Merge pull request #2124 from zh794390558/new_api

[asr] optimize attention cache used for attention ; 0-dim tensor for model export

demos/streaming_asr_server/local/rtf_from_log.py

@@ -38,4 +38,4 @@ if __name__ == '__main__':
         T += m['T']
         P += m['P']
 
-    print(f"RTF: {P/T}")
+    print(f"RTF: {P/T}, utts: {n}")

paddlespeech/cli/asr/infer.py

@@ -365,7 +365,7 @@ class ASRExecutor(BaseExecutor):
         except Exception as e:
             logger.exception(e)
             logger.error(
-                "can not open the audio file, please check the audio file format is 'wav'. \n \
+                f"can not open the audio file, please check the audio file({audio_file}) format is 'wav'. \n \
                  you can try to use sox to change the file format.\n \
                  For example: \n \
                  sample rate: 16k \n \

paddlespeech/cli/executor.py

@@ -108,19 +108,20 @@ class BaseExecutor(ABC):
             Dict[str, Union[str, os.PathLike]]: A dict with ids and inputs.
         """
         if self._is_job_input(input_):
+            # .job/.scp/.txt file
             ret = self._get_job_contents(input_)
         else:
+            # job from stdin
             ret = OrderedDict()
-
             if input_ is None:  # Take input from stdin
                 if not sys.stdin.isatty(
                 ):  # Avoid getting stuck when stdin is empty.
                     for i, line in enumerate(sys.stdin):
                         line = line.strip()
-                        if len(line.split(' ')) == 1:
+                        if len(line.split()) == 1:
                             ret[str(i + 1)] = line
-                        elif len(line.split(' ')) == 2:
-                            id_, info = line.split(' ')
+                        elif len(line.split()) == 2:
+                            id_, info = line.split()
                             ret[id_] = info
                         else:  # No valid input info from one line.
                             continue
@@ -170,7 +171,8 @@ class BaseExecutor(ABC):
             bool: return `True` for job input, `False` otherwise.
         """
         return input_ and os.path.isfile(input_) and (input_.endswith('.job') or
-                                                      input_.endswith('.txt'))
+                                                      input_.endswith('.txt') or
+                                                      input_.endswith('.scp'))
 
     def _get_job_contents(
             self, job_input: os.PathLike) -> Dict[str, Union[str, os.PathLike]]:
@@ -189,7 +191,7 @@ class BaseExecutor(ABC):
                 line = line.strip()
                 if not line:
                     continue
-                k, v = line.split(' ')
+                k, v = line.split() # space or \t
                 job_contents[k] = v
         return job_contents
 

paddlespeech/s2t/models/u2/u2.py

@@ -605,29 +605,42 @@ class U2BaseModel(ASRInterface, nn.Layer):
             xs: paddle.Tensor,
             offset: int,
             required_cache_size: int,
-            subsampling_cache: Optional[paddle.Tensor]=None,
-            elayers_output_cache: Optional[List[paddle.Tensor]]=None,
-            conformer_cnn_cache: Optional[List[paddle.Tensor]]=None,
-    ) -> Tuple[paddle.Tensor, paddle.Tensor, List[paddle.Tensor], List[
-            paddle.Tensor]]:
+            att_cache: paddle.Tensor = paddle.zeros([0, 0, 0, 0]),
+            cnn_cache: paddle.Tensor = paddle.zeros([0, 0, 0, 0]),
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
         """ Export interface for c++ call, give input chunk xs, and return
             output from time 0 to current chunk.
+
         Args:
-            xs (paddle.Tensor): chunk input
-            subsampling_cache (Optional[paddle.Tensor]): subsampling cache
-            elayers_output_cache (Optional[List[paddle.Tensor]]):
-                transformer/conformer encoder layers output cache
-            conformer_cnn_cache (Optional[List[paddle.Tensor]]): conformer
-                cnn cache
+            xs (paddle.Tensor): chunk input, with shape (b=1, time, mel-dim),
+                where `time == (chunk_size - 1) * subsample_rate + \
+                        subsample.right_context + 1`
+            offset (int): current offset in encoder output time stamp
+            required_cache_size (int): cache size required for next chunk
+                compuation
+                >=0: actual cache size
+                <0: means all history cache is required
+            att_cache (paddle.Tensor): cache tensor for KEY & VALUE in
+                transformer/conformer attention, with shape
+                (elayers, head, cache_t1, d_k * 2), where
+                `head * d_k == hidden-dim` and
+                `cache_t1 == chunk_size * num_decoding_left_chunks`.
+                `d_k * 2` for att key & value.
+            cnn_cache (paddle.Tensor): cache tensor for cnn_module in conformer,
+                (elayers, b=1, hidden-dim, cache_t2), where
+                `cache_t2 == cnn.lorder - 1`
+
         Returns:
-            paddle.Tensor: output, it ranges from time 0 to current chunk.
-            paddle.Tensor: subsampling cache
-            List[paddle.Tensor]: attention cache
-            List[paddle.Tensor]: conformer cnn cache
+            paddle.Tensor: output of current input xs,
+                with shape (b=1, chunk_size, hidden-dim).
+            paddle.Tensor: new attention cache required for next chunk, with
+                dynamic shape (elayers, head, T(?), d_k * 2)
+                depending on required_cache_size.
+            paddle.Tensor: new conformer cnn cache required for next chunk, with
+                same shape as the original cnn_cache.
         """
         return self.encoder.forward_chunk(
-            xs, offset, required_cache_size, subsampling_cache,
-            elayers_output_cache, conformer_cnn_cache)
+            xs, offset, required_cache_size, att_cache, cnn_cache)
 
     # @jit.to_static
     def ctc_activation(self, xs: paddle.Tensor) -> paddle.Tensor:

paddlespeech/s2t/models/u2_st/u2_st.py

@@ -401,29 +401,42 @@ class U2STBaseModel(nn.Layer):
             xs: paddle.Tensor,
             offset: int,
             required_cache_size: int,
-            subsampling_cache: Optional[paddle.Tensor]=None,
-            elayers_output_cache: Optional[List[paddle.Tensor]]=None,
-            conformer_cnn_cache: Optional[List[paddle.Tensor]]=None,
-    ) -> Tuple[paddle.Tensor, paddle.Tensor, List[paddle.Tensor], List[
-            paddle.Tensor]]:
+            att_cache: paddle.Tensor = paddle.zeros([0, 0, 0, 0]),
+            cnn_cache: paddle.Tensor = paddle.zeros([0, 0, 0, 0]),
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
         """ Export interface for c++ call, give input chunk xs, and return
             output from time 0 to current chunk.
+
         Args:
-            xs (paddle.Tensor): chunk input
-            subsampling_cache (Optional[paddle.Tensor]): subsampling cache
-            elayers_output_cache (Optional[List[paddle.Tensor]]):
-                transformer/conformer encoder layers output cache
-            conformer_cnn_cache (Optional[List[paddle.Tensor]]): conformer
-                cnn cache
+            xs (paddle.Tensor): chunk input, with shape (b=1, time, mel-dim),
+                where `time == (chunk_size - 1) * subsample_rate + \
+                        subsample.right_context + 1`
+            offset (int): current offset in encoder output time stamp
+            required_cache_size (int): cache size required for next chunk
+                compuation
+                >=0: actual cache size
+                <0: means all history cache is required
+            att_cache (paddle.Tensor): cache tensor for KEY & VALUE in
+                transformer/conformer attention, with shape
+                (elayers, head, cache_t1, d_k * 2), where
+                `head * d_k == hidden-dim` and
+                `cache_t1 == chunk_size * num_decoding_left_chunks`.
+                `d_k * 2` for att key & value.
+            cnn_cache (paddle.Tensor): cache tensor for cnn_module in conformer,
+                (elayers, b=1, hidden-dim, cache_t2), where
+                `cache_t2 == cnn.lorder - 1`
+
         Returns:
-            paddle.Tensor: output, it ranges from time 0 to current chunk.
-            paddle.Tensor: subsampling cache
-            List[paddle.Tensor]: attention cache
-            List[paddle.Tensor]: conformer cnn cache
+            paddle.Tensor: output of current input xs,
+                with shape (b=1, chunk_size, hidden-dim).
+            paddle.Tensor: new attention cache required for next chunk, with
+                dynamic shape (elayers, head, T(?), d_k * 2)
+                depending on required_cache_size.
+            paddle.Tensor: new conformer cnn cache required for next chunk, with
+                same shape as the original cnn_cache.
         """
         return self.encoder.forward_chunk(
-            xs, offset, required_cache_size, subsampling_cache,
-            elayers_output_cache, conformer_cnn_cache)
+            xs, offset, required_cache_size, att_cache, cnn_cache)
 
