Introduce convolution-augmented Emformer layer prototype (#2324)

Summary:
Introduces prototype of convolution-augmented Emformer layer. At a high level, it incorporates Conformer's macaron feedforward network structure and convolution module with Emformer.

Pull Request resolved: https://github.com/pytorch/audio/pull/2324

Reviewed By: mthrok

Differential Revision: D35734252

Pulled By: hwangjeff

fbshipit-source-id: c7ea0bdcfe53a948b00881a74f1f1e1928f5ac57

torchaudio/prototype/models/conv_emformer.py

+import math
+from typing import List, Optional, Tuple
+
+import torch
+from torchaudio.models.emformer import _EmformerAttention
+
+
+def _get_activation_module(activation: str) -> torch.nn.Module:
+    if activation == "relu":
+        return torch.nn.ReLU()
+    elif activation == "gelu":
+        return torch.nn.GELU()
+    elif activation == "silu":
+        return torch.nn.SiLU()
+    else:
+        raise ValueError(f"Unsupported activation {activation}")
+
+
+class _ResidualContainer(torch.nn.Module):
+    def __init__(self, module: torch.nn.Module, output_weight: int):
+        super().__init__()
+        self.module = module
+        self.output_weight = output_weight
+
+    def forward(self, input: torch.Tensor):
+        output = self.module(input)
+        return output * self.output_weight + input
+
+
+class _ConvolutionModule(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        segment_length: int,
+        right_context_length: int,
+        kernel_size: int,
+        activation: str = "silu",
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.input_dim = input_dim
+        self.segment_length = segment_length
+        self.right_context_length = right_context_length
+        self.state_size = kernel_size - 1
+
+        self.pre_conv = torch.nn.Sequential(
+            torch.nn.LayerNorm(input_dim), torch.nn.Linear(input_dim, 2 * input_dim, bias=True), torch.nn.GLU()
+        )
+        self.conv = torch.nn.Conv1d(
+            in_channels=input_dim,
+            out_channels=input_dim,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=0,
+            groups=input_dim,
+        )
+        self.post_conv = torch.nn.Sequential(
+            torch.nn.LayerNorm(input_dim),
+            _get_activation_module(activation),
+            torch.nn.Linear(input_dim, input_dim, bias=True),
+            torch.nn.Dropout(p=dropout),
+        )
+
+    def _split_right_context(self, utterance: torch.Tensor, right_context: torch.Tensor) -> torch.Tensor:
+        T, B, D = right_context.size()
+        assert T % self.right_context_length == 0
+        num_segments = T // self.right_context_length
+        # (num_segments, right context length, B, D)
+        right_context_segments = right_context.reshape(num_segments, self.right_context_length, B, D)
+        right_context_segments = right_context_segments.permute(0, 2, 1, 3).reshape(
+            num_segments * B, self.right_context_length, D
+        )
+        pad_segments = [
+            utterance[(idx + 1) * self.segment_length : (idx + 1) * self.segment_length + self.state_size, :, :]
+            for idx in range(0, num_segments)
+        ]  # [(kernel_size - 1, B, D), ...]
+        pad_segments = torch.cat(pad_segments, dim=1).permute(1, 0, 2)  # (num_segments * B, kernel_size - 1, D)
+        return torch.cat([pad_segments, right_context_segments], dim=1).permute(0, 2, 1)
+
+    def _merge_right_context(self, right_context: torch.Tensor, B: int) -> torch.Tensor:
+        # (num_segments * B, D, right_context_length)
+        right_context = right_context.reshape(-1, B, self.input_dim, self.right_context_length)
+        right_context = right_context.permute(0, 3, 1, 2)
+        return right_context.reshape(-1, B, self.input_dim)  # (right_context_length * num_segments, B, D)
+
+    def forward(
+        self, utterance: torch.Tensor, right_context: torch.Tensor, state: Optional[torch.Tensor]
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        input = torch.cat((right_context, utterance))  # input: (T, B, D)
+        x = self.pre_conv(input)
+        x_right_context, x_utterance = x[: right_context.size(0), :, :], x[right_context.size(0) :, :, :]
+        x_utterance = x_utterance.permute(1, 2, 0)  # (B, D, T_utterance)
+
+        if state is None:
+            state = torch.zeros(
+                input.size(1),
+                input.size(2),
+                self.state_size,
+                device=input.device,
+                dtype=input.dtype,
+            )  # (B, D, T)
+        state_x_utterance = torch.cat([state, x_utterance], dim=2)
+        # (B * num_segments, D, right_context_length + kernel_size - 1)
+        right_context_block = self._split_right_context(state_x_utterance.permute(2, 0, 1), x_right_context)
+        conv_right_context_block = self.conv(right_context_block)  # (B * num_segments, D, right_context_length)
