cant replicated mnist_100_full

CuriousAI / ladder

Ladder network is a deep learning algorithm that combines supervised and unsupervised learning

MIT License

516 stars 142 forks source link

cant replicated mnist_100_full #1

Closed youralien closed 9 years ago

youralien commented 9 years ago

When running the command for mnist_100_full,

# Full
run.py train --encoder-layers 1000-500-250-250-250-10 --decoder-spec sig --denoising-cost-x 1000,10,0.1,0.1,0.1,0.1,0.1 --labeled-samples 100 --unlabeled-samples 60000 --seed 1 -- mnist_100_full

I get 16.7% error, which is nothing close to the 1.13% reported in the paper.

arasmus commented 9 years ago

The problem is most likely in the way Fuel converts MNIST data. It should convert it to float32 to work. If it is in unsigned int format, it doesn't work. Could that be the cause? Some people had this problem.

If you make sure you use Fuel from tag v0.0.1 and use the command specified in the readme, it should work.

youralien commented 9 years ago

Thank you for your quick response! I did some debugging and it indeed was the case that the features were not in the 0 - 1 range. Now, I am getting 1.1% error! :bulb: