network.py

# Copyright (c) 2022 NVIDIA CORPORATION. 
#   Licensed under the MIT license.

import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F

from util import weight_scaling_init

torch.manual_seed(0)
np.random.seed(0)


# Transformer (encoder) https://github.com/jadore801120/attention-is-all-you-need-pytorch
# Original Copyright 2017 Victor Huang
#  MIT License (https://opensource.org/licenses/MIT)

class ScaledDotProductAttention(nn.Module):
    ''' Scaled Dot-Product Attention '''

    def __init__(self, temperature, attn_dropout=0.1):
        super().__init__()
        self.temperature = temperature
        self.dropout = nn.Dropout(attn_dropout)

    def forward(self, q, k, v, mask=None):

        attn = torch.matmul(q / self.temperature, k.transpose(2, 3))

        if mask is not None:
            _MASKING_VALUE = -1e9 if attn.dtype == torch.float32 else -1e4
            attn = attn.masked_fill(mask == 0, _MASKING_VALUE)

        attn = self.dropout(F.softmax(attn, dim=-1))
        output = torch.matmul(attn, v)

        return output, attn


class MultiHeadAttention(nn.Module):
    ''' Multi-Head Attention module '''

    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
        super().__init__()

        self.n_head = n_head
        self.d_k = d_k
        self.d_v = d_v

        self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)
        self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
        self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
        self.fc = nn.Linear(n_head * d_v, d_model, bias=False)

        self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)

        self.dropout = nn.Dropout(dropout)
        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)


    def forward(self, q, k, v, mask=None):

        d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
        sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)

        residual = q

        # Pass through the pre-attention projection: b x lq x (n*dv)
        # Separate different heads: b x lq x n x dv
        q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
        k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
        v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)

        # Transpose for attention dot product: b x n x lq x dv
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)

        if mask is not None:
            mask = mask.unsqueeze(1)   # For head axis broadcasting.

        q, attn = self.attention(q, k, v, mask=mask)

        # Transpose to move the head dimension back: b x lq x n x dv
        # Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
        q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
        q = self.dropout(self.fc(q))
        q += residual

        q = self.layer_norm(q)

        return q, attn


class PositionwiseFeedForward(nn.Module):
    ''' A two-feed-forward-layer module '''

    def __init__(self, d_in, d_hid, dropout=0.1):
        super().__init__()
        self.w_1 = nn.Linear(d_in, d_hid) # position-wise
        self.w_2 = nn.Linear(d_hid, d_in) # position-wise
        self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):

        residual = x

        x = self.w_2(F.relu(self.w_1(x)))
        x = self.dropout(x)
        x += residual

        x = self.layer_norm(x)

        return x


def get_subsequent_mask(seq):
    ''' For masking out the subsequent info. '''
    sz_b, len_s = seq.size()
    subsequent_mask = (1 - torch.triu(
        torch.ones((1, len_s, len_s), device=seq.device), diagonal=1)).bool()
    return subsequent_mask


class PositionalEncoding(nn.Module):

    def __init__(self, d_hid, n_position=200):
        super(PositionalEncoding, self).__init__()

        # Not a parameter
        self.register_buffer('pos_table', self._get_sinusoid_encoding_table(n_position, d_hid))

    def _get_sinusoid_encoding_table(self, n_position, d_hid):
        ''' Sinusoid position encoding table '''
        # TODO: make it with torch instead of numpy

        def get_position_angle_vec(position):
            return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]

        sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)])
        sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
        sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1

        return torch.FloatTensor(sinusoid_table).unsqueeze(0)

    def forward(self, x):
        return x + self.pos_table[:, :x.size(1)].clone().detach()


class EncoderLayer(nn.Module):
    ''' Compose with two layers '''

    def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.0):
        super(EncoderLayer, self).__init__()
        self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
        self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout)

    def forward(self, enc_input, slf_attn_mask=None):
        enc_output, enc_slf_attn = self.slf_attn(
            enc_input, enc_input, enc_input, mask=slf_attn_mask)
        enc_output = self.pos_ffn(enc_output)
        return enc_output, enc_slf_attn


class TransformerEncoder(nn.Module):
    ''' A encoder model with self attention mechanism. '''

    def __init__(
            self, d_word_vec=512, n_layers=2, n_head=8, d_k=64, d_v=64,
            d_model=512, d_inner=2048, dropout=0.1, n_position=624, scale_emb=False):

        super().__init__()

