Skip to content

Latest commit

 

History

History
85 lines (62 loc) · 3.75 KB

README.md

File metadata and controls

85 lines (62 loc) · 3.75 KB
Raw

How to use ngraph

Simple! Just add import ngraph_bridge after building it

The simplest hello-world example can be found in axpy.py. For real world examples checkout the instructions below to run tf_cnn_benchmarks and models from Tensorflow Hub

tf_cnn_benchmarks: High performance benchmarks

tf_cnn_benchmarks contains implementations of several popular convolutional models, and is designed to be as fast as possible. tf_cnn_benchmarks supports both running on a single machine or running in distributed mode across multiple hosts. See the High-Performance models guide for more information.

These models utilize many of the strategies in the TensorFlow Performance Guide. Benchmark results can be found here.

These models are designed for performance. For models that have clean and easy-to-read implementations, see the TensorFlow Official Models.

Running examples from tf_cnn_benchmarks

Use the following instructions

git clone https://github.com/tensorflow/benchmarks.git
git checkout 4c7b09ad87bbfc4b1f89650bcee40b3fc5e7dfed
cd benchmarks/scripts/tf_cnn_benchmarks/

Next enable nGraph by editing the convnet_builder.py by adding import ngraph_bridge right after the import tensorflow line.

Train for a few iterations:

KMP_BLOCKTIME=0  OMP_NUM_THREADS=56 KMP_AFFINITY=granularity=fine,compact,1,0 \
    python tf_cnn_benchmarks.py --data_format NCHW \
    --num_inter_threads 2 --train_dir=./modelsavepath/ \
    --model=resnet50 --num_batches 10 --batch_size=128

Evaluate the model (Inference pass):

KMP_BLOCKTIME=0  OMP_NUM_THREADS=56 KMP_AFFINITY=granularity=fine,compact,1,0 \
    python tf_cnn_benchmarks.py --data_format NCHW \
    --num_inter_threads 1 --train_dir=$(pwd)/modelsavepath \
    --eval --model=resnet50 --batch_size=128```

Tips

  • Use --data_format NCHW to get better performance. Avoid NHWC if possible.
  • Change the batch_size to 128 for batch inference performance and batch_size=1 for real time inference
  • Change the --model flag to test for different topologies. The (by no means non-exhaustive) list of models known to run as of now are:
  1. vgg11
  2. vgg16
  3. vgg19
  4. lenet
  5. googlenet
  6. overfeat
  7. alexnet
  8. trivial
  9. inception3
  10. inception4
  11. resnet50
  12. resnet50_v1.5
  13. resnet50_v2
  14. resnet101
  15. resnet101_v2
  16. resnet152
  17. resnet152_v2
  18. mobilenet

Please feel free to run more models and let us know if you run across any issues.

  • A more involved example of the run command KMP_BLOCKTIME=0 OMP_NUM_THREADS=28 KMP_AFFINITY=granularity=fine,proclist=[0-27] python tf_cnn_benchmarks.py --model=resnet50 --eval --num_inter_threads=1 --batch_size=128 --train_dir /nfs/fm/disks/aipg_trained_dataset/ngraph_tensorflow/partially_trained/resnet50 --data_format NCHW --num_epochs=1 --data_name=imagenet --data_dir /mnt/data/TF_ImageNet_latest/ --datasets_use_prefetch=False

Tensorflow Hub:

Tensorflow Hub models should also work. For example, you can try out network retraining by following instructions from here on, lets say, inceptionv3

Keras models:

Keras (with Tensorflow backend) too should also work out of the box with ngraph, once one adds import ngraph_bridge to the script. Here is an example.