DagNN Why adding convolution layer reduces the intensity of the output values?

1 view (last 30 days)
h612
h612 on 10 Oct 2017
I'm trying to estimate density map using Deep learning Architecture with DagNN (MatConvnet). I've seen output abruptly reduce in intensity after passing the convolution layer. I tried using LRN normalization to enhance the output- however the running error and objective resulted in 'NaN' or 'Inf'.
The conv layer is defined as:
net.addLayer('conv5', dagnn.Conv('size', [1,1,8,1], 'hasBias', true, 'stride', [1, 1], 'pad', [1 1 1 1]), {'lrn4'}, {'prediction'}, {'conv5f' 'conv5b'});
net.params(9).value= 0.1*scal*randn(1,1,8,1, 'single');
net.params(10).value= 0.001*init_bias*ones(1, 1, 'single');%'biases',
net.params(9).learningRate=1;net.params(9).weightDecay=1;
net.params(10).learningRate=2;net.params(10).weightDecay=0;

Answers (0)

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!