Last active
March 4, 2020 17:54
-
-
Save dongzhuoyao/9d4f0e75904f47b941bef958080d3220 to your computer and use it in GitHub Desktop.
This file contains 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
| import torch | |
| from torch import nn | |
| from torch.nn import functional as F | |
| import math | |
| class LayerNorm(nn.Module): | |
| def __init__(self, eps=1e-6): | |
| super().__init__() | |
| self.eps = eps | |
| def forward(self, x): | |
| mean = x.mean(-1, keepdim=True) | |
| std = x.std(-1, keepdim=True) | |
| return (x - mean) / (std + self.eps) | |
| class BasicBlock(nn.Module): | |
| def __init__( | |
| self, | |
| in_channels, | |
| inter_channels=None, | |
| sub_sample=False, | |
| bn_layer=True, | |
| dropout=0.2, | |
| out_channels=512, | |
| ): | |
| super(BasicBlock, self).__init__() | |
| self.sub_sample = sub_sample | |
| self.in_channels = in_channels | |
| self.out_channels = out_channels | |
| self.inter_channels = inter_channels | |
| if self.inter_channels is None: | |
| self.inter_channels = in_channels # // 2,TODO | |
| if self.inter_channels == 0: | |
| self.inter_channels = 1 | |
| conv_nd = nn.Linear(self.in_channels, self.inter_channels) | |
| max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2)) | |
| self.g = nn.Linear(self.in_channels, self.inter_channels) | |
| if bn_layer: | |
| self.W = nn.Linear(self.inter_channels, out_channels) | |
| else: | |
| self.W = conv_nd( | |
| in_channels=self.inter_channels, | |
| out_channels=self.in_channels, | |
| kernel_size=1, | |
| stride=1, | |
| padding=0, | |
| ) | |
| nn.init.constant(self.W.weight, 0) | |
| nn.init.constant(self.W.bias, 0) | |
| self.theta = nn.Linear(self.in_channels, self.inter_channels) | |
| self.phi = nn.Linear(self.in_channels, self.inter_channels) | |
| self.dropout = nn.Dropout(dropout) | |
| self.ln = LayerNorm() | |
| if sub_sample: | |
| self.g = nn.Sequential(self.g, max_pool_layer) | |
| self.phi = nn.Sequential(self.phi, max_pool_layer) | |
| def forward(self, query, support): | |
| """ | |
| :param query: (b, q, c), support: (b,s,c) | |
| :return: (b,q,c) | |
| """ | |
| assert query.size(2) == support.size(2) | |
| assert query.size(0) == support.size(0) | |
| g_x = self.g(support) | |
| theta_x = self.theta(query) | |
| phi_x = self.phi(support) | |
| phi_x = phi_x.permute(0, 2, 1) | |
| f = torch.matmul(theta_x, phi_x) # Bxqxs | |
| f = f / math.sqrt(self.in_channels) # rescale | |
| f_div_C = F.softmax(f, dim=-1) | |
| y = torch.matmul(f_div_C, g_x) # BxqxC | |
| y = self.ln(y) # layer normalization in last dim | |
| tmp = F.relu(y) | |
| tmp = self.W(tmp) | |
| W_y = self.dropout(tmp) | |
| z = W_y + query | |
| return z | |
| if __name__ == "__main__": | |
| from torch.autograd import Variable | |
| import torch | |
| sub_sample = True | |
| bn_layer = False |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment