def __init__(
self,
in_channels,
out_channels,
intra_model,
inter_model,
num_layers=1,
norm="ln",
K=200,
num_spks=2,
skip_around_intra=True,
linear_layer_after_inter_intra=True,
use_global_pos_enc=False,
max_length=20000,
):
super(Dual_Path_Model, self).__init__()
self.K = K
self.num_spks = num_spks
self.num_layers = num_layers
self.norm = select_norm(norm, in_channels, 3)
self.conv1d = nn.Conv1d(in_channels, out_channels, 1, bias=False)
self.use_global_pos_enc = use_global_pos_enc
if self.use_global_pos_enc:
self.pos_enc = PositionalEncoding(max_length)
self.dual_mdl = nn.ModuleList([])
for i in range(num_layers):
self.dual_mdl.append(
copy.deepcopy(
Dual_Computation_Block(
intra_model,
inter_model,
out_channels,
norm,
skip_around_intra=skip_around_intra,
linear_layer_after_inter_intra=linear_layer_after_inter_intra,
)
)
)
self.conv2d = nn.Conv2d(
out_channels, out_channels * num_spks, kernel_size=1
)
self.end_conv1x1 = nn.Conv1d(out_channels, in_channels, 1, bias=False)
self.prelu = nn.PReLU()
self.activation = nn.ReLU()
# gated output layer
self.output = nn.Sequential(
nn.Conv1d(out_channels, out_channels, 1), nn.Tanh()
)
self.output_gate = nn.Sequential(
nn.Conv1d(out_channels, out_channels, 1), nn.Sigmoid()
)
def __init__(
self,
num_layers,
d_model,
nhead,
d_ffn=2048,
input_shape=None,
kdim=None,
vdim=None,
dropout=0.1,
activation="swish",
kernel_size=31,
bias=True,
use_positional_encoding=False,
):
super(SBConformerEncoderBlock, self).__init__()
self.use_positional_encoding = use_positional_encoding
if activation == "relu":
activation = nn.ReLU
elif activation == "gelu":
activation = nn.GELU
elif activation == "swish":
activation = Swish
else:
raise ValueError("unknown activation")
self.mdl = ConformerEncoder(
num_layers=num_layers,
nhead=nhead,
d_ffn=d_ffn,
input_shape=input_shape,
d_model=d_model,
kdim=kdim,
vdim=vdim,
dropout=dropout,
activation=activation,
kernel_size=kernel_size,
bias=bias,
)
if use_positional_encoding:
self.pos_enc = PositionalEncoding(input_size=d_model)
def __init__(
self,
num_layers,
d_model,
nhead,
d_ffn=2048,
input_shape=None,
kdim=None,
vdim=None,
dropout=0.1,
activation="relu",
return_attention=False,
num_modules=1,
use_group_comm=False,
use_positional_encoding=False,
norm_before=False,
):
super(SBTransformerBlock, self).__init__()
self.use_positional_encoding = use_positional_encoding
if activation == "relu":
activation = nn.ReLU
elif activation == "gelu":
activation = nn.GELU
else:
raise ValueError("unknown activation")
self.mdl = TransformerEncoder(
num_layers=num_layers,
nhead=nhead,
d_ffn=d_ffn,
input_shape=input_shape,
d_model=d_model,
kdim=kdim,
vdim=vdim,
dropout=dropout,
activation=activation,
normalize_before=norm_before,
)
if use_positional_encoding:
self.pos_enc = PositionalEncoding(input_size=d_model)