From 8516d3e850671a35c0031b55b1884074453c331e Mon Sep 17 00:00:00 2001
From: hnluo <haoneng.lhn@alibaba-inc.com>
Date: 星期二, 19 九月 2023 19:06:49 +0800
Subject: [PATCH] Merge pull request #970 from alibaba-damo-academy/dev_lhn

---
 funasr/modules/attention.py                        |   67 +++++++++
 funasr/models/encoder/sanm_encoder.py              |  118 +++++++++++++---
 funasr/models/decoder/sanm_decoder.py              |  146 ++++++++++++++++++--
 funasr/bin/asr_inference_launch.py                 |   68 ++++++++-
 funasr/datasets/large_datasets/datapipes/batch.py  |    2 
 funasr/datasets/large_datasets/datapipes/filter.py |    4 
 funasr/datasets/large_datasets/datapipes/map.py    |    2 
 7 files changed, 350 insertions(+), 57 deletions(-)

diff --git a/funasr/bin/asr_inference_launch.py b/funasr/bin/asr_inference_launch.py
index ad657eb..50b9886 100644
--- a/funasr/bin/asr_inference_launch.py
+++ b/funasr/bin/asr_inference_launch.py
@@ -842,37 +842,72 @@
             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, "chunk_size": chunk_size}
         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, "chunk_size": chunk_size}
             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,
@@ -901,24 +936,34 @@
         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
         raw_inputs = torch.unsqueeze(raw_inputs, axis=0)
         asr_result_list = []
-        cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1)
+        cache = _prepare_cache(cache, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back, 
+                               decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
         item = {}
         if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "sound":
             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, encoder_chunk_look_back=encoder_chunk_look_back, 
+                                   decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
             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:
@@ -939,7 +984,8 @@
 
         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, encoder_chunk_look_back=encoder_chunk_look_back, 
+                                 decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
         return asr_result_list
 
     return _forward
diff --git a/funasr/datasets/large_datasets/datapipes/batch.py b/funasr/datasets/large_datasets/datapipes/batch.py
index 8ec43e9..35e5dba 100644
--- a/funasr/datasets/large_datasets/datapipes/batch.py
+++ b/funasr/datasets/large_datasets/datapipes/batch.py
@@ -39,7 +39,7 @@
         self.batch_mode = batch_mode
 
     def set_epoch(self, epoch):
-        self.epoch = epoch
+        self.datapipe.set_epoch(epoch)
 
     def __iter__(self):
         buffer = []
diff --git a/funasr/datasets/large_datasets/datapipes/filter.py b/funasr/datasets/large_datasets/datapipes/filter.py
index e79934d..6fe7153 100644
--- a/funasr/datasets/large_datasets/datapipes/filter.py
+++ b/funasr/datasets/large_datasets/datapipes/filter.py
@@ -13,7 +13,7 @@
         self.fn = fn
 
     def set_epoch(self, epoch):
-        self.epoch = epoch
+        self.datapipe.set_epoch(epoch)
 
     def __iter__(self):
         assert callable(self.fn)
@@ -21,4 +21,4 @@
             if self.fn(data):
                 yield data
             else:
-                continue
\ No newline at end of file
+                continue
diff --git a/funasr/datasets/large_datasets/datapipes/map.py b/funasr/datasets/large_datasets/datapipes/map.py
index 6e0168d..dfcd6a0 100644
--- a/funasr/datasets/large_datasets/datapipes/map.py
+++ b/funasr/datasets/large_datasets/datapipes/map.py
@@ -14,7 +14,7 @@
         self.fn = fn
 
     def set_epoch(self, epoch):
-        self.epoch = epoch
+        self.datapipe.set_epoch(epoch)
 
     def __iter__(self):
         assert callable(self.fn)
diff --git a/funasr/models/decoder/sanm_decoder.py b/funasr/models/decoder/sanm_decoder.py
index c12e098..3e4e554 100644
--- a/funasr/models/decoder/sanm_decoder.py
+++ b/funasr/models/decoder/sanm_decoder.py
@@ -105,7 +105,50 @@
 
