Created
October 24, 2017 23:27
-
-
Save MaximumEntropy/76a236035abeb1ad07089d0e3224ac42 to your computer and use it in GitHub Desktop.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
class DilatedConvSentenceEncoder(nn.Module): | |
"""A Sentence Encoder with Dilated Convs.""" | |
def __init__( | |
self, input_dim=512, hidden_dim=4096, n_layers=7, | |
dropout=0.5, batch_first=True | |
): | |
"""Initialize params.""" | |
super(DilatedConvSentenceEncoder, self).__init__() | |
self.input_dim = input_dim | |
self.hidden_dim = hidden_dim | |
self.n_layers = n_layers | |
self.dropout = dropout | |
# 512 x 128 | |
self.conv1 = nn.Conv1d( | |
in_channels=input_dim, out_channels=hidden_dim // 4, | |
kernel_size=4, stride=2, padding=12, dilation=8, bias=False | |
) | |
# self.bn1 = nn.BatchNorm1d(num_features=input_dim * 2) | |
# 1024 x 64 | |
self.conv2 = nn.Conv1d( | |
in_channels=hidden_dim // 4, out_channels=hidden_dim // 2, | |
kernel_size=4, stride=2, padding=9, dilation=6, bias=False | |
) | |
# self.bn2 = nn.BatchNorm1d(num_features=input_dim * 4) | |
# 2048 x 32 | |
self.conv3 = nn.Conv1d( | |
in_channels=hidden_dim // 2, out_channels=hidden_dim, | |
kernel_size=4, stride=2, padding=6, dilation=4, bias=False | |
) | |
# self.bn3 = nn.BatchNorm1d(num_features=input_dim * 8) | |
# 4096 x 16 | |
self.conv4 = nn.Conv1d( | |
in_channels=hidden_dim, out_channels=hidden_dim, | |
kernel_size=4, stride=2, padding=3, dilation=2, bias=False | |
) | |
# self.bn4 = nn.BatchNorm1d(num_features=input_dim * 8) | |
# 4096 x 8 | |
self.conv5 = nn.Conv1d( | |
in_channels=hidden_dim, out_channels=hidden_dim, | |
kernel_size=4, stride=2, padding=3, dilation=2, bias=False | |
) | |
# self.bn5 = nn.BatchNorm1d(num_features=input_dim * 8) | |
# 4096 x 4 | |
self.conv6 = nn.Conv1d( | |
in_channels=hidden_dim, out_channels=hidden_dim, | |
kernel_size=4, stride=2, padding=3, dilation=2, bias=False | |
) | |
# self.bn6 = nn.BatchNorm1d(num_features=input_dim * 8) | |
# 4096 x 2 | |
self.conv7 = nn.Conv1d( | |
in_channels=hidden_dim, out_channels=hidden_dim, | |
kernel_size=4, stride=2, padding=3, dilation=2, bias=False | |
) | |
# self.bn7 = nn.BatchNorm1d(num_features=input_dim * 8) | |
def forward(self, input): | |
x = input.transpose(1, 2) | |
x = F.relu(F.dropout(self.conv1(x), p=self.dropout, training=self.training)) | |
x = F.relu(F.dropout(self.conv2(x), p=self.dropout, training=self.training)) | |
x = F.relu(F.dropout(self.conv3(x), p=self.dropout, training=self.training)) | |
x = F.relu(F.dropout(self.conv4(x), p=self.dropout, training=self.training)) | |
x = F.relu(F.dropout(self.conv5(x), p=self.dropout, training=self.training)) | |
x = F.relu(F.dropout(self.conv6(x), p=self.dropout, training=self.training)) | |
return F.relu(F.dropout(self.conv7(x), p=self.dropout, training=self.training)).squeeze() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment