I am trying to test my network on the NCS. It's taking too long.
Was wondering do you have any tips / tools to fasten the inference pass.
One way to do it would be to sparsify filter coefficients.
Do you have any tools to do that?
Thanks
链接已复制
2 回复数
@schil The NCSDK doesn't support sparsity at the moment and as such we don't have any sparsity related tools. You can use the NCS profiler included in the NCSDK in the meantime to check various information (bandwidth, time through each layer, etc) regarding the forward pass through the network.
Thanks @Tome_at_Intel
One more question - Is the MFLOPS reported Mega Floating Point Operations Per Second?
I read in one of the other discussions that NCS supports only half precision (float16_t) datatype - input, output and all coefficients would have to be in half precision.
When MFLOPS is reported, are the numbers modified to reflect that change? since "float" by convention would be full precision (float32_t)
