Can DAAL be used with Deep Learning Frameworks?

[RAtlu1]

Admins

Can Intel DAAL be used with Caffe/Torch or theano? That way it is much easier to experiment for our existing code based on GPU.

Does Intel DAAL come with any prebuilt classifiers to test?

Can Intel DAAL be used to also speed up the process of inference using DAAL built classifiers on end/edge devices based on Intel IA, for example on say Intel NUC? Say image processing inference?

Thanking you

[VictoriyaS_F_Intel]

Hello Sahil,

Thank you for your interest. Here are the answers:

Can Intel DAAL be used with Caffe/Torch or theano?

Intel DAAL has a flexible API that allows to make integration to deep learning frameworks on different levels. It is possible, for example, to replace the implementation of a particular layer of neural network by DAAL implementation, or to replace the group of layers, or to feed the model trained with Caffe, Theano or Torch as an input to scoring stage of neural network implemented with DAAL.

But at the moment we do not provide DAAL-enabled versions of Caffe, Theano or Torch.

You could also use already optimized Caffe and Theano for Intel architectures that use Intel Math Kernel Library DNN extensions underneath. For more information please follow the link: https://software.intel.com/en-us/machine-learning/deep-learning#

Does Intel DAAL come with any prebuilt classifiers to test?

Could you please clarify what you mean by “prebuilt classifiers?

If you mean some pre-trained models of deep neural networks, then the answer is currently “no”. We do not provide pre-trained models in the package. But we are planning to extend the number of neural network samples provided with DAAL in the future releases. And it will be good to know the list of topologies you are interested in.

Can Intel DAAL be used to also speed up the process of inference using DAAL built classifiers on end/edge devices based on Intel IA,

DAAL supports Intel architectures starting from SSE2. And the intension of the overall library design was to improve data analytics on the wide range of the devices, from edge devices to clusters. That’s why DAAL should fit well for doing inference/prediction on edge devices.

Best regards,

Victoriya

[RAtlu1]

Dear Victoriya

Thanks for the reply. I was interested in running a pre-trained model in say Intel Edison to detect and classify objects at say atleast 30 fps on a 640 x 480 video.

If its too much to ask probably I can move to a core m tv stick or an NUC. I will try to test out DAAL in Edison (i wonder how much of SSE2 optimization can help it increase the fps).

Do you have any SSE2 inference benchmarks you can share? That would be of tremendous help in select a right edge device.

Thanking you

[VictoriyaS_F_Intel]

Sahil,

We prepare DAAL 2017 Gold release in which we plan to include the example that shows how to run neural network based prediction with the pre-trained model.

We also consider adding samples in future versions of the library that would demonstrate how to train topologies such as AlexNet, GoogLeNet, etc and use them for prediction as well.

Although we support CPUs starting with SSE2, we encourage customers to migrate to the latest hardware such IA with Intel® AVX2 extensions to get the best performance.

Best regards,

Victoriya

[RAtlu1]

Dear Victoriya

Thank you. One last question. You mean to say even on the edge devices for neural network based prediction with the pre-trained model we still may require a cpu with AVX2? (ofcourse for decent performance)

Those will AVX2 will be quite hard to deploy as edge devices, given the TDP requirements.

Is there a way in future you could utilize not just SSE2 as well as latest SSE 4.2 but also opencl with gen9 EUs unit slices?

I hate to say this but our competitors XYIDIA (name changed) are able to delivery 1 Teraflops performance under 3 watts. And their neural network prediction with a pre-trained models achieves 30fps (real-time) on those edge devices.

I wish we collectively do something in that direction.