search

tdeboissiere / DeepLearningImplementations

Implementation of recent Deep Learning papers
MIT License
1.81k stars 652 forks source link

Not receiving Error rate of 7%, for Depth of 40 and k =12 on CIFAR10 #63

Closed nbansal90 closed 6 years ago

nbansal90 commented 6 years ago

Hello,

I just ran the program provided,for a Depth of 40 and Growth rate of 12. Accordingly, the number of dense blocks were kept as 3, Number of filters for 1st convolution layer was kept as 24(2*k) But still after 300 epochs the test accuracy achieved is 90.78% for CIFAR10, far from what is described in the paper.I have not changed any thing else, could you please let me know, if I am missing something here.

tdeboissiere commented 6 years ago

I may have made a mistake when I updated the code to be compatible with keras 2 (esp. if batch norm is involved).

You could try to get back to commit a122bd3fd749645a147e023b5ed7e62aa30927b1 and keras 1.2 and see if that improves things.