I try to converte mxnet_model to caffe_model,and use tensorrt to parse caffe_model,but have some problem now.
In mxnet,the output shape from pooling_layer or conv_layer is: floor((height+2padding[0]-dilation[0](kernel_size[0]-1)-1)/stride[0])+1 , but In caffe ，it’s ceil(…)
;
such as,The input shape is （108108）,convolution is(k:3,p:1,s:2）,so the the output_shape from mxnet is floor( (108+21-3)/2 +1 )=floor(54.5) = 54, the output_shape from caffe is 55,
if I set pad=0 manually in caffe model,the output_shape is same with mxnet but the result is not same,because it changed the field of perception.