VDCNN Convolutional Block (Conneau et al. 2017)

conv_block_vdcnn.py

# Convolutional block from http://aclweb.org/anthology/E17-1104
# Anton Melnikov

import numpy as np
import torch
from torch import nn

class ConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels, *,
                 kernel_size, stride=1):
        super().__init__()
        # "same" padding (preserve original temporal dimension length)
        padding = int(np.floor((kernel_size - 1) / 2))
        
        self.conv_1 = nn.Conv1d(in_channels, out_channels, kernel_size,
                                stride=stride, padding=padding)
        self.batch_norm_1 = nn.BatchNorm1d(out_channels)
        self.relu_1 = nn.ReLU()
        
        self.conv_2 = nn.Conv1d(out_channels, out_channels,
                                kernel_size, stride=stride, padding=padding)
        self.batch_norm_2 = nn.BatchNorm1d(out_channels)
        self.relu_2 = nn.ReLU()
        self.init_conv()
    
    def init_conv(self):
        nn.init.kaiming_uniform(self.conv_1.weight.data)
        nn.init.kaiming_uniform(self.conv_2.weight.data)
    
    def forward(self, X):
        X = self.conv_1(X)
        X = self.batch_norm_1(X)
        X = self.relu_1(X)
        X = self.conv_2(X)
        X = self.batch_norm_2(X)
        X = self.relu_2(X)
        return X