     # @jit.to_static
     def ctc_activation(self, xs: paddle.Tensor) -> paddle.Tensor:

paddlespeech/s2t/modules/attention.py

@@ -86,7 +86,8 @@ class MultiHeadedAttention(nn.Layer):
     def forward_attention(self,
             value: paddle.Tensor, 
             scores: paddle.Tensor,
-                          mask: Optional[paddle.Tensor]) -> paddle.Tensor:
+            mask: paddle.Tensor = paddle.ones([0, 0, 0], dtype=paddle.bool),
+        ) -> paddle.Tensor:
         """Compute attention context vector.
         Args:
             value (paddle.Tensor): Transformed value, size
@@ -94,14 +95,23 @@ class MultiHeadedAttention(nn.Layer):
             scores (paddle.Tensor): Attention score, size
                 (#batch, n_head, time1, time2).
             mask (paddle.Tensor): Mask, size (#batch, 1, time2) or
-                (#batch, time1, time2).
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
         Returns:
-            paddle.Tensor: Transformed value weighted
-                by the attention score, (#batch, time1, d_model).
+            paddle.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
         """
         n_batch = value.shape[0]
-        if mask is not None:
+
+        # When `if mask.size(2) > 0` be True:
+        # 1. training.
+        # 2. oonx(16/4, chunk_size/history_size), feed real cache and real mask for the 1st chunk.
+        # When will `if mask.size(2) > 0` be False?
+        # 1. onnx(16/-1, -1/-1, 16/0)
+        # 2. jit (16/-1, -1/-1, 16/0, 16/4)
+        if paddle.shape(mask)[2] > 0: # time2 > 0
             mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            # for last chunk, time2 might be larger than scores.size(-1)
+            mask = mask[:, :, :, :paddle.shape(scores)[-1]]
             scores = scores.masked_fill(mask, -float('inf'))
             attn = paddle.softmax(
                 scores, axis=-1).masked_fill(mask,
@@ -121,21 +131,66 @@ class MultiHeadedAttention(nn.Layer):
                 query: paddle.Tensor,
                 key: paddle.Tensor,
                 value: paddle.Tensor,
-                mask: Optional[paddle.Tensor]) -> paddle.Tensor:
+                mask: paddle.Tensor = paddle.ones([0,0,0], dtype=paddle.bool),
+                pos_emb: paddle.Tensor = paddle.empty([0]),
+                cache: paddle.Tensor = paddle.zeros([0,0,0,0])
+                ) -> Tuple[paddle.Tensor, paddle.Tensor]:
         """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
+            query (paddle.Tensor): Query tensor (#batch, time1, size).
+            key (paddle.Tensor): Key tensor (#batch, time2, size).
+            value (paddle.Tensor): Value tensor (#batch, time2, size).
+            mask (paddle.Tensor): Mask tensor (#batch, 1, time2) or
                 (#batch, time1, time2).
+                1.When applying cross attention between decoder and encoder,
+                the batch padding mask for input is in (#batch, 1, T) shape.
+                2.When applying self attention of encoder,
+                the mask is in (#batch, T, T)  shape.
+                3.When applying self attention of decoder,
+                the mask is in (#batch, L, L)  shape.
+                4.If the different position in decoder see different block
+                of the encoder, such as Mocha, the passed in mask could be
+                in (#batch, L, T) shape. But there is no such case in current
+                Wenet.
+            cache (paddle.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
         Returns:
-            torch.Tensor: Output tensor (#batch, time1, d_model).
+            paddle.Tensor: Output tensor (#batch, time1, d_model).
+            paddle.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+
         """
         q, k, v = self.forward_qkv(query, key, value)
+
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if paddle.shape(cache)[0] > 0:
+            # last dim `d_k * 2` for (key, val)
+            key_cache, value_cache = paddle.split(cache, 2, axis=-1)
+            k = paddle.concat([key_cache, k], axis=2)
+            v = paddle.concat([value_cache, v], axis=2)
+        # We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = paddle.concat((k, v), axis=-1)
+
         scores = paddle.matmul(q,
                                k.transpose([0, 1, 3, 2])) / math.sqrt(self.d_k)
-        return self.forward_attention(v, scores, mask)
+        return self.forward_attention(v, scores, mask), new_cache
 
 
 class RelPositionMultiHeadedAttention(MultiHeadedAttention):
@@ -192,23 +247,55 @@ class RelPositionMultiHeadedAttention(MultiHeadedAttention):
                 query: paddle.Tensor,
                 key: paddle.Tensor,
                 value: paddle.Tensor,
-                pos_emb: paddle.Tensor,
-                mask: Optional[paddle.Tensor]):
+                mask: paddle.Tensor = paddle.ones([0,0,0], dtype=paddle.bool),
+                pos_emb: paddle.Tensor = paddle.empty([0]),
+                cache: paddle.Tensor = paddle.zeros([0,0,0,0])
+                ) -> Tuple[paddle.Tensor, paddle.Tensor]:
         """Compute 'Scaled Dot Product Attention' with rel. positional encoding.
         Args:
             query (paddle.Tensor): Query tensor (#batch, time1, size).
             key (paddle.Tensor): Key tensor (#batch, time2, size).
             value (paddle.Tensor): Value tensor (#batch, time2, size).
-            pos_emb (paddle.Tensor): Positional embedding tensor
-                (#batch, time1, size).
             mask (paddle.Tensor): Mask tensor (#batch, 1, time2) or
-                (#batch, time1, time2).
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
+            pos_emb (paddle.Tensor): Positional embedding tensor
+                (#batch, time2, size).
+            cache (paddle.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
         Returns:
             paddle.Tensor: Output tensor (#batch, time1, d_model).
+            paddle.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
         """
         q, k, v = self.forward_qkv(query, key, value)
         q = q.transpose([0, 2, 1, 3])  # (batch, time1, head, d_k)
 
