CNN model on cifar10 dataset don't seems to be trained

Question

I just wanted to take a look for a simple tutorial on cnn and I want to train my network composed with 5 filters(3 conv layers and 2 dense layers) followed each by max_pool layers and dropout layers.

My problem is that my network don't seems to be trained at all, and I cant find where is the problem....

I'll put all my codes (train.py test.py model.py and main.py) in the following link on github.

https://github.com/mmecoco/tensorflow_cnn_cifar10

Help me...

Answer 1

I just found my problem, this was caused from my last fc layer, the problem was that i tryed to apply a softmax activation on my output, and if I use a softmax cross entropy loss function, the value of my loss will be on the side of -inf or inf. Because of this, the networks stops the learning process and this is why I failed my training.