The Test Error of my experiment is (1-Test_Acc)= 0.0276 on mnist rotated images database.
...but the paper has put forward that the Test error is 0.0169, for the same dataset.
...
My experiment:
I downloaded the software from this git account a few days ago. I noticed that it will not run under the latest stable python(3)...So I fixed the python version errors by the python official code, 2to3. Then I
executed
python run_mnist.py --combine-train....
as suggested in the readme file.
Any idea on why the code cannot repeat the performance of the cvpr17 paper or how to reach that performance?
Hi
There was no log in the logs directory.
The Test Error of my experiment is (1-Test_Acc)= 0.0276 on mnist rotated images database.
...but the paper has put forward that the Test error is 0.0169, for the same dataset.
...
My experiment:
I downloaded the software from this git account a few days ago. I noticed that it will not run under the latest stable python(3)...So I fixed the python version errors by the python official code, 2to3. Then I executed python run_mnist.py --combine-train.... as suggested in the readme file.
Any idea on why the code cannot repeat the performance of the cvpr17 paper or how to reach that performance?