Pyramid-Attention-Networks-pytorch
Pyramid-Attention-Networks-pytorch copied to clipboard
JaveyWang
Implementation of Pyramid Attention Networks for Semantic Segmentation.
In networks.py, lines 123 - 124: ```python x3_upsample = self.relu(self.bn_upsample_3(self.conv_upsample_3(x3_2))) x2_merge = self.relu(x2_2 + x3_upsample) ``` I know that x2_2 has a linear activation, why does x3_upsample have a relu,...
I run another dataset, image size(3, 320, 640) and get this error: in FPA forward(self, x) 131 print(x2_2.shape) 132 print(x3_upsample.shape) --> 133 x2_merge = self.relu(x2_2 + x3_upsample) 134 x2_upsample =...
Thanks for your work! I removed the classification module and its related loss and the performance is about 77%. I wonder if you have done the experiments with classification loss...
self.conv_upsample_2 = nn.ConvTranspose2d(channels_mid, channels_mid, kernel_size=4, stride=2, padding=1, bias=False) 改为： self.conv_upsample_2 = nn.ConvTranspose2d(channels_mid, channels_mid, kernel_size=3, stride=2, padding=1, bias=False) 这样修改正确吗？
could you share the cls_labels.npy? thank you
请问，FPA模块中的Global Pooling那个分支最后是否有upsample。论文的Figure3(b)中是有upsample这一模块的，而作者您的实现是直接将1x1的结果加入到最后的预测结果中。我想知道到底这个全局池化的结果是直接加还是要上采样至原图大小再与预测结果相加。谢谢~
The file cls_labels.npy saves the class label of each image, format like {'filename' : '1, 0, 1, ..., 0'}. For simplicity, you can use the label from the x_mask, I...
image size(3,512,512) mask_pred = self.mask_classifier(out_ss) print(mask_pred.shape) #torch.Size([16, 5, 128, 128]) target mask 512, predict mask ony 128 self.gau_block1 = GAU(channels_blocks[0], channels_blocks[1], upsample=False) self.gau_block2 = GAU(channels_blocks[1], channels_blocks[2]) self.gau_block3 = GAU(channels_blocks[2], channels_blocks[3])...
In line 216 of networks.py, should you use += instead of = ?
Hello sir , I didn't found the Color_Classifier and loading of checkpoints while doing the eval
