How can I use NCS to convert and implement my own caffe layers ?
Or is there a way to use host's CPU to execute the layers ?
@cuihao1993 Thanks for your interest. Currently there isn't support for custom layers in the current NCSDK (v 2.05) and we cannot provide a road map or eta for this feature right now.
@cuihao1993 Maybe you could achieve your goal by using mvNCCompile and the -in and -on options. You could compile a model with
-in input and then use
-on <output node> to specify the output node. This would result in compiling a graph file that takes input from the input node and its output node would be the specified output node. So instead of processing the data through the entire model, you can get the output of the intermediate results of whatever node you choose.
You can view the nodes in your model using the mvNCProfile tool. More information on the mvNCProfile tool can be found at: https://movidius.github.io/ncsdk/tools/profile.html
hello, @Tome_at_Intel .
And I have another question , in
Regular Convolution - 1x1s1, 3x3s1, 5x5s1, 7x7s1, 7x7s2, 7x7s4
Does this mean that 3*3s2 and 5x5s2 is not supported?
I use mvNCProfile to get the runtime of a deploy.prototxt which contains conv 5x5s2 conv 3x3s2 successful.
So I think it support 3×3s2 and 5x5s2 . Am I correct?
Thank you !