+        # (T_right_context, B, D)
+        conv_right_context = self._merge_right_context(conv_right_context_block, input.size(1))
+
+        conv_utterance = self.conv(state_x_utterance)  # (B, D, T_utterance)
+        conv_utterance = conv_utterance.permute(2, 0, 1)
+
+        y = torch.cat([conv_right_context, conv_utterance], dim=0)
+        output = self.post_conv(y) + input
+        new_state = state_x_utterance[:, :, -self.state_size :]
+        return output[right_context.size(0) :], output[: right_context.size(0)], new_state
+
+    def infer(
+        self, utterance: torch.Tensor, right_context: torch.Tensor, state: Optional[torch.Tensor]
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        input = torch.cat((utterance, right_context))
+        x = self.pre_conv(input)  # (T, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, T)
+
+        if state is None:
+            state = torch.zeros(
+                input.size(1),
+                input.size(2),
+                self.state_size,
+                device=input.device,
+                dtype=input.dtype,
+            )  # (B, D, T)
+        state_x = torch.cat([state, x], dim=2)
+        conv_out = self.conv(state_x)
+        conv_out = conv_out.permute(2, 0, 1)  # T, B, D
+        output = self.post_conv(conv_out) + input
+        new_state = state_x[:, :, -self.state_size - right_context.size(0) : -right_context.size(0)]
+        return output[: utterance.size(0)], output[utterance.size(0) :], new_state
+
+
+class _ConvEmformerLayer(torch.nn.Module):
+    r"""Convolution-augmented Emformer layer that constitutes ConvEmformer.
+
+    Args:
+        input_dim (int): input dimension.
+        num_heads (int): number of attention heads.
+        ffn_dim: (int): hidden layer dimension of feedforward network.
+        segment_length (int): length of each input segment.
+        kernel_size (int): size of kernel to use in convolution module.
+        dropout (float, optional): dropout probability. (Default: 0.0)
+        ffn_activation (str, optional): activation function to use in feedforward network.
+            Must be one of ("relu", "gelu", "silu"). (Default: "relu")
+        left_context_length (int, optional): length of left context. (Default: 0)
+        right_context_length (int, optional): length of right context. (Default: 0)
+        max_memory_size (int, optional): maximum number of memory elements to use. (Default: 0)
+        weight_init_gain (float or None, optional): scale factor to apply when initializing
+            attention module parameters. (Default: ``None``)
+        tanh_on_mem (bool, optional): if ``True``, applies tanh to memory elements. (Default: ``False``)
+        negative_inf (float, optional): value to use for negative infinity in attention weights. (Default: -1e8)
+        conv_activation (str, optional): activation function to use in convolution module.
+            Must be one of ("relu", "gelu", "silu"). (Default: "silu")
+    """
+
+    def __init__(
+        self,
+        input_dim: int,
+        num_heads: int,
+        ffn_dim: int,
+        segment_length: int,
+        kernel_size: int,
+        dropout: float = 0.0,
+        ffn_activation: str = "relu",
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        max_memory_size: int = 0,
+        weight_init_gain: Optional[float] = None,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+        conv_activation: str = "silu",
+    ):
+        super().__init__()
+        # TODO: implement talking heads attention.
+        self.attention = _EmformerAttention(
+            input_dim=input_dim,
+            num_heads=num_heads,
+            dropout=dropout,
+            weight_init_gain=weight_init_gain,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+        )
+        self.dropout = torch.nn.Dropout(dropout)
+        self.memory_op = torch.nn.AvgPool1d(kernel_size=segment_length, stride=segment_length, ceil_mode=True)
+
+        activation_module = _get_activation_module(ffn_activation)
+        self.ffn0 = _ResidualContainer(
+            torch.nn.Sequential(
+                torch.nn.LayerNorm(input_dim),
+                torch.nn.Linear(input_dim, ffn_dim),
+                activation_module,
+                torch.nn.Dropout(dropout),
+                torch.nn.Linear(ffn_dim, input_dim),
+                torch.nn.Dropout(dropout),
+            ),
+            0.5,
+        )
+        self.ffn1 = _ResidualContainer(
+            torch.nn.Sequential(
+                torch.nn.LayerNorm(input_dim),
+                torch.nn.Linear(input_dim, ffn_dim),
+                activation_module,
+                torch.nn.Dropout(dropout),
+                torch.nn.Linear(ffn_dim, input_dim),
+                torch.nn.Dropout(dropout),
