add paraformer online opt infer code

funasr/bin/asr_inference_launch.py

@@ -840,38 +840,73 @@ def inference_paraformer_online(
             data = yaml.load(f, Loader=yaml.Loader)
         return data
 
-    def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+    def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], encoder_chunk_look_back=0,
+                       decoder_chunk_look_back=0, batch_size=1):
         if len(cache) > 0:
             return cache
         config = _read_yaml(asr_train_config)
         enc_output_size = config["encoder_conf"]["output_size"]
         feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
         cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
-                    "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+                    "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size,
+                    "encoder_chunk_look_back": encoder_chunk_look_back, "last_chunk": False, "opt": None,
                     "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
         cache["encoder"] = cache_en
 
-        cache_de = {"decode_fsmn": None}
+        cache_de = {"decode_fsmn": None, "decoder_chunk_look_back": decoder_chunk_look_back, "opt": None}
         cache["decoder"] = cache_de
 
         return cache
 
-    def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+    def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], encoder_chunk_look_back=0,
+                     decoder_chunk_look_back=0, batch_size=1):
         if len(cache) > 0:
             config = _read_yaml(asr_train_config)
             enc_output_size = config["encoder_conf"]["output_size"]
             feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
             cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
-                        "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
-                        "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)),
-                        "tail_chunk": False}
+                        "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size,
+                        "encoder_chunk_look_back": encoder_chunk_look_back, "last_chunk": False, "opt": None,
+                        "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
             cache["encoder"] = cache_en
 
-            cache_de = {"decode_fsmn": None}
+            cache_de = {"decode_fsmn": None, "decoder_chunk_look_back": decoder_chunk_look_back, "opt": None}
             cache["decoder"] = cache_de
 
         return cache
 
+    #def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+    #    if len(cache) > 0:
+    #        return cache
+    #    config = _read_yaml(asr_train_config)
+    #    enc_output_size = config["encoder_conf"]["output_size"]
+    #    feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
+    #    cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
+    #                "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+    #                "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
+    #    cache["encoder"] = cache_en
+
+    #    cache_de = {"decode_fsmn": None}
+    #    cache["decoder"] = cache_de
+
+    #    return cache
+
+    #def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+    #    if len(cache) > 0:
+    #        config = _read_yaml(asr_train_config)
+    #        enc_output_size = config["encoder_conf"]["output_size"]
+    #        feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
+    #        cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
+    #                    "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+    #                    "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)),
+    #                    "tail_chunk": False}
+    #        cache["encoder"] = cache_en
+
+    #        cache_de = {"decode_fsmn": None}
+    #        cache["decoder"] = cache_de
+
+    #    return cache
+
     def _forward(
             data_path_and_name_and_type,
             raw_inputs: Union[np.ndarray, torch.Tensor] = None,
@@ -899,12 +934,20 @@ def inference_paraformer_online(
         is_final = False
         cache = {}
         chunk_size = [5, 10, 5]
+        encoder_chunk_look_back = 0
+        decoder_chunk_look_back = 0
         if param_dict is not None and "cache" in param_dict:
             cache = param_dict["cache"]
         if param_dict is not None and "is_final" in param_dict:
             is_final = param_dict["is_final"]
         if param_dict is not None and "chunk_size" in param_dict:
             chunk_size = param_dict["chunk_size"]
+        if param_dict is not None and "encoder_chunk_look_back" in param_dict:
+            encoder_chunk_look_back = param_dict["encoder_chunk_look_back"]
+            if encoder_chunk_look_back > 0:
+                chunk_size[0] = 0
+        if param_dict is not None and "decoder_chunk_look_back" in param_dict:
+            decoder_chunk_look_back = param_dict["decoder_chunk_look_back"]
 
         # 7 .Start for-loop
         # FIXME(kamo): The output format should be discussed about
@@ -916,7 +959,8 @@ def inference_paraformer_online(
             sample_offset = 0
             speech_length = raw_inputs.shape[1]
             stride_size = chunk_size[1] * 960
-            cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1)
+            cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1, 
+                                   encoder_chunk_look_back=encoder_chunk_look_back, decoder_chunk_look_back=decoder_chunk_look_back)
             final_result = ""
             for sample_offset in range(0, speech_length, min(stride_size, speech_length - sample_offset)):
                 if sample_offset + stride_size >= speech_length - 1:
@@ -937,7 +981,8 @@ def inference_paraformer_online(
 
         asr_result_list.append(item)
         if is_final:
-            cache = _cache_reset(cache, chunk_size=chunk_size, batch_size=1)
+            cache = _cache_reset(cache, chunk_size=chunk_size, batch_size=1,
+                                 encoder_chunk_look_back=encoder_chunk_look_back, decoder_chunk_look_back=decoder_chunk_look_back)
         return asr_result_list
 
     return _forward

funasr/models/encoder/sanm_encoder.py

@@ -114,9 +114,45 @@ class EncoderLayerSANM(nn.Module):
         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
+
+
 class SANMEncoder(AbsEncoder):
     """
     Author: Speech Lab of DAMO Academy, Alibaba Group
@@ -837,11 +873,56 @@ class SANMEncoderChunkOpt(AbsEncoder):
         cache["feats"] = overlap_feats[:, -(cache["chunk_size"][0] + cache["chunk_size"][2]):, :]
         return overlap_feats
 
