src/Spike_driven/Q_transformer_encoder.py

# Copyright 2019 Shigeki Karita
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Transformer encoder definition."""

from typing import List, Optional, Tuple

import torch
import torch.nn as nn
from typeguard import typechecked

from espnet2.asr.ctc import CTC
from espnet2.asr.encoder.abs_encoder import AbsEncoder
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet2.asr.encoder.Spike_driven.Spike_driven_modules.Q_attention import *

from espnet.nets.pytorch_backend.transformer.embedding import PositionalEncoding
from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
# from espnet2.asr_transducer.normalization import RMSNorm
from espnet.nets.pytorch_backend.transformer.multi_layer_conv import (
    Conv1dLinear,
    MultiLayeredConv1d,
)
from espnet2.asr.encoder.Spike_driven.Spike_driven_modules.Q_positionwise_feed_forward import Q_PositionwiseFeedForward, Q_GLU
from espnet.nets.pytorch_backend.transformer.repeat import repeat
from espnet.nets.pytorch_backend.transformer.subsampling import (
    Conv1dSubsampling2,
    Conv2dSubsampling,
    Conv2dSubsampling1,
    Conv2dSubsampling2,
    Conv2dSubsampling6,
    Conv2dSubsampling8,
    TooShortUttError,
    check_short_utt,
)
from espnet2.asr.encoder.Spike_driven.Q_trick import MultiSpike

class Q_Transformer_EncoderLayer(nn.Module):
    """Encoder layer module.

    Args:
        size (int): Input dimension.
        self_attn (torch.nn.Module): Self-attention module instance.
            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention` instance
            can be used as the argument.
        feed_forward (torch.nn.Module): Feed-forward module instance.
            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
            can be used as the argument.
        dropout_rate (float): Dropout rate.
        normalize_before (bool): Whether to use layer_norm before the first block.
        concat_after (bool): Whether to concat attention layer's input and output.
            if True, additional linear will be applied.
            i.e. x -> x + linear(concat(x, att(x)))
            if False, no additional linear will be applied. i.e. x -> x + att(x)
        stochastic_depth_rate (float): Proability to skip this layer.
            During training, the layer may skip residual computation and return input
            as-is with given probability.
    """

    def __init__(
        self,
        size,
        self_attn,
        feed_forward,
        dropout_rate,
        normalize_before=True,
        concat_after=False,
        stochastic_depth_rate=0.0,
    ):
        """Construct an EncoderLayer object."""
        super(Q_Transformer_EncoderLayer, self).__init__()
        self.self_attn = self_attn
        self.feed_forward = feed_forward
        self.norm1 = LayerNorm(size)
        self.norm2 = LayerNorm(size)
        self.dropout = nn.Dropout(dropout_rate)
        self.size = size
        self.normalize_before = normalize_before
        self.concat_after = concat_after
        if self.concat_after:
            self.concat_linear = nn.Linear(size + size, size)
        self.stochastic_depth_rate = stochastic_depth_rate
        self.ATT_sn = MultiSpike(size)
        self.FFN_sn = MultiSpike(size)

    def forward(self, x, mask, iiter=None, cache=None):
        """Compute encoded features.

        Args:
            x_input (torch.Tensor): Input tensor (#batch, time, size).
            mask (torch.Tensor): Mask tensor for the input (#batch, 1, 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, 1, time).

        """
        skip_layer = False
        # with stochastic depth, residual connection `x + f(x)` becomes
        # `x <- x + 1 / (1 - p) * f(x)` at training time.
        stoch_layer_coeff = 1.0
        if self.training and self.stochastic_depth_rate > 0:
            skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
            stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)

        if skip_layer:
            if cache is not None:
                x = torch.cat([cache, x], dim=1)
            return x, mask

        residual = x
        if self.normalize_before:
            x = self.norm1(x)

        if cache is None:
            x_q = x
        else:
            assert cache.shape == (x.shape[0], x.shape[1] - 1, self.size)
            x_q = x[:, -1:, :]
            residual = residual[:, -1:, :]
            mask = None if mask is None else mask[:, -1:, :]

        x_q = self.ATT_sn(x_q, iiter)
        x = self.ATT_sn(x, iiter)
        if self.concat_after:
            x_concat = torch.cat((x, self.self_attn(x_q, x, x, mask, iiter)), dim=-1)
            x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
        else:
            x = residual + stoch_layer_coeff * self.dropout(
                self.self_attn(x_q, x, x, mask, iiter)
            )
        if not self.normalize_before:
            x = self.norm1(x)

        residual = x
        x = self.FFN_sn(x, iiter)
        if self.normalize_before:
            x = self.norm2(x)
        x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x, iiter))
        if not self.normalize_before:
            x = self.norm2(x)

        if cache is not None:
            x = torch.cat([cache, x], dim=1)

        return x, mask


class Q_TransformerEncoder(AbsEncoder):
    """Transformer encoder module.

    Args:
        input_size: input dim
        output_size: dimension of attention
        attention_heads: the number of heads of multi head attention
        linear_units: the number of units of position-wise feed forward
        num_blocks: the number of decoder blocks
        dropout_rate: dropout rate
        attention_dropout_rate: dropout rate in attention
        positional_dropout_rate: dropout rate after adding positional encoding
        input_layer: input layer type
        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
        normalize_before: whether to use layer_norm before the first block
        concat_after: whether to concat attention layer's input and output
            if True, additional linear will be applied.
            i.e. x -> x + linear(concat(x, att(x)))
            if False, no additional linear will be applied.
            i.e. x -> x + att(x)
        positionwise_layer_type: linear of conv1d
        positionwise_conv_kernel_size: kernel size of positionwise conv1d layer
        padding_idx: padding_idx for input_layer=embed
    """

    @typechecked
    def __init__(
        self,
        input_size: int,
        output_size: int = 256,
        attention_heads: int = 4,
        attention_layer_type: str = "selfattn",
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        attention_dropout_rate: float = 0.0,
        input_layer: Optional[str] = "conv2d",
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
        positionwise_layer_type: str = "FFN",
        padding_idx: int = -1,
        interctc_layer_idx: List[int] = [],
        interctc_use_conditioning: bool = False,
        layer_drop_rate: float = 0.0,
    ):
        super().__init__()
        self._output_size = output_size

        if input_layer == "linear":
            self.embed = torch.nn.Sequential(
                torch.nn.Linear(input_size, output_size),
                torch.nn.LayerNorm(output_size),
                torch.nn.Dropout(dropout_rate),
                torch.nn.ReLU(),
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "conv1d2":
            self.embed = Conv1dSubsampling2(
                input_size,
                output_size,
                dropout_rate,
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "conv2d":
            self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
        elif input_layer == "conv2d1":
            self.embed = Conv2dSubsampling1(input_size, output_size, dropout_rate)
        elif input_layer == "conv2d2":
            self.embed = Conv2dSubsampling2(input_size, output_size, dropout_rate)
        elif input_layer == "conv2d6":
            self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
        elif input_layer == "conv2d8":
            self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
        elif input_layer == "embed":
            self.embed = torch.nn.Sequential(
                torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer is None:
            if input_size == output_size:
                self.embed = None
            else:
                self.embed = torch.nn.Linear(input_size, output_size)
        else:
            raise ValueError("unknown input_layer: " + input_layer)
        self.normalize_before = normalize_before
        if attention_layer_type == "selfattn":
            encoder_selfattn_layer = Q_MultiHeadedAttention
            encoder_selfattn_layer_args = (
                attention_heads, 
                output_size, 
                attention_dropout_rate
            )
        elif attention_layer_type == "selfattn_woSoftMax":
            encoder_selfattn_layer = Q_MultiHeadedAttention_woSoftMax
            encoder_selfattn_layer_args = (
                attention_heads, 
                output_size, 
                attention_dropout_rate
            ) 
        elif attention_layer_type == "HierDecayv2":
            encoder_selfattn_layer = Q_MultiHeadedAttention_HierDecay
            encoder_selfattn_layer_args = (
                attention_heads,
                output_size,
                attention_dropout_rate,
            )
        elif attention_layer_type == "HierDecay_woSoftMax":
            encoder_selfattn_layer = Q_MultiHeadedAttention_HierDecay_woSoftMax
            encoder_selfattn_layer_args = (
                attention_heads,
                output_size,
                attention_dropout_rate,
            )
        
