Intel TensorFlow CNN Benchmarking Script

intel_tf_cnn_benchmarks.sh

# !/bin/bash

# Run TensorFlow's tf_cnn benchmarks in tensorflow_p36 virtualenv

# Activate TensorFlow virtual environment
#source activate tensorflow_p36

# Clone benchmark scripts
git clone https://github.com/tensorflow/benchmarks.git   
cd benchmarks/scripts/tf_cnn_benchmarks/  
#rm *.log # remove logs from any previous benchmark runs

# Assign num_cores to the number of physical cores on your machine
num_cores=36

# Set environment variables
export KMP_AFFINITY=granularity=fine,verbose,compact,1,0
export KMP_BLOCKTIME=1
export KMP_SETTINGS=1
export OMP_NUM_THREADS=$num_cores
export OMP_PROC_BIND=true



## Run training benchmark scripts

#networks=( alexnet googlenet inception3 resnet50 resnet152 vgg16 )
networks=( googlenet inception3 resnet50 resnet152 vgg16 )
batch_sizes=( 1 32 64 96 128 )

for network in "${networks[@]}" ; do
  for bs in "${batch_sizes[@]}"; do
    echo -e "\n\n #### Starting $network with batch size = $bs ####\n\n"
    
    # Run Inference benchmark, please note the data format is NCHW is optimized for MKL
    time python tf_cnn_benchmarks.py  --data_format NCHW --model $network --batch_size $bs --num_batches 30 \
                --num_intra_threads $num_cores --num_inter_threads 1 --forward_only=True \
                2>&1 | tee net_"$network"_bs_"$bs".log
                
  done
done

## Print training benchmark throughput

echo -e "\n Network batch_size images/second \n"

for network in "${networks[@]}" ; do
  for bs in "${batch_sizes[@]}"; do
    fps=$(grep  "total images/sec:"  net_"$network"_bs_"$bs".log | cut -d ":" -f2 | xargs)
    echo "$network $bs $fps"
  done
    echo -e "\n"
done

# Deactivate virtual environment
#source deactivate