Reader of this guide will be introduced to go through two demonstrative examples of image classification which comes with the WML-A system. Self-practice is recommended before the user feel comfortable to adapt the workflow for their own datasets or models.

Common Steps

Network Requirements

The AI-service system,, can be accessed from within HKU campus network (physically connected to HKU campus network, or using SSID “HKU” while on campus. If you need off-campus access, please use HKUVPN2FA service).

Logging onto the server via SSH

You have to use the Secure Shell (SSH) in order to connect to the server for uploading or downloading data. If you use command line interface, you may also download data from the Internet onto the server directly. Please refer to SSH Login and File Transfer Guide for those details.

Web Interface

Point your browser to , and then logon to the system with your HKU portal ID (tmchan only, not

Navigating the File System

User data should be put in /dli/u/tmchan (when your HKU portal ID is tmchan).

(Note: If you also use HPC2021, you will find that it is actually the same storage backing the /dli/ on ai-service and the /lustre1/ over there HPC2021. Files are both viewable/editable by both systems.)

Mode of Operation

For the time being, we are only supporting the “Deep Learning” portion of the system. This mode of operation supports a nearly no-code approach for using the power of deep learning.

Basic Examples of the “Deep Learning Impact” mode of operation

You are strongly recommended to run the following examples in verbatim and get familiar with the basic Web UI before attempting to use your own data or modifying the source code.

  • Flower using TensorFlow (Image Classification with default VGG19)
    • Could be changed to Inception v3, Mobilenet v2, Resnet 50, Resnet 50 v2, Resnet 101, Resnet 101 v2, Resnet 152, Resnet 152 v2, Densenet 121, Densenet 169, Densenet 201
  • CIFAR10 using PyTorch (Image Classification with default Resnet18; Work in Progress)
    • Could be changed to Resnet 18, Resnet 34, Resnet 50, Resnet 101, Resnet 152

Troubleshooting and How to Report Problems

(Work in Progress) If you faced problems/error when you are using the system, you may first check if you are facing known items with quick solutions. At the end of the page there will be instructions on what you can do before sending e-mail to us for technical assistance.

Simple Modifications so that it works better for Your Data

(Work in Progress)

What’s Next in the “Deep Learning Impact” mode of operation?

The ai-service utilize IBM’s Watson Machine Learning Accelerator 1.2.2. Besides the “Basic Examples” above, IBM also provided some other samples at WMLA 1.2.2 samples. Make sure you are NOT using samples from elsewhere as even samples for version 1.2.3 samples could NOT be used without modifications.

If you intend to edit the source code extensively, you may also need to go over the IBM’s description of the TensorFlow and PyTorch training models, and elastic distributed learning order to modify the sample codes appropriately. You may find the exact versions of Python, TensorFlow, Keras, Caffe and PyTorch which are used to support the service here . This information maybe necessary so that you could find the right documentation pages when you are modifying the Python source code.