Thank you for sharing this code I was wondering why did you comment that forward section in soft decision tree and why you are not using the 0.5 test for left or right nodes ??

just another question output dimension in args is the number of classes right?

best regards

Really great implementation! I have a question about the implementation. In the last paragraph of section of regularizers, the authors mention ' exponential increase in the temporal scale of the window used to compute the running average'. Is this feature implemented in this codebase? I didn't find it. Thanks:)

I try to use my dataset, my data is a table with many discrete data, such as 0,1,2. I found the loss is nan

Train Epoch: 1 [0/49626 (0%)]	Loss: nan, Accuracy: 28/64 (43.0000%)
Train Epoch: 1 [640/49626 (1%)]	Loss: nan, Accuracy: 30/64 (46.0000%)
Train Epoch: 1 [1280/49626 (3%)]	Loss: nan, Accuracy: 36/64 (56.0000%)
Train Epoch: 1 [1920/49626 (4%)]	Loss: nan, Accuracy: 35/64 (54.0000%)
Train Epoch: 1 [2560/49626 (5%)]	Loss: nan, Accuracy: 40/64 (62.0000%)
Train Epoch: 1 [3200/49626 (6%)]	Loss: nan, Accuracy: 29/64 (45.0000%)
Train Epoch: 1 [3840/49626 (8%)]	Loss: nan, Accuracy: 33/64 (51.0000%)
Train Epoch: 1 [4480/49626 (9%)]	Loss: nan, Accuracy: 28/64 (43.0000%)
Train Epoch: 1 [5120/49626 (10%)]	Loss: nan, Accuracy: 30/64 (46.0000%)
Train Epoch: 1 [5760/49626 (12%)]	Loss: nan, Accuracy: 28/64 (43.0000%)
Train Epoch: 1 [6400/49626 (13%)]	Loss: nan, Accuracy: 38/64 (59.0000%)
Train Epoch: 1 [7040/49626 (14%)]	Loss: nan, Accuracy: 36/64 (56.0000%)
Train Epoch: 1 [7680/49626 (15%)]	Loss: nan, Accuracy: 27/64 (42.0000%)
Train Epoch: 1 [8320/49626 (17%)]	Loss: nan, Accuracy: 29/64 (45.0000%)
Train Epoch: 1 [8960/49626 (18%)]	Loss: nan, Accuracy: 31/64 (48.0000%)
Train Epoch: 1 [9600/49626 (19%)]	Loss: nan, Accuracy: 32/64 (50.0000%)
Train Epoch: 1 [10240/49626 (21%)]	Loss: nan, Accuracy: 34/64 (53.0000%)
Train Epoch: 1 [10880/49626 (22%)]	Loss: nan, Accuracy: 34/64 (53.0000%)
Train Epoch: 1 [11520/49626 (23%)]	Loss: nan, Accuracy: 29/64 (45.0000%)
Train Epoch: 1 [12160/49626 (24%)]	Loss: nan, Accuracy: 34/64 (53.0000%)
Train Epoch: 1 [12800/49626 (26%)]	Loss: nan, Accuracy: 39/64 (60.0000%)
Train Epoch: 1 [13440/49626 (27%)]	Loss: nan, Accuracy: 34/64 (53.0000%)

Hi, I don't want to classify images, just 16-element tensors per datum. Will this work out of the box with that? If not, where should I look to change?

First of all, thank you for such a good code. I want to ask, when I input a larger size, such as 224 * 224 * 3, I find that the training has no effect, is it necessary to change some parts of the code?