tonyreina

import numpy as np
import tensorflow as tf

num_channels = 3

def create_model():
    """
     Generic fully-convolutional (FCN) model
     Model inputs can be defined with size "None" so that they can
      use different shapes each time. This works for FCN models

tonyreina

import numpy as np
import tensorflow as tf
from tensorflow.python.client import timeline
import json

def create_model():
    # Just a simple CNN
    # Alternatively, just
    # model = tf.keras.models.load_model("mymodel.hdf5")

tonyreina

## TO BUILD CONTAINER:
## Make sure you have downloaded the Linux version of OpenVINO: https://software.intel.com/en-us/openvino-toolkit/choose-download/free-download-linux
## Place the downloaded OpenVINO installer in the same directory as this Dockerfile.
## docker build -t 3d_unet_decathlon --build-arg HTTP_PROXY=${HTTP_PROXY} --build-arg HTTPS_PROXY=${HTTPS_PROXY} --build-arg NO_PROXY=${NO_PROXY} --build-arg http_proxy=${http_proxy} --build-arg https_proxy=${https_proxy} --build-arg no_proxy=${no_proxy} .
## TO RUN BUILT CONTAINER:
## For CPU - `docker run -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp.X11-unix  -it 3d_unet_decathlon`
FROM ubuntu:16.04

ENV APP_DIR /app
ADD . ${APP_DIR}

tonyreina

There are several options here. For Keras, just simply load the Keras model and use sess = keras.get_session() to link the TF session variable to the Keras model (model.output and model.input are the output and input ops)

# Option 1: Saved TensorFlow model
tf.saved_model.simple_save(session, export_dir, inputs={"x":x, "y":y}, outputs={"z": z})
mo_tf.py -saved_model_dir export_dir

# Option 2: Saved TensorFlow checkpoint
saver = tf.train.Saver()
saver.save(sess, "./model.ckpt")

tonyreina

source /opt/intel/openvino/bin/setupvars.sh
alias downloader='python3 ${INTEL_OPENVINO_DIR}/deployment_tools/model_downloader/downloader.py'
alias mo='${INTEL_OPENVINO_DIR}/deployment_tools/model_optimizer/mo.py'

tonyreina

FROM intelaipg/openvino-model-server

# Create this working directory in container
# All commands after this are relative to WORKDIR
WORKDIR /usr/openvino

# Copy everything from local machine directory into container
COPY . .

# Install h5py within container

tonyreina

import nibabel as nib
import numpy as np

data = np.ones((32, 32, 15, 100), dtype=np.int16) # dummy data in numpy matrix
img = nib.Nifti1Image(data, np.eye(4))  # Save axis for data (just identity)

img.header.get_xyzt_units()
img.to_filename(os.path.join('build','test4d.nii.gz'))  # Save as NiBabel file

tonyreina

#Prints FPS from from the logs emitted from tf_cnn_bench_6nets scripts.
#Usage: ./print_fps_tf_cnn_bench_6nets.sh 96
echo -e "\n Net BZ FPS \n"

num_cores=`grep -c ^processor /proc/cpuinfo`   # Get number of cores

for network in googlenet inception3 resnet50 resnet152 vgg16 ; do
  for bz in 1 32 64 96 128; do
    fps=$(grep  "total images/sec:"  net_${network}_bz_${bz}_numcores_${num_cores}.log | cut -d ":" -f2 | xargs)
    echo "$network $bz $fps"

tonyreina

#!/bin/bash
#sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'
date > start_benchmark.txt

# Get number of sockets (set interthread to num sockets)
inter=`grep -i "physical id" /proc/cpuinfo | sort -u | wc -l`
# Get number of physical cores (set intra to num physical cores)
num_cores=`grep -c ^processor /proc/cpuinfo`   # Get number of cores
for network in googlenet inception3 resnet50 resnet152 vgg16 ; do
  for bz in 1; do