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if paddle.shape(cache)[0] > 0:
+            # last dim `d_k * 2` for (key, val)
+            key_cache, value_cache = paddle.split(cache, 2, axis=-1)
+            k = paddle.concat([key_cache, k], axis=2)
+            v = paddle.concat([value_cache, v], axis=2)
+        # We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = paddle.concat((k, v), axis=-1)
+
         n_batch_pos = pos_emb.shape[0]
         p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
         p = p.transpose([0, 2, 1, 3])  # (batch, head, time1, d_k)
@@ -234,4 +321,4 @@ class RelPositionMultiHeadedAttention(MultiHeadedAttention):
         scores = (matrix_ac + matrix_bd) / math.sqrt(
             self.d_k)  # (batch, head, time1, time2)
 
-        return self.forward_attention(v, scores, mask)
+        return self.forward_attention(v, scores, mask), new_cache
\ No newline at end of file

paddlespeech/s2t/modules/conformer_convolution.py

@@ -108,15 +108,17 @@ class ConvolutionModule(nn.Layer):
 
     def forward(self,
                 x: paddle.Tensor,
-                mask_pad: Optional[paddle.Tensor]=None,
-                cache: Optional[paddle.Tensor]=None
+                mask_pad: paddle.Tensor= paddle.ones([0,0,0], dtype=paddle.bool),
+                cache: paddle.Tensor= paddle.zeros([0,0,0]),
                 ) -> Tuple[paddle.Tensor, paddle.Tensor]:
         """Compute convolution module.
         Args:
             x (paddle.Tensor): Input tensor (#batch, time, channels).
-            mask_pad (paddle.Tensor): used for batch padding, (#batch, channels, time).
+            mask_pad (paddle.Tensor): used for batch padding (#batch, 1, time),
+                (0, 0, 0) means fake mask.
             cache (paddle.Tensor): left context cache, it is only
-                used in causal convolution. (#batch, channels, time')
+                used in causal convolution (#batch, channels, cache_t),
+                (0, 0, 0) meas fake cache.
         Returns:
             paddle.Tensor: Output tensor (#batch, time, channels).
             paddle.Tensor: Output cache tensor (#batch, channels, time')
@@ -125,11 +127,11 @@ class ConvolutionModule(nn.Layer):
         x = x.transpose([0, 2, 1])  # [B, C, T]
 
         # mask batch padding
-        if mask_pad is not None:
+        if paddle.shape(mask_pad)[2] > 0: # time > 0
             x = x.masked_fill(mask_pad, 0.0)
 
         if self.lorder > 0:
-            if cache is None:
+            if paddle.shape(cache)[2] == 0: # cache_t == 0
                 x = nn.functional.pad(
                     x, [self.lorder, 0], 'constant', 0.0, data_format='NCL')
             else:
@@ -143,7 +145,7 @@ class ConvolutionModule(nn.Layer):
             # It's better we just return None if no cache is requried,
             # However, for JIT export, here we just fake one tensor instead of
             # None.
-            new_cache = paddle.zeros([1], dtype=x.dtype)
+            new_cache = paddle.zeros([0, 0, 0], dtype=x.dtype)
 
         # GLU mechanism
         x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
@@ -159,7 +161,7 @@ class ConvolutionModule(nn.Layer):
         x = self.pointwise_conv2(x)
 
         # mask batch padding
-        if mask_pad is not None:
+        if paddle.shape(mask_pad)[2] > 0: # time > 0
             x = x.masked_fill(mask_pad, 0.0)
 
         x = x.transpose([0, 2, 1])  # [B, T, C]

paddlespeech/s2t/modules/decoder_layer.py

@@ -121,11 +121,11 @@ class DecoderLayer(nn.Layer):
 
         if self.concat_after:
             tgt_concat = paddle.cat(
-                (tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)), dim=-1)
+                (tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0]), dim=-1)
             x = residual + self.concat_linear1(tgt_concat)
         else:
             x = residual + self.dropout(
-                self.self_attn(tgt_q, tgt, tgt, tgt_q_mask))
+                self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
         if not self.normalize_before:
             x = self.norm1(x)
 
@@ -134,11 +134,11 @@ class DecoderLayer(nn.Layer):
             x = self.norm2(x)
         if self.concat_after:
             x_concat = paddle.cat(
-                (x, self.src_attn(x, memory, memory, memory_mask)), dim=-1)
+                (x, self.src_attn(x, memory, memory, memory_mask)[0]), dim=-1)
             x = residual + self.concat_linear2(x_concat)
         else:
             x = residual + self.dropout(
-                self.src_attn(x, memory, memory, memory_mask))
+                self.src_attn(x, memory, memory, memory_mask)[0])
         if not self.normalize_before:
             x = self.norm2(x)
 

paddlespeech/s2t/modules/embedding.py

@@ -131,7 +131,7 @@ class PositionalEncoding(nn.Layer, PositionalEncodingInterface):
             offset (int): start offset
             size (int): requried size of position encoding
         Returns:
-            paddle.Tensor: Corresponding position encoding
+            paddle.Tensor: Corresponding position encoding, #[1, T, D].
         """
         assert offset + size < self.max_len
         return self.dropout(self.pe[:, offset:offset + size])

paddlespeech/s2t/modules/encoder.py

@@ -177,7 +177,7 @@ class BaseEncoder(nn.Layer):
             decoding_chunk_size, self.static_chunk_size,
             num_decoding_left_chunks)
         for layer in self.encoders:
-            xs, chunk_masks, _ = layer(xs, chunk_masks, pos_emb, mask_pad)
+            xs, chunk_masks, _, _ = layer(xs, chunk_masks, pos_emb, mask_pad)
         if self.normalize_before:
             xs = self.after_norm(xs)
         # Here we assume the mask is not changed in encoder layers, so just
@@ -190,30 +190,31 @@ class BaseEncoder(nn.Layer):
             xs: paddle.Tensor,
             offset: int,
             required_cache_size: int,
-            subsampling_cache: Optional[paddle.Tensor]=None,
-            elayers_output_cache: Optional[List[paddle.Tensor]]=None,
-            conformer_cnn_cache: Optional[List[paddle.Tensor]]=None,
-    ) -> Tuple[paddle.Tensor, paddle.Tensor, List[paddle.Tensor], List[
-            paddle.Tensor]]:
+            att_cache: paddle.Tensor = paddle.zeros([0,0,0,0]),
+            cnn_cache: paddle.Tensor = paddle.zeros([0,0,0,0]),
+            att_mask: paddle.Tensor = paddle.ones([0,0,0], dtype=paddle.bool),
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
         """ Forward just one chunk
         Args:
-            xs (paddle.Tensor): chunk input, [B=1, T, D]
+            xs (paddle.Tensor): chunk audio feat input, [B=1, T, D], where 
+                `T==(chunk_size-1)*subsampling_rate + subsample.right_context + 1`
             offset (int): current offset in encoder output time stamp
             required_cache_size (int): cache size required for next chunk
                 compuation
                 >=0: actual cache size
                 <0: means all history cache is required
-            subsampling_cache (Optional[paddle.Tensor]): subsampling cache
-            elayers_output_cache (Optional[List[paddle.Tensor]]):
-                transformer/conformer encoder layers output cache
-            conformer_cnn_cache (Optional[List[paddle.Tensor]]): conformer
-                cnn cache
+            att_cache(paddle.Tensor): cache tensor for key & val in 
+                transformer/conformer attention. Shape is 
+                (elayers, head, cache_t1, d_k * 2), where`head * d_k == hidden-dim` 
+                and `cache_t1 == chunk_size * num_decoding_left_chunks`.
+            cnn_cache (paddle.Tensor): cache tensor for cnn_module in conformer, 
+                (elayers, B=1, hidden-dim, cache_t2), where `cache_t2 == cnn.lorder - 1`
         Returns:
-            paddle.Tensor: output of current input xs
-            paddle.Tensor: subsampling cache required for next chunk computation
-            List[paddle.Tensor]: encoder layers output cache required for next
-                chunk computation
-            List[paddle.Tensor]: conformer cnn cache
+            paddle.Tensor: output of current input xs, (B=1, chunk_size, hidden-dim)
+            paddle.Tensor: new attention cache required for next chunk, dyanmic shape 
+                (elayers, head, T, d_k*2) depending on required_cache_size
+            paddle.Tensor: new conformer cnn cache required for next chunk, with
+                same shape as the original cnn_cache
         """
         assert xs.shape[0] == 1  # batch size must be one
         # tmp_masks is just for interface compatibility
@@ -225,50 +226,50 @@ class BaseEncoder(nn.Layer):
         if self.global_cmvn is not None:
             xs = self.global_cmvn(xs)
 
-        xs, pos_emb, _ = self.embed(
-            xs, tmp_masks, offset=offset)  #xs=(B, T, D), pos_emb=(B=1, T, D)
+        # before embed, xs=(B, T, D1), pos_emb=(B=1, T, D)
+        xs, pos_emb, _ = self.embed(xs, tmp_masks, offset=offset) 
+        # after embed, xs=(B=1, chunk_size, hidden-dim)
 
-        if subsampling_cache is not None:
-            cache_size = subsampling_cache.shape[1]  #T
-            xs = paddle.cat((subsampling_cache, xs), dim=1)
-        else:
-            cache_size = 0
+        elayers = paddle.shape(att_cache)[0]
+        cache_t1 = paddle.shape(att_cache)[2]
+        chunk_size = paddle.shape(xs)[1]
+        attention_key_size = cache_t1 + chunk_size
 
