def forward(self, signal):
if is_stft(self.config.data.features) or is_mel(self.config.data.features):
signal = compute_torch_stft(
signal.squeeze(-1),
self.config.data.features
)
if is_stft(self.config.data.features):
signal = torch.log(signal + 1e-4)
if is_mel(self.config.data.features):
signal = nn.functional.conv1d(
signal,
self.filterbanks.unsqueeze(-1)
)
signal = torch.log(signal + 1e-4)
signal = signal.unsqueeze(1)
signal = signal.repeat(1, 3, 1, 1)
signal = self.input_norm(signal)
h = self.backbone.features(signal)
features = self.global_maxpool(h).squeeze(-1).squeeze(-1)
class_logits = self.output_transform(features)
r = dict(
class_logits=class_logits
)
return r
def __init__(self, experiment, device="cuda"):
super().__init__()
self.device = device
self.experiment = experiment
self.config = experiment.config
if is_mel(self.config.data.features):
self.filterbanks = torch.from_numpy(
make_mel_filterbanks(self.config.data.features)).to(self.device)
self.input_norm = nn.BatchNorm2d(3)
if self.config.network.backbone == "resnet18":
self.backbone = resnet18(pretrained=None)
elif self.config.network.backbone == "resnet34":
self.backbone = resnet34(pretrained=None)
self.global_maxpool = nn.AdaptiveMaxPool2d(1)
total_depth = self.backbone.last_linear.in_features
self.output_transform = nn.Sequential(
nn.BatchNorm1d(total_depth),
nn.Linear(total_depth, total_depth),
nn.BatchNorm1d(total_depth),
nn.PReLU(total_depth),
nn.Dropout(p=self.config.network.output_dropout),
nn.Linear(total_depth, self.config.data._n_classes)
)
self.to(self.device)
def __init__(self, experiment, device="cuda"):
super().__init__()
self.device = device
self.experiment = experiment
self.config = experiment.config
if is_mel(self.config.data.features):
self.filterbanks = torch.from_numpy(
make_mel_filterbanks(self.config.data.features)).to(
self.device)
self.conv_modules = torch.nn.ModuleList()
self.rnns = torch.nn.ModuleList()
total_depth = 0
for k in range(self.config.network.num_conv_blocks):
input_size = 2 if not k else depth
depth = int(self.config.network.growth_rate**k *
self.config.network.conv_base_depth)
rnn_size = 128
if k >= self.config.network.start_deep_supervision_on:
if self.config.network.aggregation_type == "max":
total_depth += depth
elif self.config.network.aggregation_type == "rnn":
total_depth += rnn_size * 2
self.rnns.append(
nn.Sequential(
nn.LayerNorm((depth, )),
nn.GRU(depth,
rnn_size,
batch_first=True,
bidirectional=True)))
modules = [nn.BatchNorm2d(input_size)]
modules.extend([
nn.Conv2d(input_size, depth, kernel_size=3, padding=1),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.BatchNorm2d(depth),
nn.PReLU(depth),
ResnetBlock2d(depth)
])
self.conv_modules.append(nn.Sequential(*modules))
self.global_maxpool = nn.AdaptiveMaxPool2d(1)
self.output_transform = nn.Sequential(
nn.BatchNorm1d(total_depth), nn.Linear(total_depth, total_depth),
nn.BatchNorm1d(total_depth), nn.PReLU(total_depth),
nn.Dropout(p=self.config.network.output_dropout),
nn.Linear(total_depth, 2))
self.to(self.device)
def forward(self, signal):
if is_stft(self.config.data.features) or is_mel(self.config.data.features):
signal = compute_torch_stft(
signal.squeeze(-1),
self.config.data.features
)
if is_stft(self.config.data.features):
signal = torch.log(signal + 1e-4)
if is_mel(self.config.data.features):
signal = nn.functional.conv1d(
signal,
self.filterbanks.unsqueeze(-1)
)
signal = torch.log(signal + 1e-4)
signal = signal.unsqueeze(1)
signal = self._add_frequency_encoding(signal)
features = []
h = signal
for k, module in enumerate(self.conv_modules):
h = module(h)
if k >= self.config.network.start_deep_supervision_on:
if self.config.network.aggregation_type == "max":
features.append(self.global_maxpool(h).squeeze(-1).squeeze(-1))
elif self.config.network.aggregation_type == "rnn":
rnn_input = torch.mean(h, 2).permute(0, 2, 1)
outputs, state = self.rnns[
k - self.config.network.start_deep_supervision_on](rnn_input)
features.append(
state.permute(1, 0, 2).contiguous().view(rnn_input.size(0), -1))
features = torch.cat(features, -1)
class_logits = self.output_transform(features)
r = dict(
class_logits=class_logits
)
return r