1. Install TensorFlow core 2.9

Docker within LinuxONE

List of most recent IBM z containers

Usage Notes

Exposed Ports: 8888

To start container from image & start an application in production mode:

docker run --name tensorflow -it icr.io/ibmz/tensorflow:[version] bash

Prouction mode with exposed ports:

docker run --name tensorflow -it -p 8888:8888 icr.io/ibmz/tensorflow:[version]

Building from Source

Procedure is delineated here

2. Install zDNN >= 0.4.0

Under Linux:

apt-get update && apt-get install -y libzdnn-dev

Or build from source following the official zDNN repo

3. Install IBM-zDNN-Plugin

Using Python 3:

python3 -m pip3 install ibm-zdnn-plugin

Which will install the IBM zDNN plugin into your local Python 3 environment. If you're using a virtual Python env, install it there accordingly.

Example Models

Source document

• BERT
• Biomedical Image Segmentation
• Credit Card Fraud
• DenseNet121
• DenseNet169
• DenseNet201
• InceptionResNet
• InceptionV3
• NASNetLarge
• NMT
• ResNet101
• ResNet152
• ResNet50
• VGG16
• VGG19
• Xception
• YOLOV3
• YOLOv4

4. GitHub Example with Fashion MNIST

IBM-zDNN-Plugin supports inferencing workloads only. Meaning, the accelerator is only active in the latter part of the example.

Preamble

Once both TensorFlow Core, zDNN, and the IBM-zDNN-Plugin are installed, clone the repo here using git under Z:

git clone https://github.com/IBM/ibm-zdnn-plugin.git

Navigate into the samples/fashion-mnist directory

cd ./samples/fashion-mnist

Training and Running

First, train and save the model to disk with the fashion_mnist_training.py script. This will download the fashion MNIST data set and create a model in the current directory.

Training will take some time. The epoch number in the output will indicate progress.

python3 fashion_mnist_training.py

Once the model has been trained, run the fashion_mnist.py script to run inference against the model.

python3 fashion_mnist.py

The script will report a prediction for some sample images.