Hi, I compared your implementation with the original DGCNN and I find out that, in yours the function "get_graph_feature" is only used before the first convolution operation, instead of using it before every convolution layer.
I am wondering if there is some reason behind this difference.
|
def forward(self, xyz): |
|
|
|
xyz = xyz.permute(0, 2, 1).contiguous() # (B, 3, N) |
|
|
|
batch_size, num_dims, num_points = xyz.size() |
|
x = get_graph_feature(xyz, self.features, self.neighboursnum) # (B, C, N, n) |
|
|
|
x = F.relu(self.bn1(self.conv1(x))) |
|
x1 = x.max(dim=-1, keepdim=True)[0] |
|
|
|
x = F.relu(self.bn2(self.conv2(x))) |
|
x2 = x.max(dim=-1, keepdim=True)[0] |
|
|
|
x = F.relu(self.bn3(self.conv3(x))) |
|
x3 = x.max(dim=-1, keepdim=True)[0] |
|
|
|
x = F.relu(self.bn4(self.conv4(x))) |
|
x4 = x.max(dim=-1, keepdim=True)[0] |
|
|
|
x = torch.cat((x1, x2, x3, x4), dim=1) |
|
x_node = x.squeeze(-1) |
|
|
|
x_edge = F.relu(self.bn5(self.conv5(x))).view(batch_size, -1, num_points) |
|
# if torch.sum(torch.isnan(x_edge)): |
|
# print('discover nan value') |
|
return x_node, x_edge |