         return x, tgt_mask, memory, memory_mask, cache
 
-    def forward_chunk(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
+    #def forward_chunk(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
+    #    """Compute decoded features.
+
+    #    Args:
+    #        tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
+    #        tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
+    #        memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
+    #        memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
+    #        cache (List[torch.Tensor]): List of cached tensors.
+    #            Each tensor shape should be (#batch, maxlen_out - 1, size).
+
+    #    Returns:
+    #        torch.Tensor: Output tensor(#batch, maxlen_out, size).
+    #        torch.Tensor: Mask for output tensor (#batch, maxlen_out).
+    #        torch.Tensor: Encoded memory (#batch, maxlen_in, size).
+    #        torch.Tensor: Encoded memory mask (#batch, maxlen_in).
+
+    #    """
+    #    # tgt = self.dropout(tgt)
+    #    residual = tgt
+    #    if self.normalize_before:
+    #        tgt = self.norm1(tgt)
+    #    tgt = self.feed_forward(tgt)
+
+    #    x = tgt
+    #    if self.self_attn:
+    #        if self.normalize_before:
+    #            tgt = self.norm2(tgt)
+    #        if self.training:
+    #            cache = None
+    #        x, cache = self.self_attn(tgt, tgt_mask, cache=cache)
+    #        x = residual + self.dropout(x)
+
+    #    if self.src_attn is not None:
+    #        residual = x
+    #        if self.normalize_before:
+    #            x = self.norm3(x)
+
+    #        x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
+
+
+    #    return x, tgt_mask, memory, memory_mask, cache
+
+    def forward_chunk(self, tgt, memory, fsmn_cache=None, opt_cache=None, chunk_size=None, look_back=0):
         """Compute decoded features.
 
         Args:
@@ -123,7 +166,6 @@
             torch.Tensor: Encoded memory mask (#batch, maxlen_in).
 
         """
-        # tgt = self.dropout(tgt)
         residual = tgt
         if self.normalize_before:
             tgt = self.norm1(tgt)
@@ -133,9 +175,7 @@
         if self.self_attn:
             if self.normalize_before:
                 tgt = self.norm2(tgt)
-            if self.training:
-                cache = None
-            x, cache = self.self_attn(tgt, tgt_mask, cache=cache)
+            x, fsmn_cache = self.self_attn(tgt, None, fsmn_cache)
             x = residual + self.dropout(x)
 
         if self.src_attn is not None:
@@ -143,10 +183,11 @@
             if self.normalize_before:
                 x = self.norm3(x)
 
-            x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
+            x, opt_cache = self.src_attn.forward_chunk(x, memory, opt_cache, chunk_size, look_back)
+            x = residual + x
 
+        return x, memory, fsmn_cache, opt_cache
 
-        return x, tgt_mask, memory, memory_mask, cache
 
 class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
     """
@@ -992,6 +1033,65 @@
         )
         return logp.squeeze(0), state
 
+    #def forward_chunk(
+    #    self,
+    #    memory: torch.Tensor,
+    #    tgt: torch.Tensor,
+    #    cache: dict = None,
+    #) -> Tuple[torch.Tensor, torch.Tensor]:
+    #    """Forward decoder.
+
+    #    Args:
+    #        hs_pad: encoded memory, float32  (batch, maxlen_in, feat)
+    #        hlens: (batch)
+    #        ys_in_pad:
+    #            input token ids, int64 (batch, maxlen_out)
+    #            if input_layer == "embed"
+    #            input tensor (batch, maxlen_out, #mels) in the other cases
+    #        ys_in_lens: (batch)
+    #    Returns:
+    #        (tuple): tuple containing:
+
+    #        x: decoded token score before softmax (batch, maxlen_out, token)
+    #            if use_output_layer is True,
+    #        olens: (batch, )
+    #    """
+    #    x = tgt
+    #    if cache["decode_fsmn"] is None:
+    #        cache_layer_num = len(self.decoders)
+    #        if self.decoders2 is not None:
+    #            cache_layer_num += len(self.decoders2)
+    #        new_cache = [None] * cache_layer_num
+    #    else:
+    #        new_cache = cache["decode_fsmn"]
+    #    for i in range(self.att_layer_num):
+    #        decoder = self.decoders[i]
+    #        x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
+    #            x, None, memory, None, cache=new_cache[i]
+    #        )
+    #        new_cache[i] = c_ret
+
+    #    if self.num_blocks - self.att_layer_num > 1:
+    #        for i in range(self.num_blocks - self.att_layer_num):
+    #            j = i + self.att_layer_num
+    #            decoder = self.decoders2[i]
+    #            x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
+    #                x, None, memory, None, cache=new_cache[j]
+    #            )
+    #            new_cache[j] = c_ret
+
+    #    for decoder in self.decoders3:
+
+    #        x, tgt_mask, memory, memory_mask, _ = decoder.forward_chunk(
+    #            x, None, memory, None, cache=None
+    #        )
+    #    if self.normalize_before:
+    #        x = self.after_norm(x)
+    #    if self.output_layer is not None:
+    #        x = self.output_layer(x)
+    #    cache["decode_fsmn"] = new_cache
+    #    return x
+
     def forward_chunk(
         self,
         memory: torch.Tensor,
@@ -1020,35 +1120,43 @@
             cache_layer_num = len(self.decoders)
             if self.decoders2 is not None:
                 cache_layer_num += len(self.decoders2)
-            new_cache = [None] * cache_layer_num
+            fsmn_cache = [None] * cache_layer_num
         else:
-            new_cache = cache["decode_fsmn"]
+            fsmn_cache = cache["decode_fsmn"]
+
+        if cache["opt"] is None:
+            cache_layer_num = len(self.decoders)
+            opt_cache = [None] * cache_layer_num
+        else:
+            opt_cache = cache["opt"]
+
         for i in range(self.att_layer_num):
             decoder = self.decoders[i]
-            x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
-                x, None, memory, None, cache=new_cache[i]
+            x, memory, fsmn_cache[i], opt_cache[i] = decoder.forward_chunk(
+                x, memory, fsmn_cache=fsmn_cache[i], opt_cache=opt_cache[i],
+                chunk_size=cache["chunk_size"], look_back=cache["decoder_chunk_look_back"]
             )
-            new_cache[i] = c_ret
 
