Use Pre-Trained Tensorflow Models with a GPU in SAP AI Core

[gianfranco-s]

Tutorials: https://developers.sap.com/tutorials/ai-core-tensorflow-byod.html

I'm using the Free Tier. I'm fairly certain that this tutorial cannot be completed with GPU nodes The tutorial can be completed without GPU nodes, although ATM I don't have any way to try it with GPU nodes.

I haven't been able to complete this tutorial because there are multiple issues. I have solved all the issues with the tutorial. I'll fork the project, and create a PR.

The first issue appears when trying to register the artifact. It is my understanding (after having tried it multiple times) that the application must be created before registering the model.

Here are my proposed changes

Order	Current	Proposed
1	Step 1	Step 1
2	Step 2	Step 2
3	Step 3	Step 3
4	Step 4	Step 6
5	Step 5	Step 7
6	Step 6	Step 4
7	Step 7	Step 5
8	Step 8	Step 8
9	Step 9	Step 9

Additional comments on current steps

• Step 6 says to set resourcePlan to infer.s which will enable GPU. Free tier only allows for starter resourcePlan. Maybe add this to the prerequisites?
• Step 7 has a typo: executables.ai.sap.com/name: "serve-executuable" -> executables.ai.sap.com/name: "serve-executable".

[gianfranco-s]

• Step 5: command docker build -t ends in error: ERROR: failed to solve: process "/bin/sh -c pip3 install -r requirements.txt" did not complete successfully: exit code: 1

I suggest leaving requirements.txt without versions

# requirements.txt
scikit-learn
Flask
gunicorn

And using flag '--ignore-installed' in dockerfile to avoid conflicts with base image packages

RUN pip3 install --ignore-installed -r requirements.txt

[gianfranco-s]

• Step 3: Error: Could not connect to the endpoint URL

Attribute endpoint is the problem. Should be s3. instead of s3-. Source

• Original: endpoint = "s3-eu-central-1.amazonaws.com", # Change this
• Correct: endpoint = "s3.eu-central-1.amazonaws.com", # Change this

[gianfranco-s]

• Step 5: using latest tensor flow Docker image results in conflict while loading model.h5.

I suggest changing the Dockerfile to this:

FROM tensorflow/tensorflow:2.10.0-gpu-jupyter

This will require to import pad_sequences from keras.utils (instead of keras.preprocessing.sequence)