Single-node Caffe scoring and training based on the intel® Xeon E5 series processor

Original link

The importance of Deep neural network (DNN) applications is increasing in many areas, such as Internet search engines and medical imaging. Pradeep Dubey An overview of Intel in its blog post? Architecture Machine Learning vision. Intel is implementing the machine learning vision outlined in the Pradeep Dubey blog post and is working on developing software solutions to accelerate machine learning workloads. Will these solutions be included in future versions of Intel? Library of Mathematical core functions (Intel? MKL) and Intel? Data analytics Acceleration Library (Intel? DAAL).  This technical preview shows the performance that the Intel platform will expect to achieve with the software we are developing. This version is only available in support of Intel? Advanced Vector Extension Instruction set 2 (Intel? AVX2) running on the processor. In future articles, we will cover the benefits of a distributed multi-node configuration.

The preview package described in this article is limited in functionality and is not designed for production use. The features discussed here are now available in the Intel MKL 2017 Beta and intel® Caffe Branch (fork).

Caffe is a deep learning framework developed by the Berkeley Vision and Learning Center (Berkeley Vision and Learning Center, BVLC) and is one of the most commonly used community frameworks for image recognition. Caffe is typically used as a performance benchmark with AlexNet (an image recognition neural network topology) and ImageNet (a label image database).

Caffe can take full advantage of the math routines optimized in Intel MKL, and will also be able to further improve Intel-based technology by applying code modernization technologies. Xeon? The performance of the processor's system. With the proper use of Intel MKL, Vectorization, and parallelization, optimized solutions are expected to increase training performance by up to 11 times times and improve classification performance by up to 10 times times compared to the Caffe solution that is not optimized.

With these optimizations, the alexnet* network is trained across the ILSVRC-2012 dataset to achieve the top five accuracy in 80% of the time, from 58 days to approximately 5 days.

Begin

We are working to develop new features for software products, and now you can use the Technology Preview package included with this article to reproduce the performance results that are presented and even train AlexNet with your own datasets.

The preview package supports the AlexNet topology and introduces the "Intel_alexnet" model, which is similar to bvlc_alexnet, adding 2 new "Intelpack" and "intelunpack" layers, as well as an optimized convolution, pooling, and normalization layer. In addition, we changed the validation parameters to improve vectorization performance, increased the number of validation minibatch from 50 to 256, and reduced the test iterations from 1000 to 200, so that the amount of images used in the validation run remained the same. The preview package adds the Intel_alexnet model to the following files:

• Models/intel_alexnet/deploy.prototxt
• Models/intel_alexnet/solver.prototxt
• Models/intel_alexnet/train_val.prototxt.

The "intel_alexnet" model allows you to train and test ILSVRC-2012 training sets.

When you start using the preview package, make sure that all of the general Caffe dependencies listed in system requirements and restrictions are installed on the system, and then:

• Unpack the preview package.
• Specify the path for the database, snapshot location, and image mean file in the following "intel_alexnet" model file.
  • Models/intel_alexnet/deploy.prototxt
  • Models/intel_alexnet/solver.prototxt
  • Models/intel_alexnet/train_val.prototxt
• Set up the runtime environment for the software tools listed in the System requirements and Restrictions section.
• Add the./build/lib/libcaffe.so path in the LD_LIBRARY_PATH environment variable
• To set the threading environment:
  $> Export Omp_num_threads=<n_processors * n_cores>
  $> Export Kmp_affinity=compact,granularity=fine
• Use the following command to perform timings on a single node:
  $>./build/tools/caffe time \
  -iterations <number of iterations> \
  --model=models/intel_alexnet/train_val.prototxt
• Use the following command to perform the training on a single node:
  $>./build/tools/caffe train \
  --solver=models/intel_alexnet/solver.prototxt

System Requirements and Limitations

The preview package has the same software dependencies as the Caffe that are not optimized:

• boost* 1.53.0
• opencv* 2.4.9
• protobuf* 3.0.0-beta1
• glog* 0.3.4
• gflags* 2.1.2
• lmdb* 0.9.16
• leveldb* 1.18
• hdf5* 1.8.15
• Red Hat Enterprise Linux * 6.5 or later

and Intel MKL version 11.3 or later.

Hardware compatibility:

• Fourth generation smart Intel? Cool core? Processor (Code Haswell)

This software is only validated using the AlexNet topology and may not be available for other configurations.

Support:

If you have any questions or suggestions about this preview package, please contact: mailto:[email protected].

Single-node Caffe scoring and training based on the intel® Xeon E5 series processor