When using reconstruction flag, I get the following error:

net.py:28: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument. c = F.softmax(self.b) net.py:38: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument. c = F.softmax(b_batch.view(-1, output_caps)).view(-1, input_caps, output_caps, 1) Traceback (most recent call last): File "net.py", line 274, in train(epoch) File "net.py", line 230, in train output, probs = model(data, target) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call result = self.forward(*input, **kwargs) File "net.py", line 137, in forward reconstruction = self.reconstruction_net(x, target) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call result = self.forward(*input, **kwargs) File "net.py", line 119, in forward mask.scatter_(1, target.data.view(-1, 1), 1.) RuntimeError: scatter_() received an invalid combination of arguments - got (int, torch.cuda.LongTensor, float), but expected one of:

• (int dim, Variable index, Variable src) didn't match because some of the arguments have invalid types: (int, torch.cuda.LongTensor, float)
• (int dim, Variable index, float value) didn't match because some of the arguments have invalid types: (int, torch.cuda.LongTensor, float)

I have NVIDIA GTX 1060 with 6GB RAM and cpu RAM of 16GB. Getting the following error after 1st epoch:

THCudaCheck FAIL file=/opt/conda/conda-bld/pytorch_1512387374934/work/torch/lib/THC/generic/THCStorage.cu line=58 error=2 : out of memory Traceback (most recent call last): File "net.py", line 275, in test_loss = test() File "net.py", line 260, in test output, probs = model(data) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call result = self.forward(*input, **kwargs) File "net.py", line 101, in forward x = self.digitCaps(x) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call result = self.forward(input, **kwargs) File "net.py", line 62, in forward u_predict = caps_output.matmul(self.weights) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/autograd/variable.py", line 386, in matmul return torch.matmul(self, other) File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/functional.py", line 219, in matmul tensor2_expanded = tensor2.expand((expand_batch_portion + tensor2_exp_size[-2:]))
File "/home/khan/anaconda3/envs/kt/lib/python3.6/site-packages/torch/autograd/variable.py", line 280, in contiguous self.data = self.data.contiguous() RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1512387374934/work/torch/lib/THC/generic/THCStorage.cu:58

Hi, For every routing call, logits b_ij needs to set to 0 as per the paper. In the code, it is initialized only once at the beginning. Or. Am I missing something here?

After the train, I encountered the following problems in the reconstruction_visualization process.

KeyError: 'unexpected key "conv1.weight" in state_dict'

When I run reconstruction_visualization.ipynb on jupyter-notebook, I find a mistake in In [4]. The error is in reconstructionnet = ReconstructionNet(16, 10), when I run it, there will be a mistake as follows ：FileNotFoundError: [Errno 2] No such file or directory: '229_model_dict_3routing_reconstructionTrue.pth' How should I correct it？

1. fix a mismatch between default and actual value of lr param.
2. some changes to default values for low memory.
3. compatibility change from .data[0] to item() for newer PyTorch versions.

In inference time , why use the label to reconstruct the image while not use the prediction of the model and the digit capsules as the input of reconstruction module? If the predict is wrong, the reconstruction module will reconstruct the wrong class of image?

/home/raul/Desktop/CapsNet-pytorch/net.py:28: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument. c = F.softmax(self.b) /home/raul/Desktop/CapsNet-pytorch/net.py:38: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument. c = F.softmax(b_batch.view(-1, output_caps)).view(-1, input_caps, output_caps, 1) Traceback (most recent call last): File "/home/raul/Desktop/CapsNet-pytorch/net.py", line 275, in train(epoch) File "/home/raul/Desktop/CapsNet-pytorch/net.py", line 244, in train 100. * batch_idx / len(train_loader), loss.data[0])) IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number

I have a question: according to the paper, the squash function only be used after the sum of prediction u-hat? and in this code, there is a squash after the primary capsule. I got really confused. class PrimaryCapsLayer(nn.Module): def __init__(self, input_channels, output_caps, output_dim, kernel_size, stride): super(PrimaryCapsLayer, self).__init__() self.conv = nn.Conv2d(input_channels, output_caps * output_dim, kernel_size=kernel_size, stride=stride) # input_channels = 256,output_caps = 32, output_dim = 8, kernel_size = 9, stride = 2 self.input_channels = input_channels self.output_caps = output_caps self.output_dim = output_dim def forward(self, input): out = self.conv(input) N, C, H, W = out.size() out = out.view(N, self.output_caps, self.output_dim, H, W) # will output N x OUT_CAPS x OUT_DIM out = out.permute(0, 1, 3, 4, 2).contiguous() out = out.view(out.size(0), -1, out.size(4)) out = squash(out) #####QUESTION?? return out