        # self.src_word_emb = nn.Embedding(n_src_vocab, d_word_vec, padding_idx=pad_idx)
        if n_position > 0:
            self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position)
        else:
            self.position_enc = lambda x: x
        self.dropout = nn.Dropout(p=dropout)
        self.layer_stack = nn.ModuleList([
            EncoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
            for _ in range(n_layers)])
        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
        self.scale_emb = scale_emb
        self.d_model = d_model

    def forward(self, src_seq, src_mask, return_attns=False):

        enc_slf_attn_list = []

        # -- Forward
        # enc_output = self.src_word_emb(src_seq)
        enc_output = src_seq
        if self.scale_emb:
            enc_output *= self.d_model ** 0.5
        enc_output = self.dropout(self.position_enc(enc_output))
        enc_output = self.layer_norm(enc_output)

        for enc_layer in self.layer_stack:
            enc_output, enc_slf_attn = enc_layer(enc_output, slf_attn_mask=src_mask)
            enc_slf_attn_list += [enc_slf_attn] if return_attns else []

        if return_attns:
            return enc_output, enc_slf_attn_list
        return enc_output


# CleanUNet architecture


def padding(x, D, K, S):
    """padding zeroes to x so that denoised audio has the same length"""

    L = x.shape[-1]
    for _ in range(D):
        if L < K:
            L = 1
        else:
            L = 1 + np.ceil((L - K) / S)

    for _ in range(D):
        L = (L - 1) * S + K
    
    L = int(L)
    x = F.pad(x, (0, L - x.shape[-1]))
    return x


class CleanUNet(nn.Module):
    """ CleanUNet architecture. """

    def __init__(self, channels_input=1, channels_output=1,
                 channels_H=64, max_H=768,
                 encoder_n_layers=8, kernel_size=4, stride=2,
                 tsfm_n_layers=3, 
                 tsfm_n_head=8,
                 tsfm_d_model=512, 
                 tsfm_d_inner=2048):
        
        """
        Parameters:
        channels_input (int):   input channels
        channels_output (int):  output channels
        channels_H (int):       middle channels H that controls capacity
        max_H (int):            maximum H
        encoder_n_layers (int): number of encoder/decoder layers D
        kernel_size (int):      kernel size K
        stride (int):           stride S
        tsfm_n_layers (int):    number of self attention blocks N
        tsfm_n_head (int):      number of heads in each self attention block
        tsfm_d_model (int):     d_model of self attention
        tsfm_d_inner (int):     d_inner of self attention
        """

        super(CleanUNet, self).__init__()

        self.channels_input = channels_input
        self.channels_output = channels_output
        self.channels_H = channels_H
        self.max_H = max_H
        self.encoder_n_layers = encoder_n_layers
        self.kernel_size = kernel_size
        self.stride = stride

        self.tsfm_n_layers = tsfm_n_layers
        self.tsfm_n_head = tsfm_n_head
        self.tsfm_d_model = tsfm_d_model
        self.tsfm_d_inner = tsfm_d_inner

        # encoder and decoder
        self.encoder = nn.ModuleList()
        self.decoder = nn.ModuleList()

        for i in range(encoder_n_layers):
            self.encoder.append(nn.Sequential(
                nn.Conv1d(channels_input, channels_H, kernel_size, stride),
                nn.ReLU(),
                nn.Conv1d(channels_H, channels_H * 2, 1), 
                nn.GLU(dim=1)
            ))
            channels_input = channels_H

            if i == 0:
                # no relu at end
                self.decoder.append(nn.Sequential(
                    nn.Conv1d(channels_H, channels_H * 2, 1), 
                    nn.GLU(dim=1),
                    nn.ConvTranspose1d(channels_H, channels_output, kernel_size, stride)
                ))
            else:
                self.decoder.insert(0, nn.Sequential(
                    nn.Conv1d(channels_H, channels_H * 2, 1), 
                    nn.GLU(dim=1),
                    nn.ConvTranspose1d(channels_H, channels_output, kernel_size, stride),
                    nn.ReLU()
                ))
            channels_output = channels_H
            
            # double H but keep below max_H
            channels_H *= 2
            channels_H = min(channels_H, max_H)
        
        # self attention block
        self.tsfm_conv1 = nn.Conv1d(channels_output, tsfm_d_model, kernel_size=1)
        self.tsfm_encoder = TransformerEncoder(d_word_vec=tsfm_d_model, 
                                               n_layers=tsfm_n_layers, 
                                               n_head=tsfm_n_head, 
                                               d_k=tsfm_d_model // tsfm_n_head, 
                                               d_v=tsfm_d_model // tsfm_n_head, 
                                               d_model=tsfm_d_model, 
                                               d_inner=tsfm_d_inner, 
                                               dropout=0.0, 
                                               n_position=0, 
                                               scale_emb=False)
        self.tsfm_conv2 = nn.Conv1d(tsfm_d_model, channels_output, kernel_size=1)

        # weight scaling initialization
        for layer in self.modules():
            if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
                weight_scaling_init(layer)

    def forward(self, noisy_audio):
        # (B, L) -> (B, C, L)
        if len(noisy_audio.shape) == 2:
            noisy_audio = noisy_audio.unsqueeze(1)
        B, C, L = noisy_audio.shape
        assert C == 1
        
        # normalization and padding
        std = noisy_audio.std(dim=2, keepdim=True) + 1e-3
        noisy_audio /= std
        x = padding(noisy_audio, self.encoder_n_layers, self.kernel_size, self.stride)
        
        # encoder
        skip_connections = []
        for downsampling_block in self.encoder:
            x = downsampling_block(x)
            skip_connections.append(x)
        skip_connections = skip_connections[::-1]

        # attention mask for causal inference; for non-causal, set attn_mask to None
        len_s = x.shape[-1]  # length at bottleneck
        attn_mask = (1 - torch.triu(torch.ones((1, len_s, len_s), device=x.device), diagonal=1)).bool()

        x = self.tsfm_conv1(x)  # C 1024 -> 512
        x = x.permute(0, 2, 1)
        x = self.tsfm_encoder(x, src_mask=attn_mask)
        x = x.permute(0, 2, 1)
        x = self.tsfm_conv2(x)  # C 512 -> 1024

        # decoder
        for i, upsampling_block in enumerate(self.decoder):
            skip_i = skip_connections[i]
            x += skip_i[:, :, :x.shape[-1]]
            x = upsampling_block(x)

        x = x[:, :, :L] * std
        return x


if __name__ == '__main__':
    import json
    import argparse 
    import os

    parser = argparse.ArgumentParser()
    parser.add_argument('-c', '--config', type=str, default='configs/DNS-large-full.json', 
                        help='JSON file for configuration')
    args = parser.parse_args()

    with open(args.config) as f:
        data = f.read()
    config = json.loads(data)
    network_config = config["network_config"]

    model = CleanUNet(**network_config).cuda()
    from util import print_size
    print_size(model, keyword="tsfm")
    
    input_data = torch.ones([4,1,int(4.5*16000)]).cuda()
    output = model(input_data)
    print(output.shape)

    y = torch.rand([4,1,int(4.5*16000)]).cuda()
    loss = torch.nn.MSELoss()(y, output)
    loss.backward()
    print(loss.item())