+            ),
+            0.5,
+        )
+        self.layer_norm_input = torch.nn.LayerNorm(input_dim)
+        self.layer_norm_output = torch.nn.LayerNorm(input_dim)
+
+        self.conv = _ConvolutionModule(
+            input_dim=input_dim,
+            kernel_size=kernel_size,
+            activation=conv_activation,
+            dropout=dropout,
+            segment_length=segment_length,
+            right_context_length=right_context_length,
+        )
+
+        self.left_context_length = left_context_length
+        self.segment_length = segment_length
+        self.max_memory_size = max_memory_size
+        self.input_dim = input_dim
+        self.kernel_size = kernel_size
+        self.use_mem = max_memory_size > 0
+
+    def _init_state(self, batch_size: int, device: Optional[torch.device]) -> List[torch.Tensor]:
+        empty_memory = torch.zeros(self.max_memory_size, batch_size, self.input_dim, device=device)
+        left_context_key = torch.zeros(self.left_context_length, batch_size, self.input_dim, device=device)
+        left_context_val = torch.zeros(self.left_context_length, batch_size, self.input_dim, device=device)
+        past_length = torch.zeros(1, batch_size, dtype=torch.int32, device=device)
+        conv_cache = torch.zeros(
+            batch_size,
+            self.input_dim,
+            self.kernel_size - 1,
+            device=device,
+        )
+        return [empty_memory, left_context_key, left_context_val, past_length, conv_cache]
+
+    def _unpack_state(self, state: List[torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        past_length = state[3][0][0].item()
+        past_left_context_length = min(self.left_context_length, past_length)
+        past_mem_length = min(self.max_memory_size, math.ceil(past_length / self.segment_length))
+        pre_mems = state[0][self.max_memory_size - past_mem_length :]
+        lc_key = state[1][self.left_context_length - past_left_context_length :]
+        lc_val = state[2][self.left_context_length - past_left_context_length :]
+        conv_cache = state[4]
+        return pre_mems, lc_key, lc_val, conv_cache
+
+    def _pack_state(
+        self,
+        next_k: torch.Tensor,
+        next_v: torch.Tensor,
+        update_length: int,
+        mems: torch.Tensor,
+        conv_cache: torch.Tensor,
+        state: List[torch.Tensor],
+    ) -> List[torch.Tensor]:
+        new_k = torch.cat([state[1], next_k])
+        new_v = torch.cat([state[2], next_v])
+        state[0] = torch.cat([state[0], mems])[-self.max_memory_size :]
+        state[1] = new_k[new_k.shape[0] - self.left_context_length :]
+        state[2] = new_v[new_v.shape[0] - self.left_context_length :]
+        state[3] = state[3] + update_length
+        state[4] = conv_cache
+        return state
+
+    def _apply_pre_attention(
+        self, utterance: torch.Tensor, right_context: torch.Tensor, summary: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        x = torch.cat([right_context, utterance, summary])
+        ffn0_out = self.ffn0(x)
+        layer_norm_input_out = self.layer_norm_input(ffn0_out)
+        layer_norm_input_right_context, layer_norm_input_utterance, layer_norm_input_summary = (
+            layer_norm_input_out[: right_context.size(0)],
+            layer_norm_input_out[right_context.size(0) : right_context.size(0) + utterance.size(0)],
+            layer_norm_input_out[right_context.size(0) + utterance.size(0) :],
+        )
+        return ffn0_out, layer_norm_input_right_context, layer_norm_input_utterance, layer_norm_input_summary
+
+    def _apply_post_attention(
+        self,
+        rc_output: torch.Tensor,
+        ffn0_out: torch.Tensor,
+        conv_cache: Optional[torch.Tensor],
+        rc_length: int,
+        utterance_length: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        result = self.dropout(rc_output) + ffn0_out[: rc_length + utterance_length]
+        conv_utterance, conv_right_context, conv_cache = self.conv(result[rc_length:], result[:rc_length], conv_cache)
+        result = torch.cat([conv_right_context, conv_utterance])
+        result = self.ffn1(result)
+        result = self.layer_norm_output(result)
+        output_utterance, output_right_context = result[rc_length:], result[:rc_length]
+        return output_utterance, output_right_context, conv_cache
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        lengths: torch.Tensor,
+        right_context: torch.Tensor,
+        mems: torch.Tensor,
+        attention_mask: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        r"""Forward pass for training.
+
+        B: batch size;
+        D: feature dimension of each frame;
+        T: number of utterance frames;
+        R: number of right context frames;
+        M: number of memory elements.
+
+        Args:
