peak[0][len_time - 1 - i] - self.threshold

                last_fire = True

                break

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

        cache["cif_hidden"] = hidden[:, last_fire_place:, :]

        cache["cif_alphas"] = torch.cat((last_fire_remainds.unsqueeze(0), alphas[:, last_fire_place+1:]), -1)

        if last_fire:

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

           cache["cif_hidden"] = hidden[:, last_fire_place:, :]

           cache["cif_alphas"] = torch.cat((last_fire_remainds.unsqueeze(0), alphas[:, last_fire_place+1:]), -1)

        else:

           cache["cif_hidden"] = hidden

           cache["cif_alphas"] = alphas

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

        return acoustic_embeds[:, 0:token_num_int, :], token_num, alphas, cif_peak