TensorFlow AWS AMI aka TFAMI

Are you frustrated that many AMIs are not working to get TensorFlow working? Or that you have to waste time setting up your instances?

This is an AMI that I hope people, including myself, would continually monitor and ensure it works all the time. I call it TFAMI! How innovative.

Also, I have included Keras and OpenAI Gym as a bonus! Other packages can be easily installed with pip install package_name if you need anything else.

How to find?

Under community AMI, search for TFAMI.v3 or TFAMI.v2. Please take note that TFAMI.v3 is very new and I am still deploying it across different regions.

TFAMI.v3 (latest AMI)

This works on both p2 and g2 instances but I am working on deploying on all regions.

• Tensorflow 0.12 head
• Keras 1.1.0
• TensorLayer 1.2.7
• CUDA 8.0
• CuDNN 5.1
• Python 2.7
• Ubuntu 16.04

TFAMI.v2 (stable AMI)

If you are running p2 instance, you should use this.

• Tensorflow 0.10.0
• Keras 1.1.0
• CUDA 8.0
• CuDNN 5.1
• Python 2.7
• Ubuntu 16.04

TFAMI (deprecated)

If you are running g2 instance, you should use this.

• TensorFlow 0.8.0
• Keras 1.0.4
• OpenAI Gym
• Python 2.7
• CUDA 7.5
• CuDNN 5
• Ubuntu 14.04

How to install?

If you are a unfamiliar with Amazon AWS GPU instance, I suggest you use this guide that is made for beginners. Trust me, you cannot go wrong with the guide and this AMI! Instead of using the AMI that the website recommends (you will know when you reach that section how to search for an AMI), just search for TFAMI.v3 or TFAMI.v2 instead.

Please take note we have updated TensorFlow and all of its dependencies. The new version is available as TFAMI.v3.

FAQs (Relating to Driver Issues)

1. Error: Failed to initialize NVML: Driver/library version mismatch
   • Run sudo reboot.
2. Error: NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
   • Run sudo apt-get update.
3. Why does p2 instance work with only TFAMI.v2.
   • This is due to a configuration I did when I manually compiled TensorFlow and indicated compute capability 3.5 where g2 instance only has compute capability 3.0.
   • Both g2 instance and p2 instance would work on the official TFAMI.v3.
4. Error: Mismatch of Kernel with DSO
   • Run sudo reboot.

Regions (TFAMI.v3)

• N. Virginia ami-0e969619
• Ohio ami-9cc69cf9
• N. California ami-08451468
• Oregon ami-52bb0c32
• Canada ami-cb2694af
• Singapore ami-2c0bad4f
• Ireland ami-b34566c0
• Frankfurt ami-515c9c3e
• London ami-d4e2e8b0
• Singapore ami-e6862885
• Sydney ami-baa19ad9
• Seoul ami-8fa87ee1
• Tokyo ami-96e982f1
• Mumbai ami-3fcdba50
• Sao Paulo ami-19198175

Regions (TFAMI.v2)

• N. Virginia ami-a96634be
• Ohio ami-73045e16
• N. California ami-813a71e1
• Oregon ami-ac8b2fcc
• Singapore ami-2c0bad4f
• Ireland ami-19d49a6a
• Frankfurt ami-155ca57a
• Tokyo ami-b701a7d6
• Seoul ami-80ec38ee
• Sydney ami-ec201d8f
• Mumbai ami-1b562274
• Sao Paulo ami-0c28b560

Regions (TFAMI)

• N. Virginia ami-d0e4adc7
• N. California ami-5ce2ab3c
• Oregon ami-92af74f2
• Singapore ami-4362c520
• Ireland ami-b8f7b4cb
• Frankfurt ami-d09b65bf
• Tokyo ami-9331eaf2
• Seoul ami-45a3772b
• Sydney ami-be596bdd
• Mumbai ami-fe1a6e91
• Sao Paulo ami-cd2fbda1

TFAMI.v3 Instance Specifications

• 100 GB EBS
• Can be used on any GPU instances including the new p2 instances and the old g2 instances You can easily change your EBS volume with this guide.

TFAMI.v2 Instance Specifications

• 100 GB EBS
• Can be used only with g2 instances due to compute capability 3.5 You can easily change your EBS volume with this guide.

TFAMI Instance Specifications

• 40 GB EBS
• Can be used on any GPU instances including the new p2 instances.

Have a bug to report?

Raise an issue here and we'll update TFAMI to make sure it works or enable it across different regions.

License

MIT