        else:
            raise ValueError("unknown encoder_attn_layer: " + attention_layer_type)

        positionwise_layer = Q_PositionwiseFeedForward
        positionwise_layer_args = (
            output_size,
            linear_units,
            dropout_rate,
        )
        

        if "HierDecay" in attention_layer_type:
            self.encoders = repeat(
                num_blocks,
                lambda lnum: Q_Transformer_EncoderLayer(
                    output_size,
                    encoder_selfattn_layer(*encoder_selfattn_layer_args, lnum),
                    positionwise_layer(*positionwise_layer_args),
                    dropout_rate,
                    normalize_before,
                    concat_after,
                ),
                layer_drop_rate,
            )
        else:
            self.encoders = repeat(
                num_blocks,
                lambda lnum: Q_Transformer_EncoderLayer(
                    output_size,
                    encoder_selfattn_layer(*encoder_selfattn_layer_args),
                    positionwise_layer(*positionwise_layer_args),
                    dropout_rate,
                    normalize_before,
                    concat_after,
                ),
                layer_drop_rate,
            )
        
        if self.normalize_before:
            self.after_norm = LayerNorm(output_size)

        self.interctc_layer_idx = interctc_layer_idx
        if len(interctc_layer_idx) > 0:
            assert 0 < min(interctc_layer_idx) and max(interctc_layer_idx) < num_blocks
        self.interctc_use_conditioning = interctc_use_conditioning
        self.conditioning_layer = None

    def output_size(self) -> int:
        return self._output_size

    def forward(
        self,
        xs_pad: torch.Tensor,
        ilens: torch.Tensor,
        iiter: int = 0,
        prev_states: torch.Tensor = None,
        ctc: CTC = None,
        return_all_hs: bool = False,
    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
        """Embed positions in tensor.

        Args:
            xs_pad: input tensor (B, L, D)
            ilens: input length (B)
            prev_states: Not to be used now.
            ctc (CTC): ctc module for intermediate CTC loss
            return_all_hs (bool): whether to return all hidden states

        Returns:
            position embedded tensor and mask
        """
        masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
        # print('iiter:{}'.format(iiter))
        if self.embed is None:
            xs_pad = xs_pad
        elif (
            isinstance(self.embed, Conv2dSubsampling)
            or isinstance(self.embed, Conv1dSubsampling2)
            or isinstance(self.embed, Conv2dSubsampling1)
            or isinstance(self.embed, Conv2dSubsampling2)
            or isinstance(self.embed, Conv2dSubsampling6)
            or isinstance(self.embed, Conv2dSubsampling8)
        ):
            short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
            if short_status:
                raise TooShortUttError(
                    f"has {xs_pad.size(1)} frames and is too short for subsampling "
                    + f"(it needs more than {limit_size} frames), return empty results",
                    xs_pad.size(1),
                    limit_size,
                )
            xs_pad, masks = self.embed(xs_pad, masks)
        else:
            xs_pad = self.embed(xs_pad)
        intermediate_outs = []
        if len(self.interctc_layer_idx) == 0:
            for layer_idx, encoder_layer in enumerate(self.encoders):
                xs_pad, masks = encoder_layer(xs_pad, masks, iiter)
                if return_all_hs:
                    if isinstance(xs_pad, tuple):
                        intermediate_outs.append(xs_pad[0])
                    else:
                        intermediate_outs.append(xs_pad)

                
        else:
            for layer_idx, encoder_layer in enumerate(self.encoders):
                xs_pad, masks = encoder_layer(xs_pad, masks, iiter)

                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)

        olens = masks.squeeze(1).sum(1)
        # from IPython import embed; embed()
        if len(intermediate_outs) > 0:
            return (xs_pad, intermediate_outs), olens, None
        
        return xs_pad, olens, None