(python352) C:\Users\ZQ>python
Python 3.5.2 |Continuum Analytics, Inc.| (default, Jul 5 2016, 11:41:13) [MSC v.1900 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

import torch
print(torch.version)
0.3.1.post2
quit()

(python352) C:\Users\ZQ>python D:\soft_decision_tree\main.py
directory ./data already exists
Traceback (most recent call last):
File "D:\soft_decision_tree\main.py", line 53, in
transforms.Normalize((0.1307,), (0.3081,))
File "D:\Anaconda\envs\python352\lib\site-packages\torchvision\datasets\mnist.py", line 54, in init
os.path.join(self.root, self.processed_folder, self.training_file))
File "D:\Anaconda\envs\python352\lib\site-packages\torch\serialization.py", line 267, in load
return _load(f, map_location, pickle_module)
File "D:\Anaconda\envs\python352\lib\site-packages\torch\serialization.py", line 420, in _load
result = unpickler.load()
AttributeError: Can't get attribute '_rebuild_tensor_v2' on <module 'torch._utils' from 'D:\Anaconda\envs\python352\lib\site-packages\torch\_utils.py'>

Traceback (most recent call last):
File "C:/Users/woai_fish/Desktop/android/soft-decision-tree-master/main.py", line 100, in
model.train_(train_loader, epoch)
File "C:\Users\woai_fish\Desktop\android\soft-decision-tree-master\model.py", line 193, in train_
loss, output = self.cal_loss(data, self.target_onehot)
File "C:\Users\woai_fish\Desktop\android\soft-decision-tree-master\model.py", line 133, in cal_loss
leaf_accumulator = self.root.cal_prob(x, self.path_prob_init)
File "C:\Users\woai_fish\Desktop\android\soft-decision-tree-master\model.py", line 58, in cal_prob
self.prob = self.forward(x) # probability of selecting right node
File "C:\Users\woai_fish\Desktop\android\soft-decision-tree-master\model.py", line 48, in forward
return (F.sigmoid(self.beta * self.fc(x)))
File "E:\Conda\envs\android\lib\site-packages\torch\autograd\variable.py", line 757, in mul
return self.mul(other)
File "E:\Conda\envs\android\lib\site-packages\torch\autograd\variable.py", line 301, in mul
return Mul.apply(self, other)
File "E:\Conda\envs\android\lib\site-packages\torch\autograd_functions\basic_ops.py", line 50, in forward
return a.mul(b)
RuntimeError: inconsistent tensor size at d:\downloads\pytorch-master-1\torch\lib\th\generic/THTensorMath.c:847

Once all epochs are complete, I get this error
save_result() missing 1 required positional argument: 'acc'

Python verison 3.6.3
Torch - 0.3.0.post4

So this seems to be an implementation of the tree as described in the paper, however, how then do we distill the knowledge from a neural network into the tree?

Thanks

Thank you for bothering you again.

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:)

Thank you for your code I was wondering what I should change to use the nn.CrossEntropy loss instead of the loss stated in the paper ?

Thank you

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?

Is that possible to retrieve feature importance, as original decision trees

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%)

I want to visualize the tree to look how it stimulate the NN , just like the picture in img folder. But the source code seems to do not have this function ? what should I do?

where code should I to change?thank u

How to load the model and perform inference

python3 main.py --max-depth 4 --lr 1
...
Train Epoch: 3 [17280/60000 (29%)] Loss: 0.522095, Accuracy: 54/64 (84.0000%)
Traceback (most recent call last):
File "main.py", line 79, in
model.train_(train_loader, epoch)
File "/content/model.py", line 190, in train_
correct += pred.eq(target.data).cpu().sum()
RuntimeError: Expected object of backend CPU but got backend CUDA for argument #2 'other'

This is so stange and I can't fix it.

when initialize the weight in a different range then the default case I get the node probability as NaN did any one experienced that ?