+            utterance (torch.Tensor): utterance frames, with shape `(T, B, D)`.
+            lengths (torch.Tensor): with shape `(B,)` and i-th element representing
+                number of valid frames for i-th batch element in ``utterance``.
+            right_context (torch.Tensor): right context frames, with shape `(R, B, D)`.
+            mems (torch.Tensor): memory elements, with shape `(M, B, D)`.
+            attention_mask (torch.Tensor): attention mask for underlying attention module.
+
+        Returns:
+            (Tensor, Tensor, Tensor):
+                Tensor
+                    encoded utterance frames, with shape `(T, B, D)`.
+                Tensor
+                    updated right context frames, with shape `(R, B, D)`.
+                Tensor
+                    updated memory elements, with shape `(M, B, D)`.
+        """
+        if self.use_mem:
+            summary = self.memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)
+        else:
+            summary = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+
+        (
+            ffn0_out,
+            layer_norm_input_right_context,
+            layer_norm_input_utterance,
+            layer_norm_input_summary,
+        ) = self._apply_pre_attention(utterance, right_context, summary)
+
+        rc_output, output_mems = self.attention(
+            utterance=layer_norm_input_utterance,
+            lengths=lengths,
+            right_context=layer_norm_input_right_context,
+            summary=layer_norm_input_summary,
+            mems=mems,
+            attention_mask=attention_mask,
+        )
+
+        output_utterance, output_right_context, _ = self._apply_post_attention(
+            rc_output, ffn0_out, None, right_context.size(0), utterance.size(0)
+        )
+
+        return output_utterance, output_right_context, output_mems
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        lengths: torch.Tensor,
+        right_context: torch.Tensor,
+        state: Optional[List[torch.Tensor]],
+        mems: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]:
+        r"""Forward pass for inference.
+
+        B: batch size;
+        D: feature dimension of each frame;
+        T: number of utterance frames;
+        R: number of right context frames;
+        M: number of memory elements.
+
+        Args:
+            utterance (torch.Tensor): utterance frames, with shape `(T, B, D)`.
+            lengths (torch.Tensor): with shape `(B,)` and i-th element representing
+                number of valid frames for i-th batch element in ``utterance``.
+            right_context (torch.Tensor): right context frames, with shape `(R, B, D)`.
+            state (List[torch.Tensor] or None): list of tensors representing layer internal state
+                generated in preceding invocation of ``infer``.
+            mems (torch.Tensor): memory elements, with shape `(M, B, D)`.
+
+        Returns:
+            (Tensor, Tensor, List[torch.Tensor], Tensor):
+                Tensor
+                    encoded utterance frames, with shape `(T, B, D)`.
+                Tensor
+                    updated right context frames, with shape `(R, B, D)`.
+                List[Tensor]
+                    list of tensors representing layer internal state
+                    generated in current invocation of ``infer``.
+                Tensor
+                    updated memory elements, with shape `(M, B, D)`.
+        """
+        if self.use_mem:
+            summary = self.memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[:1]
+        else:
+            summary = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+
+        (
+            ffn0_out,
+            layer_norm_input_right_context,
+            layer_norm_input_utterance,
+            layer_norm_input_summary,
+        ) = self._apply_pre_attention(utterance, right_context, summary)
+
+        if state is None:
+            state = self._init_state(layer_norm_input_utterance.size(1), device=layer_norm_input_utterance.device)
+        pre_mems, lc_key, lc_val, conv_cache = self._unpack_state(state)
+
+        rc_output, next_m, next_k, next_v = self.attention.infer(
+            utterance=layer_norm_input_utterance,
+            lengths=lengths,
+            right_context=layer_norm_input_right_context,
+            summary=layer_norm_input_summary,
+            mems=pre_mems,
+            left_context_key=lc_key,
+            left_context_val=lc_val,
+        )
+
+        output_utterance, output_right_context, conv_cache = self._apply_post_attention(
+            rc_output, ffn0_out, conv_cache, right_context.size(0), utterance.size(0)
+        )
+        output_state = self._pack_state(next_k, next_v, utterance.size(0), mems, conv_cache, state)
+        return output_utterance, output_right_context, output_state, next_m