I confirmed that currently CLDNN does not support direct model importing for CAFFE, tensorflow, and so on referring to link.
Would you guide me a right way to convert the training model of tensorflow to a runnable model in CLDNN step-by-step?
I went over the example code for alexnet. It needs the blob file (alexnet.blobs). How can I produce such file based on my training model of tensorflow?
For your information, I am going to use R-CAR H3 with IMG PowerVR Series6XT GX6650.
Thank you in advance.
Hi,
Hmmm, this is not a trivial question. In the future we will expose an Intermediate Representation file exported from Tensorflow, Caffe2, etc. that will be loaded directly in the API. Sadly, we do not have this capability in the current release of clDNN. We will not have a proper cross-network solution until our next release of imgDNN SDK.
AlexNet is easy because they supply the binary file with labels, weights and offsets.
Exporting from Tensorflow is somehow convoluted. It has to be done using protobuf (Google’s Protocol Buffers). There might be other mechanisms like using the Tensorflow’s Saver or, maybe, ONNX.
You will need to understand these formats to be able to feed the required data to clDNN.
Regards.
Carlos.
