Training on NCS

[idata]

Can the NCS be used to train a model ? If we can train a model then are there already available examples for this ?

[idata]

Hi @darrengreene,

The NCS can't be used for training, its designed to be an inference acceleration device only.

Neal

[idata]

Thanks for the response @neal_at_intel. I am new to using the NCS, so please excuse the basic nature of my questions. Just to clarify, the inability to train is a hardware restriction and not an SDK/API shortcoming right ?

[idata]

Hi @darrengreene,

No worries, to clarify, the hardware could be programmed to do about anything I suppose. But the firmware, the NCSDK and associated API only provide the capability to use the device for its intended purpose which is accelerating inferences.

Neal

[idata]

@neal_at_intel got it. Thanks.

[idata]

On a different note, I have been playing around with some of the already available examples that are part of the NCS SDK. I noticed that with NCAPI v2, we support multiple inputs/graphs using fifo class and 16 or 32 bit FP. The example I have been testing is caffe/googlenet nad that seems to be using NCSAPI v1. Wondering if there are examples already migrated or specifically written with NCSAPI v2 ?

[idata]

Hi @darrengreene

The ncappzoo github repository here: https://github.com/movidius/ncappzoo has an ncsdk2 branch that corresponds with NCSDK2 / NCAPI v2.

Also, if you have existing NCAPI v1 programs, you can make use of the python shim module in that branch to get it running on NCAPI v2 fairly easily. Its available here: https://github.com/movidius/ncappzoo/tree/ncsdk2/ncapi2_shim

This git clone command will set you up on the ncsdk2 branch of the ncappzoo repo:

git clone -b ncsdk2 https://github.com/movidius/ncappzoo.git

[idata]

Thanks @neal_at_intel, just what I was looking for !