Adelaide Deep Racer Hints and Tips

start: 3Nov2021 9am Adelaide time ends: 10Nov2021 noon

Deep Racer References

Documentation

Algorithms

Reward Graphs

Tuning and Optimization

A short 15min video with some great guidance on training times and reward graph complexity. The results can be surprising. DeepRacer Expert Bootcamp: How much training is enough? https://www.youtube.com/watch?v=qOEqW-xTClE

Applying Credits to Offset Training Costs

Once you request credits, they will be emailed to you. Then you can apply them to your account using this link https://docs.aws.amazon.com/awsaccountbilling/latest/aboutv2/useconsolidatedbilling-credits.html

Getting Help

Slack

When you join the AWS Machine Learning Community on Slack via this link https://aws-ml-community.slack.com/archives/C02JZ7ZKHB6 the AWS race organizers will then add you to the private Slack channel titled ‘deepracer-adelaide’ where you’ll find your fellow racers.

How you can use Slack

We have created a private slack channel for this event and invited all those who have registered. We are also providing support, question answering and keeping you up to date with race results in this channel.

Additionally you can access the global Deep Racer hive mind, over 40k people at the following links:

Welcome to AWS Machine Learning Community Slack, hope you will like it here! Please be respectful to other community members, we’re doing our best to make it a friendly and supportive place. If you have joined this Slack because of an event that you are participating in, make sure you check the instructions on who to message and which channel to join (for Udacity it is #learning-udacity). We are a big group and most members will not be able to help you with the event itself, so it’s best that you follow the instructions to get to the right place 🙂 Here are a few channels that you might want to join: • #introducing-everyone say “Hi” and share a few words about yourself • #random to chat about just about anything • private channel sometimes we share things that we make and we appreciate it when you help us share them with the World • Channels starting with meetup - check if you can connect with members in your area for an in-person event To help your journey with DeepRacer: • #aws-deepracer general DeepRacer talks • #dr-getting-started-with-deepracer is helpful when you have just started your DeepRacer adventure • #dr-training is handy for general training questions (reward function, action space, algorithms, log analysis - you name it) • #dr-training-local how to set up you DeepRacer training outside the AWS DeepRacer Console Generally channels starting with dr- are DeepRacer channels. For learning and more general AWS channels you can go to: • #aws-sagemaker for SageMaker • #aws-deepcomposer for DeepComposer • #learning-ml and #learning-rl-theory for an opportunity to learn with others We hope you will enjoy learning with us!

Accessing the Race

You need to do the following:

  • Have an AWS Account
  • Race link
  • Train a model
  • Submit your trained model to the Leaderboard

Get an AWS Account

You can use an existing AWS account, but ensure you’re not sharing with anyone as you’ll be stuck with a single racer name and may hit simultaneous model training limits. We’ll be providing credits for costs incurred after you work through the 10 hours free Deep Racer credits.

You can also sign up for an AWS account when you click the race link. https://console.aws.amazon.com/deepracer/home#raceToken/JFupLMLcTvW3Q8964wa_Ew You’ll notice a small charge on your credit card when you sign up; this will be returned as part of the merchant process with your bank.

NOTE: We’ll be providing Deep Racer credits in increments of $25. You can request credits from the Slack channel. Note that if you spin up a dozen EC2 instances to do some crypto mining, or other non Deep Racer activity this won’t be covered by credits. We stronly suggest you setup some cost alerts so you can monitor spend. See https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/monitor_estimated_charges_with_cloudwatch.html or https://github.com/tarasowski/aws-billing-alarm-cloudformation-template and don’t forget to confirm your billing alert subscription in your email.

Also please create an admin user, or similar and don’t use the root user in your account. This is not good security practice. Follow the instructions in https://docs.aws.amazon.com/IAM/latest/UserGuide/getting-started_create-admin-group.html to create an admin user which has permission to also access billing information.Your root user credentials (that you created the account with) should be then managed as important secrets.

When you click on the race link you’ll be taken the race entry page in the AWS Console (in your browser). https://console.aws.amazon.com/deepracer/home#raceToken/JFupLMLcTvW3Q8964wa_Ew

Here’s how you access our classic race format https://github.com/awsdocs/aws-deepracer-developer-guide/blob/master/doc_source/deepracer-join-community-race.md

NOTE: You can’t modify your racer name after you’re in the race. To avoid having a weird portmanteau of words and numbers make sure you label yourself with an appropriate cool name as you’re joining.

Managing Costs

NOTE that we are providing AWS credits to cover your Deep Racer expenses. Request those in the Slack channel and we’ll email them out to you.

THere are a few ways you can keep an eye on costs:

Cost Explorer

If you go to the AWS Cost Management link in your account, then filter the cost explorer view for sagemaker service you’ll see costs. Daily view will probably help here. It also shows a total. Filter for the race date range.

From the Command Line

Using https://docs.aws.amazon.com/cli/latest/reference/ce/get-cost-and-usage.html as a reference you can run the following cli query which reads filters.json

CLI

aws ce get-cost-and-usage –time-period Start=2021-11-02,End=2021-11-04 –granularity MONTHLY –metrics “BlendedCost” “UnblendedCost” “UsageQuantity” –group-by Type=DIMENSION,Key=SERVICE Type=TAG,Key=Environment –filter file://filters.json

Note from Understanding your AWS Cost Datasets: A Cheat Sheet (circa Nov 2019) https://aws.amazon.com/blogs/aws-cloud-financial-management/understanding-your-aws-cost-datasets-a-cheat-sheet/

  • Unblended Cost - actual usage costs
  • Blended Cost - Used for multi account billing agregration and not useful here
  • Usage Quantity - only relates to RIs and prepayments where you may not use all of your allocated hours. Not useful here.

filters.json

Add the following to your filters.json file and run a linter over it as this page probably mangles the tabbing. { “Dimensions”: { “Key”: “SERVICE”, “Values”: [ “Amazon SageMaker” ] } }

Continue reading articles in my Amazon Web Services series