DEV Community: T Suzuki

Run Unity ml-agents Examples on Mac

T Suzuki — Sun, 20 Jun 2021 12:24:08 +0000

After watching several videos about how reinforcement learning can be demonstrated using Unity's 3D simulated physics, I tried to set up Unity for the first time. After hours of researching and debugging, I finally was able to compile the code to start learning RL. Since this didn't work out of the box and there was no document that summarized the steps and how to debug etc to make this work (some were partial), I am sharing how you can fix these issues.

Firstly, I followed the steps in this tutorial to download/install Unity and clone the repository from ml-agents:
At this time, you should have copied ml-agents/Project/Assets/ML-Agents folder under $PROJ_DIR/Assets folder.

When you press play at the top, it should succeed but in my case, I faced some compile error such as below:

The type or namespace name 'Input' does not exist in the namespace 'Unity.MLAgents.Extensions'

This means, the extensions are provided from Unity-Technologies/ml-agents repository but this was not copied to the project. So let's copy by going to Window -> Package Manager -> plus icon -> Add package from disk -> Select ml-agents/com.unity.ml-agents.extensions/package.json.

If this still gives you error, load ml-agents/com.unity.ml-agents/package.json and select the correct version according to the release notes.

Then I get an error that IInputActionAssetProvider was not read by the compiler, so once I searched for the code and it looks like there is a macro flag MLA_INPUT_SYSTEM which enables this to be read. This is only available using the Input System of 1.1.0-preview.3 according to this code.

So, I updated Input System's version to 1.1.0-preview.3 in the Package Manager. You cannot preview this version unless you enabled preview in the steps above.

Lastly, I came across this error:

You are trying to read Input using the UnityEngine.Input class, but you have switched active Input handling to Input System package in Player Settings.

This error occurs because there is a new and old way of handling input system packages. (why is that? I don't know) In my case, this was pointing to a new version whereas it should have been pointing to the old version. So I updated that by going into Edit -> Project Settings -> Player -> Other Settings -> Configuration -> Active Input Handling -> Select Input Manager (Old) instead of Input Manager (New). This asks you if you want to restart and so you restart it. The Both setting doesn't get applied upon restart so you want to verify the setting after you restart Unity. If someone can tell me why that is, I'd be appreciated.

This finally compiled and worked!

This is my final Environment State:

ml-agents branch: release_18
Unity ML Agent version: 2.1.0-exp1
Unity ML Agent Extension version: 0.5.0-preview
Unity Input System version: 1.1.0-preview.3
Active Input Handling: Input Manager (Old)

Happy learning!

5 Minute AI Model Deployment on DIY Platforms

T Suzuki — Tue, 20 Aug 2019 00:04:54 +0000

Background

While deploying AI models has often proven to be a reasonable
challenge, DIY platforms are changing the rules of the game with
deployment made both easy and swift, even for the relatively
inexperienced.
For most developers, AI model development seems like a nice touch, but the real challenge begins when they need to create and deploy the models.

Understanding feature engineering
Understanding the mathematics behind machine learning models and choosing the correct models
Designing an interface to access predictive features of AI models
Implementing an infrastructure to serve the AI models on a cloud or self hosted platform

As you can see above, these skills require a team of engineers and DevOps which large corporations can afford to hire.
By democratizing AI building, SuperAI creates an enormous value where anyone with meaningful data is able to generate a model and start integrating it back into their application.

How it’s done

Access SuperAI, an AI platform that simplifies model building and deployment.
Signup and create a project. I selected a use case, "Document Classifier". Document classification is to predict the categories from sentences.
While following the instructions on data preparation, I drag-and-dropped the file to upload. It seems to not work if I don't follow the instructions as it is..
There were some warnings on the data; the platform asks me to choose the target column. The target column is what we want to predict the outcome of. In this case, I want to predict the sentiment so, I chose "sentiment" as the target column and the warning is gone.
Now I click the "Train" tab and I can train a model. I've selected 5 minutes as training time but this can be shorter or longer depending on the data size and complexity.
After 5 minutes, I was able to get a model here at 94.75% accuracy. That seems good enough for my application.
Now, I want to deploy this model so I click "Stage" with "Deploy" checked
After some time, it seems deployed
Now I go to "Predict" tab and test that it's working. I entered a sentence to predict then I was able to get the categorized outcomes with their confidence level.
Then I generated the API key and tested in Terminal to verify the results

Great! Now I can integrate this back into my application
It seems like the tip is to understand your data and format it in a way that they accept it. If the data does not make sense, you would expect a model with bad accuracy. If the data is not formatted correctly, it does not seem to parse well.

Conclusion

DIY AI platforms are changing the rules of the game as far as outputting an AI model relatively good and very quickly.
As long as you understand your data, the rest of the pipeline is quite straightforward. I hope this DIY AI platform improves so that it can handle more use cases and wider range of data formats.

Thanks for reading,