+    #def forward_chunk(self,
+    #                  xs_pad: torch.Tensor,
+    #                  ilens: torch.Tensor,
+    #                  cache: dict = None,
+    #                  ctc: CTC = None,
+    #                  ):
+    #    xs_pad *= self.output_size() ** 0.5
+    #    if self.embed is None:
+    #        xs_pad = xs_pad
+    #    else:
+    #        xs_pad = self.embed(xs_pad, cache)
+    #    if cache["tail_chunk"]:
+    #        xs_pad = to_device(cache["feats"], device=xs_pad.device)
+    #    else:
+    #        xs_pad = self._add_overlap_chunk(xs_pad, cache)
+    #    encoder_outs = self.encoders0(xs_pad, None, None, None, None)
+    #    xs_pad, masks = encoder_outs[0], encoder_outs[1]
+    #    intermediate_outs = []
+    #    if len(self.interctc_layer_idx) == 0:
+    #        encoder_outs = self.encoders(xs_pad, None, None, None, None)
+    #        xs_pad, masks = encoder_outs[0], encoder_outs[1]
+    #    else:
+    #        for layer_idx, encoder_layer in enumerate(self.encoders):
+    #            encoder_outs = encoder_layer(xs_pad, None, None, None, None)
+    #            xs_pad, masks = encoder_outs[0], encoder_outs[1]
+    #            if layer_idx + 1 in self.interctc_layer_idx:
+    #                encoder_out = xs_pad
+
+    #                # intermediate outputs are also normalized
+    #                if self.normalize_before:
+    #                    encoder_out = self.after_norm(encoder_out)
+
+    #                intermediate_outs.append((layer_idx + 1, encoder_out))
+
+    #                if self.interctc_use_conditioning:
+    #                    ctc_out = ctc.softmax(encoder_out)
+    #                    xs_pad = xs_pad + self.conditioning_layer(ctc_out)
+
+    #    if self.normalize_before:
+    #        xs_pad = self.after_norm(xs_pad)
+
+    #    if len(intermediate_outs) > 0:
+    #        return (xs_pad, intermediate_outs), None, None
+    #    return xs_pad, ilens, None
+
+
     def forward_chunk(self,
                       xs_pad: torch.Tensor,
                       ilens: torch.Tensor,
                       cache: dict = None,
-                      ctc: CTC = None,
                       ):
         xs_pad *= self.output_size() ** 0.5
         if self.embed is None:
@@ -852,34 +933,25 @@ class SANMEncoderChunkOpt(AbsEncoder):
             xs_pad = to_device(cache["feats"], device=xs_pad.device)
         else:
             xs_pad = self._add_overlap_chunk(xs_pad, cache)
-        encoder_outs = self.encoders0(xs_pad, None, None, None, None)
-        xs_pad, masks = encoder_outs[0], encoder_outs[1]
-        intermediate_outs = []
-        if len(self.interctc_layer_idx) == 0:
-            encoder_outs = self.encoders(xs_pad, None, None, None, None)
-            xs_pad, masks = encoder_outs[0], encoder_outs[1]
+        if cache["opt"] is None:
+            cache_layer_num = len(self.encoders0) + len(self.encoders)
+            new_cache = [None] * cache_layer_num
         else:
-            for layer_idx, encoder_layer in enumerate(self.encoders):
-                encoder_outs = encoder_layer(xs_pad, None, None, None, None)
-                xs_pad, masks = encoder_outs[0], encoder_outs[1]
-                if layer_idx + 1 in self.interctc_layer_idx:
-                    encoder_out = xs_pad
-
-                    # intermediate outputs are also normalized
-                    if self.normalize_before:
-                        encoder_out = self.after_norm(encoder_out)
+            new_cache = cache["opt"]
 
-                    intermediate_outs.append((layer_idx + 1, encoder_out))
+        for layer_idx, encoder_layer in enumerate(self.encoders0):
+            encoder_outs = encoder_layer.forward_chunk(xs_pad, new_cache[layer_idx], cache["chunk_size"], cache["encoder_chunk_look_back"])
+            xs_pad, new_cache[0] = encoder_outs[0], encoder_outs[1]
 
-                    if self.interctc_use_conditioning:
-                        ctc_out = ctc.softmax(encoder_out)
-                        xs_pad = xs_pad + self.conditioning_layer(ctc_out)
+        for layer_idx, encoder_layer in enumerate(self.encoders):
+            encoder_outs = encoder_layer.forward_chunk(xs_pad, new_cache[layer_idx+len(self.encoders0)], cache["chunk_size"], cache["encoder_chunk_look_back"])
+            xs_pad, new_cache[layer_idx+1] = encoder_outs[0], encoder_outs[1]
 
         if self.normalize_before:
             xs_pad = self.after_norm(xs_pad)
+        if cache["encoder_chunk_look_back"] > 0:
+            cache["opt"] = new_cache
 
-        if len(intermediate_outs) > 0:
-            return (xs_pad, intermediate_outs), None, None
         return xs_pad, ilens, None
 
     def gen_tf2torch_map_dict(self):

funasr/modules/attention.py

@@ -456,6 +456,38 @@ class MultiHeadedAttentionSANM(nn.Module):
         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:
+            if cache is not None:
+                k_h = torch.cat((cache["k"], k_h), dim=2)
+                v_h = torch.cat((cache["v"], v_h), dim=2)
+                cache["k"] = k_h[:, :, -(look_back * chunk_size[1]):, :]
+                cache["v"] = v_h[:, :, -(look_back * chunk_size[1]):, :]
+            else:
+                cache_tmp = {"k": k_h[:, :, -(look_back * chunk_size[1]):, :],
+                             "v": v_h[:, :, -(look_back * chunk_size[1]):, :]}
+                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 MultiHeadedAttentionSANMwithMask(MultiHeadedAttentionSANM):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)