tensorflow atrous convoltion

conv2d_same 就是专门针对 atrous convolution 的 dilation 进行的卷积 

https://blog.csdn.net/u013200703/article/details/79300217
def conv2d_same(inputs, num_outputs, kernel_size, stride, rate=1, scope=None):
  if stride == 1:
    return slim.conv2d(inputs, num_outputs, kernel_size, stride=1, rate=rate,
                       padding='SAME', scope=scope)
  else:
    kernel_size_effective = kernel_size + (kernel_size - 1) * (rate - 1)
    pad_total = kernel_size_effective - 1
    pad_beg = pad_total // 2
    pad_end = pad_total - pad_beg
    inputs = tf.pad(inputs,
                    [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
    return slim.conv2d(inputs, num_outputs, kernel_size, stride=stride,
                       rate=rate, padding='VALID', scope=scope)

其实conv2d_same是为了实现

net = slim.conv2d(inputs, num_outputs, 3, stride=1, padding='SAME')
net = subsample(net, factor=stride)

所以

pad_total = (kernel_size - 1) * rate