Distributed Deep Learning With Spark on AWS

Training deep networks is a time-consuming process, with networks for object recognition often requiring multiple days to train. For this reason, leveraging the resources of a cluster to speed up training is an important area of work. In this talk we'll show how to use an AWS Spark cluster to train a model quickly from a laptop at a very little cost.

Transcript

1. DISTRIBUTED DEEP LEARNING CAFFE ON SPARK ON AWS Vincent Van Steenbergen - @nsteenv
2. WHOAMI VINCENT VAN STEENBERGEN Playing with Scala, Akka & Spark +/- 3 years Deeply interested in Artificial Intelligence and Data Analysis
3. DISCLAIMER
4. DEEP LEARNING aka. Convolutional Neural Networks (convNet)
5. APPLICATIONS
6. IMAGE ANALYSIS
7. IMAGE GENERATION
8. GAMES
9. TRAINING A MODEL REQUIRES: 1. a lot of time (usually weeks/months) 2. a lot of computing power Ex: AlphaGo - 1202 CPU and 176 GPU - 6 weeks training
10. SO HOW CAN I DO THAT... from my laptop? for a decent cost? within a short timespan?
11. YES YOU CAN! Technicaly possible on a (high end) laptop but very slow Solution: distribute training over a cluster
12. APACHE SPARK Pool ressources from all the spark slaves on the cluser
13. AMAZON WEB SERVICES (EC2) GPU instances (g2.2xlarge, g2.8xlarge) Spot instances (on demand, generally 2-3 times cheaper than regular instances)
14. G2.8XLARGE CONFIGURATION Four NVIDIA GRID GPUs, each with 1,536 CUDA cores and 4 GB of video memory 32 vCPUs 60 GiB of memory 240 GB (2 x 120) of SSD storage Average price: $1.00 per hour
15. DEEP LEARNING FRAMEWORKS TensorFlow (Google) Caffe (Berkeley) Torch (Facebook/Deepmind)
16. CAFFE ON SPARK Distribute Caffe on a Spark cluster Developed/maintained by Yahoo (Flickr)
17. ADVANTAGES Can run on an existing cluster along other Spark jobs Leverage existing Caffe models Use SQL, DataFrames, existing LMDB files Peer-to peer communication with Message passing
18. NOT BAD... ... let's give it a go!
19. MNIST Classifying handwritten digits
21. UNDER THE HOOD
22. RESULTS
23. THANK YOU! Any questions? My email: v.vansteenbergen@gmail.com
24. RESSOURCES Sample launch scripts Caffe on Spark Yahoo blog SparkNet