memory leak with vaviable input size

[coolbei]

I have a full convolutional network. when I input variable data into the net, memory keeps growing.
when I modify the souce code, to use origin caffe convolutional layer instead of mkldnn conv layer, the memory did not grow.
any ideas?