convolution - NAN when I use ReLU activation function in convolutional neural network Lenet-5 -
i did programmed convolution neural network lenet-5. made modifications:
- i replaced activation function of output neurons in last layer rbf softmah.
- subsampling layers maxpooling layers.
- learning method backpropagation
as result, network working correctly.
after tried replace sigmoid output of each neuron in feature maps relu (rectifier linear unit). result, network began learn faster, if not choose low speed, nan value.
for small set of input data, simpler use lower speed of learning. when comes more 1,000 examples, network working, in end nan again.
why there nan when using relu? lenet architecture not relu?
Comments
Post a Comment