thoroughimages/octconv

A PyTorch Implementation for Octave Convolution in Drop an Octave

octconv's Issues

TypeError: __init__() missing 1 required positional argument: 'padding_mode'

Thank you for your work that does help me a lot. Here I have one question, if the input of octave convolution network is (batch_size, channels, H, W) where H is equal to 1000, while W is equal to one. Then I use the code self.convs = nn.Sequential(OctConv2d('first', in_channels=1, out_channels=32, kernel_size=3),;it raises a TypeError:init() missing 1 required positional argument: 'padding_mode' in
super(OctConv2d, self).__init__( in_channels, out_channels, kernel_size, stride, padding, dilation, False, _pair(0), 1, bias)
Thank you!

Please kindly help answer : at the first layer how to distinguish low-frequency and high-frequency feature maps

Please kindly help answer : at the first layer how to distinguish low-frequency and high-frequency feature maps

Strided OctConv probably wrong

OctConv/octconv.py

Line 72 in 6d4c6b4

high_group_hh = F.conv2d(high_group, self.high_hh_weight, self.high_hh_bias, self.stride,

The emerging consensus is that this is the wrong way to implement strided convolutions in OctConv. See terrychenism/OctaveConv#4

RuntimeError: The size of tensor a (11) must match the size of tensor b (10) at non-singleton dimension 3

This is my code :

class Accurate_Modle((torch.nn.Module)):
def init(self,data_kind,padding):
super(Accurate_Modle, self).init()
self.con6 = torch.nn.Conv2d(2*112,384, kernel_size=1, stride=1, padding=0)
self.con7 = torch.nn.Conv2d(384, 384, kernel_size=1, stride=1, padding=0)
self.con8 = torch.nn.Conv2d(384, 384, kernel_size=1, stride=1, padding=0)
self.con9 = torch.nn.Conv2d(384, 1, kernel_size=1, stride=1, padding=0)
if data_kind=='mb':

        # torch.nn.init.constant_(self.con5.bias, 0)
        self.conv1 = torch.nn.Sequential(OctConv2d('first', in_channels=1, out_channels=112, kernel_size=3),
                                   OctReLU(),
                                   OctConv2d('regular', in_channels=112, out_channels=112, kernel_size=3),
                                   OctReLU(),
                                   OctConv2d('regular', in_channels=112, out_channels=112, kernel_size=3),
                                   OctReLU(),
                                   OctConv2d('regular', in_channels=112, out_channels=112, kernel_size=3),
                                   OctReLU(),
                                   OctConv2d('last', in_channels=112, out_channels=112, kernel_size=3),
                                   nn.ReLU(),
                                         )
        self.full = torch.nn.Sequential(
                                      self.con6,
                                      torch.nn.ReLU(),
                                      self.con7,
                                      torch.nn.ReLU(),
                                      self.con8,
                                      torch.nn.ReLU(),
                                      self.con9,
                                      torch.nn.Sigmoid()
                                )
    # elif data_kind=='kitt':
    #    #...



def forward(self,x0,x1,flag):
   if flag=='train':
       y0 = self.conv1(x0)  # left_patch
       y1 = self.conv1(x1)
       y3 = torch.cat((y0, y1), 1)
       # print(y3.shape)
       y = self.full(y3)