|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
|
|
class ChannelAttention(nn.Module): |
|
def __init__(self, in_planes, ratio=16): |
|
super(ChannelAttention, self).__init__() |
|
self.avg_pool = nn.AdaptiveAvgPool2d(1) |
|
self.max_pool = nn.AdaptiveMaxPool2d(1) |
|
|
|
self.fc1 = nn.Conv2d(in_planes, in_planes // ratio, 1, bias=False) |
|
self.relu1 = nn.ReLU() |
|
self.fc2 = nn.Conv2d(in_planes // ratio, in_planes, 1, bias=False) |
|
|
|
self.sigmoid = nn.Sigmoid() |
|
|
|
def forward(self, x): |
|
avg_out = self.fc2(self.relu1(self.fc1(self.avg_pool(x)))) |
|
max_out = self.fc2(self.relu1(self.fc1(self.max_pool(x)))) |
|
y = avg_out + max_out |
|
y = self.sigmoid(y) |
|
|
|
return x * y.expand_as(x) |
|
|
|
|
|
class ResCell(nn.Module): |
|
def __init__(self, input_channel, output_channel, stride=1): |
|
super(ResCell, self).__init__() |
|
|
|
self.stride = stride |
|
self.input_channel = input_channel |
|
self.output_channel = output_channel |
|
|
|
if self.stride == -1: |
|
output_size = () |
|
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True) |
|
self.skip = nn.Conv2d(self.input_channel, self.output_channel, kernel_size=1, stride=1, padding=0) |
|
self.conv1 = nn.ConvTranspose2d(self.input_channel, self.output_channel, kernel_size=5, stride=2, padding=2, output_padding=1) |
|
self.conv2 = nn.ConvTranspose2d(self.output_channel, self.output_channel, kernel_size=5, padding=2) |
|
|
|
elif self.stride == 2: |
|
self.skip = nn.Conv2d(self.input_channel, self.output_channel, kernel_size=1, stride=2, padding=0) |
|
self.conv1 = nn.Conv2d(self.input_channel, self.output_channel, kernel_size=5, stride=self.stride, padding=2) |
|
self.conv2 = nn.Conv2d(self.output_channel, self.output_channel, kernel_size=5, padding=2) |
|
|
|
else: |
|
self.conv1 = nn.Conv2d(self.input_channel, self.output_channel, kernel_size=5, stride=self.stride, padding=2) |
|
self.conv2 = nn.Conv2d(self.output_channel, self.output_channel, kernel_size=5, padding=2) |
|
|
|
self.bn1 = nn.BatchNorm2d(self.output_channel) |
|
self.bn2 = nn.BatchNorm2d(self.output_channel) |
|
|
|
|
|
self.cbam = ChannelAttention(self.output_channel) |
|
|
|
def forward(self, x): |
|
if self.stride == -1: |
|
upsampled_x = self.upsample(x) |
|
skip = self.skip(upsampled_x) |
|
x = F.elu(self.bn1(self.conv1(x))) |
|
x = self.conv2(x) |
|
elif self.stride == 2: |
|
skip = self.skip(x) |
|
x = F.elu(self.bn1(self.conv1(x))) |
|
x = self.conv2(x) |
|
else: |
|
skip = x |
|
x = F.elu(self.bn1(self.conv1(x))) |
|
x = self.conv2(x) |
|
|
|
x = self.bn2(x) |
|
x = self.cbam(x) |
|
x = x + skip |
|
x = F.elu(x) |
|
|
|
return x |
|
|
|
|
|
class ResBlock(nn.Module): |
|
def __init__(self, input_channel, output_channel, upsample=False, n_cells=2): |
|
super(ResBlock, self).__init__() |
|
|
|
stride = -1 if upsample else 2 |
|
self.cells = nn.ModuleList([ResCell(input_channel, output_channel, stride=stride)]) |
|
|
|
for _ in range(n_cells - 1): |
|
self.cells.append(ResCell(input_channel, output_channel, stride=1)) |
|
|
|
def forward(self, x): |
|
for cell in self.cells: |
|
x = cell(x) |
|
return x |
|
|
|
|
|
class Encoder(nn.Module): |
|
def __init__(self, input_shape, timbre_dim, N2=0, channel_sizes=None): |
|
super(Encoder, self).__init__() |
|
|
|
if channel_sizes is None: |
|
channel_sizes = [32, 64, 64, 96, 96, 128, 160, 216] |
|
|
|
self.input_shape = input_shape |
|
self.timbre_dim = timbre_dim |
|
self.blocks = nn.ModuleList() |
|
|
|
self.blocks.append(ResBlock(input_channel=1, output_channel=channel_sizes[0], upsample=False, n_cells=1)) |
|
input_channel = channel_sizes[0] |
|
|
|
for c in channel_sizes[1:]: |
|
self.blocks.append(ResBlock(input_channel=input_channel, output_channel=c, upsample=False, n_cells=1 + N2)) |
|
input_channel = c |
|
|
|
self.flatten = nn.Flatten() |
|
self.mu_timbre = nn.Linear(self._get_flattened_dim(), timbre_dim) |
|
self.sigma_timbre = nn.Linear(self._get_flattened_dim(), timbre_dim) |
|
|
|
def _get_flattened_dim(self): |
|
x = torch.zeros((1,) + self.input_shape) |
|
for block in self.blocks: |
|
x = block(x) |
|
x = self.flatten(x) |
|
return x.shape[1] |
|
|
|
def reparameterize(self, mu, logvar): |
|
std = torch.exp(0.5*logvar) |
|
eps = torch.randn_like(std) |
|
return mu + eps*std |
|
|
|
def forward(self, x): |
|
for block in self.blocks: |
|
x = block(x) |
|
|
|
x = self.flatten(x) |
|
mu = self.mu_timbre(x) |
|
logvar = self.sigma_timbre(x) |
|
latent_vector = self.reparameterize(mu, logvar) |
|
|
|
|
|
|
|
|
|
return mu, logvar, latent_vector |
|
|
|
|
|
class Decoder(nn.Module): |
|
def __init__(self, timbre_dim, N2=0, N3=8, channel_sizes=None): |
|
super(Decoder, self).__init__() |
|
|
|
if channel_sizes is None: |
|
channel_sizes = [32, 64, 64, 96, 96, 128, 160, 216] |
|
|
|
self.conv_shape = [-1, channel_sizes[-1], 2 ** (9 - N3), 2 ** (8 - N3)] |
|
|
|
self.dense = nn.Linear(timbre_dim, self.conv_shape[1] * self.conv_shape[2] * self.conv_shape[3]) |
|
self.blocks = nn.ModuleList() |
|
|
|
input_channel = channel_sizes[-1] |
|
for c in list(reversed(channel_sizes))[1:]: |
|
self.blocks.append(ResBlock(input_channel=input_channel, output_channel=c, upsample=True, n_cells=1 + N2)) |
|
input_channel = c |
|
|
|
self.decoder_conv = nn.ConvTranspose2d(channel_sizes[0], 1, kernel_size=5, stride=2, padding=2, output_padding=1) |
|
|
|
def forward(self, x): |
|
x = F.elu(self.dense(x)) |
|
x = x.view(-1, self.conv_shape[1], self.conv_shape[2], self.conv_shape[3]) |
|
for block in self.blocks: |
|
x = block(x) |
|
|
|
x = self.decoder_conv(x) |
|
x = torch.sigmoid(x) |
|
return x |