         # only used when using `RelPositionMultiHeadedAttention`
         pos_emb = self.embed.position_encoding(
-            offset=offset - cache_size, size=xs.shape[1])
+            offset=offset - cache_t1, size=attention_key_size)
 
         if required_cache_size < 0:
             next_cache_start = 0
         elif required_cache_size == 0:
-            next_cache_start = xs.shape[1]
+            next_cache_start = attention_key_size
         else:
-            next_cache_start = xs.shape[1] - required_cache_size
-        r_subsampling_cache = xs[:, next_cache_start:, :]
-
-        # Real mask for transformer/conformer layers
-        masks = paddle.ones([1, xs.shape[1]], dtype=paddle.bool)
-        masks = masks.unsqueeze(1)  #[B=1, L'=1, T]
-        r_elayers_output_cache = []
-        r_conformer_cnn_cache = []
+            next_cache_start = max(attention_key_size - required_cache_size, 0)
+
+        r_att_cache = []
+        r_cnn_cache = []
         for i, layer in enumerate(self.encoders):
-            attn_cache = None if elayers_output_cache is None else elayers_output_cache[
-                i]
-            cnn_cache = None if conformer_cnn_cache is None else conformer_cnn_cache[
-                i]
-            xs, _, new_cnn_cache = layer(
-                xs,
-                masks,
-                pos_emb,
-                output_cache=attn_cache,
-                cnn_cache=cnn_cache)
-            r_elayers_output_cache.append(xs[:, next_cache_start:, :])
-            r_conformer_cnn_cache.append(new_cnn_cache)
+            # att_cache[i:i+1] = (1, head, cache_t1, d_k*2)
+            # cnn_cache[i] = (B=1, hidden-dim, cache_t2)
+            xs, _, new_att_cache, new_cnn_cache = layer(
+                xs, att_mask, pos_emb,
+                att_cache=att_cache[i:i+1] if elayers > 0 else att_cache,
+                cnn_cache=cnn_cache[i] if paddle.shape(cnn_cache)[0] > 0 else cnn_cache,
+            )
+            # new_att_cache = (1, head, attention_key_size, d_k*2)
+            # new_cnn_cache = (B=1, hidden-dim, cache_t2)
+            r_att_cache.append(new_att_cache[:,:, next_cache_start:, :])
+            r_cnn_cache.append(new_cnn_cache.unsqueeze(0)) # add elayer dim
+
         if self.normalize_before:
             xs = self.after_norm(xs)
 
-        return (xs[:, cache_size:, :], r_subsampling_cache,
-                r_elayers_output_cache, r_conformer_cnn_cache)
+        # r_att_cache (elayers, head, T, d_k*2)
+        # r_cnn_cache （elayers, B=1, hidden-dim, cache_t2)
+        r_att_cache = paddle.concat(r_att_cache, axis=0)
+        r_cnn_cache = paddle.concat(r_cnn_cache, axis=0)
+        return xs, r_att_cache, r_cnn_cache
+
 
     def forward_chunk_by_chunk(
             self,
@@ -313,25 +314,24 @@ class BaseEncoder(nn.Layer):
 
         num_frames = xs.shape[1]
         required_cache_size = decoding_chunk_size * num_decoding_left_chunks
-        subsampling_cache: Optional[paddle.Tensor] = None
-        elayers_output_cache: Optional[List[paddle.Tensor]] = None
-        conformer_cnn_cache: Optional[List[paddle.Tensor]] = None
+
+        att_cache: paddle.Tensor = paddle.zeros([0,0,0,0])
+        cnn_cache: paddle.Tensor = paddle.zeros([0,0,0,0])
+
         outputs = []
         offset = 0
         # Feed forward overlap input step by step
         for cur in range(0, num_frames - context + 1, stride):
             end = min(cur + decoding_window, num_frames)
             chunk_xs = xs[:, cur:end, :]
-            (y, subsampling_cache, elayers_output_cache,
-             conformer_cnn_cache) = self.forward_chunk(
-                 chunk_xs, offset, required_cache_size, subsampling_cache,
-                 elayers_output_cache, conformer_cnn_cache)
+
+            (y, att_cache, cnn_cache) = self.forward_chunk(
+                 chunk_xs, offset, required_cache_size, att_cache, cnn_cache)
+
             outputs.append(y)
             offset += y.shape[1]
         ys = paddle.cat(outputs, 1)
-        # fake mask, just for jit script and compatibility with `forward` api
-        masks = paddle.ones([1, ys.shape[1]], dtype=paddle.bool)
-        masks = masks.unsqueeze(1)
+        masks = paddle.ones([1, 1, ys.shape[1]], dtype=paddle.bool)
         return ys, masks
 