         if self.num_blocks - self.att_layer_num > 1:
             for i in range(self.num_blocks - self.att_layer_num):
                 j = i + self.att_layer_num
                 decoder = self.decoders2[i]
-                x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
-                    x, None, memory, None, cache=new_cache[j]
+                x, memory, fsmn_cache[j], _  = decoder.forward_chunk(
+                    x, memory, fsmn_cache=fsmn_cache[j]
                 )
-                new_cache[j] = c_ret
 
         for decoder in self.decoders3:
-
-            x, tgt_mask, memory, memory_mask, _ = decoder.forward_chunk(
-                x, None, memory, None, cache=None
+            x, memory, _, _ = decoder.forward_chunk(
+                x, memory
             )
         if self.normalize_before:
             x = self.after_norm(x)
         if self.output_layer is not None:
             x = self.output_layer(x)
-        cache["decode_fsmn"] = new_cache
+
+        cache["decode_fsmn"] = fsmn_cache
+        if cache["decoder_chunk_look_back"] > 0 or cache["decoder_chunk_look_back"] == -1:
+            cache["opt"] = opt_cache
         return x
 
     def forward_one_step(
diff --git a/funasr/models/encoder/sanm_encoder.py b/funasr/models/encoder/sanm_encoder.py
index 9e27d4a..e04b9e7 100644
--- a/funasr/models/encoder/sanm_encoder.py
+++ b/funasr/models/encoder/sanm_encoder.py
@@ -114,8 +114,44 @@
         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):
     """
@@ -837,11 +873,56 @@
         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 @@
             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
+            new_cache = cache["opt"]
 
-                    # intermediate outputs are also normalized
-                    if self.normalize_before:
-                        encoder_out = self.after_norm(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]
 
-                    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)
+        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+len(self.encoders0)] = encoder_outs[0], encoder_outs[1]
 
         if self.normalize_before:
             xs_pad = self.after_norm(xs_pad)
+        if cache["encoder_chunk_look_back"] > 0 or cache["encoder_chunk_look_back"] == -1:
+            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):
diff --git a/funasr/modules/attention.py b/funasr/modules/attention.py
index ab59493..b007d58 100644
--- a/funasr/modules/attention.py
+++ b/funasr/modules/attention.py
@@ -456,6 +456,44 @@
         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 MultiHeadedAttentionSANMwithMask(MultiHeadedAttentionSANM):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -667,6 +705,35 @@
         scores = torch.matmul(q_h, k_h.transpose(-2, -1))
         return self.forward_attention(v_h, scores, memory_mask)
 
+    def forward_chunk(self, x, memory, 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 = self.forward_qkv(x, memory)
+        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
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        return self.forward_attention(v_h, scores, None), cache
+
 
 class MultiHeadSelfAttention(nn.Module):
     """Multi-Head Attention layer.

--
Gitblit v1.9.1