 

paddlespeech/s2t/modules/encoder_layer.py

@@ -75,49 +75,45 @@ class TransformerEncoderLayer(nn.Layer):
             self,
             x: paddle.Tensor,
             mask: paddle.Tensor,
-            pos_emb: Optional[paddle.Tensor]=None,
-            mask_pad: Optional[paddle.Tensor]=None,
-            output_cache: Optional[paddle.Tensor]=None,
-            cnn_cache: Optional[paddle.Tensor]=None,
-    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+            pos_emb: paddle.Tensor,
+            mask_pad: paddle.Tensor= paddle.ones([0,0,0], dtype=paddle.bool),
+            att_cache: paddle.Tensor=paddle.zeros([0,0,0,0]),
+            cnn_cache: paddle.Tensor=paddle.zeros([0,0,0,0]),
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor, paddle.Tensor]:
         """Compute encoded features.
         Args:
-            x (paddle.Tensor): Input tensor (#batch, time, size).
-            mask (paddle.Tensor): Mask tensor for the input (#batch, time).
+            x (paddle.Tensor): (#batch, time, size)
+            mask (paddle.Tensor): Mask tensor for the input (#batch, time，time),
+                (0, 0, 0) means fake mask.
             pos_emb (paddle.Tensor): just for interface compatibility
                 to ConformerEncoderLayer
-            mask_pad (paddle.Tensor): not used here, it's for interface
-                compatibility to ConformerEncoderLayer
-            output_cache (paddle.Tensor): Cache tensor of the output
-                (#batch, time2, size), time2 < time in x.
-            cnn_cache (paddle.Tensor): not used here, it's for interface
-                compatibility to ConformerEncoderLayer
+            mask_pad (paddle.Tensor): does not used in transformer layer,
+                just for unified api with conformer.
+            att_cache (paddle.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
+            cnn_cache (paddle.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2), not used here, it's for interface
+                compatibility to ConformerEncoderLayer.
         Returns:
             paddle.Tensor: Output tensor (#batch, time, size).
-            paddle.Tensor: Mask tensor (#batch, time).
-            paddle.Tensor: Fake cnn cache tensor for api compatibility with Conformer (#batch, channels, time').
+            paddle.Tensor: Mask tensor (#batch, time, time).
+            paddle.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+            paddle.Tensor: cnn_cahce tensor (#batch=1, size, cache_t2).
         """
         residual = x
         if self.normalize_before:
             x = self.norm1(x)
 
-        if output_cache is None:
-            x_q = x
-        else:
-            assert output_cache.shape[0] == x.shape[0]
-            assert output_cache.shape[1] < x.shape[1]
-            assert output_cache.shape[2] == self.size
-            chunk = x.shape[1] - output_cache.shape[1]
-            x_q = x[:, -chunk:, :]
-            residual = residual[:, -chunk:, :]
-            mask = mask[:, -chunk:, :]
+        x_att, new_att_cache = self.self_attn(
+            x, x, x, mask, cache=att_cache
+        )
 
         if self.concat_after:
-            x_concat = paddle.concat(
-                (x, self.self_attn(x_q, x, x, mask)), axis=-1)
+            x_concat = paddle.concat((x, x_att), axis=-1)
             x = residual + self.concat_linear(x_concat)
         else:
-            x = residual + self.dropout(self.self_attn(x_q, x, x, mask))
+            x = residual + self.dropout(x_att)
         if not self.normalize_before:
             x = self.norm1(x)
 
@@ -128,11 +124,8 @@ class TransformerEncoderLayer(nn.Layer):
         if not self.normalize_before:
             x = self.norm2(x)
 
-        if output_cache is not None:
-            x = paddle.concat([output_cache, x], axis=1)
-
-        fake_cnn_cache = paddle.zeros([1], dtype=x.dtype)
-        return x, mask, fake_cnn_cache
+        fake_cnn_cache = paddle.zeros([0,0,0], dtype=x.dtype)
+        return x, mask, new_att_cache, fake_cnn_cache
 
 
 class ConformerEncoderLayer(nn.Layer):
@@ -192,32 +185,41 @@ class ConformerEncoderLayer(nn.Layer):
         self.size = size
         self.normalize_before = normalize_before
         self.concat_after = concat_after
+        if self.concat_after:
             self.concat_linear = Linear(size + size, size)
+        else:
+            self.concat_linear = nn.Identity()
 
     def forward(
             self,
             x: paddle.Tensor,
             mask: paddle.Tensor,
             pos_emb: paddle.Tensor,
-            mask_pad: Optional[paddle.Tensor]=None,
-            output_cache: Optional[paddle.Tensor]=None,
-            cnn_cache: Optional[paddle.Tensor]=None,
-    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+            mask_pad: paddle.Tensor= paddle.ones([0,0,0], dtype=paddle.bool),
+            att_cache: paddle.Tensor=paddle.zeros([0,0,0,0]),
+            cnn_cache: paddle.Tensor=paddle.zeros([0,0,0,0]),
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor, paddle.Tensor]:
         """Compute encoded features.
         Args:
-            x (paddle.Tensor): (#batch, time, size)
-            mask (paddle.Tensor): Mask tensor for the input (#batch, time，time).
-            pos_emb (paddle.Tensor): positional encoding, must not be None
-                for ConformerEncoderLayer.
-            mask_pad (paddle.Tensor): batch padding mask used for conv module, (B, 1, T).
-            output_cache (paddle.Tensor): Cache tensor of the encoder output
-                (#batch, time2, size), time2 < time in x.
+            x (paddle.Tensor): Input tensor (#batch, time, size).
+            mask (paddle.Tensor): Mask tensor for the input (#batch, time, time).
+                (0,0,0) means fake mask.
+            pos_emb (paddle.Tensor): postional encoding, must not be None 
+                for ConformerEncoderLayer
+            mask_pad (paddle.Tensor): batch padding mask used for conv module.
+               (#batch, 1，time), (0, 0, 0) means fake mask.
+            att_cache (paddle.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
             cnn_cache (paddle.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2)
         Returns:
            paddle.Tensor: Output tensor (#batch, time, size).
-            paddle.Tensor: Mask tensor (#batch, time).
-            paddle.Tensor: New cnn cache tensor (#batch, channels, time').
+           paddle.Tensor: Mask tensor (#batch, time, time).
+           paddle.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+           paddle.Tensor: cnn_cahce tensor (#batch, size, cache_t2).
         """
+
         # whether to use macaron style FFN
         if self.feed_forward_macaron is not None:
             residual = x
@@ -233,18 +235,8 @@ class ConformerEncoderLayer(nn.Layer):
         if self.normalize_before:
             x = self.norm_mha(x)
 
-        if output_cache is None:
-            x_q = x
-        else:
-            assert output_cache.shape[0] == x.shape[0]
-            assert output_cache.shape[1] < x.shape[1]
-            assert output_cache.shape[2] == self.size
-            chunk = x.shape[1] - output_cache.shape[1]
-            x_q = x[:, -chunk:, :]
-            residual = residual[:, -chunk:, :]
-            mask = mask[:, -chunk:, :]
-
-        x_att = self.self_attn(x_q, x, x, pos_emb, mask)
+        x_att, new_att_cache = self.self_attn(
+            x, x, x, mask, pos_emb, cache=att_cache)
 
         if self.concat_after:
             x_concat = paddle.concat((x, x_att), axis=-1)
@@ -257,7 +249,7 @@ class ConformerEncoderLayer(nn.Layer):
 
         # convolution module
         # Fake new cnn cache here, and then change it in conv_module
-        new_cnn_cache = paddle.zeros([1], dtype=x.dtype)
+        new_cnn_cache = paddle.zeros([0,0,0], dtype=x.dtype)
         if self.conv_module is not None:
             residual = x
             if self.normalize_before:
@@ -282,7 +274,4 @@ class ConformerEncoderLayer(nn.Layer):
         if self.conv_module is not None:
             x = self.norm_final(x)
 
-        if output_cache is not None:
-            x = paddle.concat([output_cache, x], axis=1)
-
-        return x, mask, new_cnn_cache
+        return x, mask, new_att_cache, new_cnn_cache
\ No newline at end of file

paddlespeech/server/engine/asr/online/python/asr_engine.py

@@ -130,9 +130,9 @@ class PaddleASRConnectionHanddler:
 
         ## conformer
         # cache for conformer online
-        self.subsampling_cache = None
-        self.elayers_output_cache = None
-        self.conformer_cnn_cache = None
+        self.att_cache = paddle.zeros([0,0,0,0])
+        self.cnn_cache = paddle.zeros([0,0,0,0])
+
         self.encoder_out = None
         # conformer decoding state
         self.offset = 0  # global offset in decoding frame unit
@@ -474,11 +474,9 @@ class PaddleASRConnectionHanddler:
             # cur chunk
             chunk_xs = self.cached_feat[:, cur:end, :]
             # forward chunk
-            (y, self.subsampling_cache, self.elayers_output_cache,
-             self.conformer_cnn_cache) = self.model.encoder.forward_chunk(
+            (y, self.att_cache, self.cnn_cache) = self.model.encoder.forward_chunk(
                  chunk_xs, self.offset, required_cache_size,
-                 self.subsampling_cache, self.elayers_output_cache,
-                 self.conformer_cnn_cache)
+                 self.att_cache, self.cnn_cache)
             outputs.append(y)
 
             # update the global offset, in decoding frame unit

setup.py

@@ -71,7 +71,8 @@ base = [
     "colorlog",
     "pathos == 0.2.8",
     "braceexpand",
-    "pyyaml"
+    "pyyaml",
+    "pybind11",
 ]
 
 server = [
@@ -90,7 +91,6 @@ requirements = {
         "gpustat",
         "paddlespeech_ctcdecoders",
         "phkit",
-        "pybind11",
         "pypi-kenlm",
         "snakeviz",
         "sox",