GPUs are very essential to work with text (LSTM) and image (ConvNet) related machine learning tasks. GPU are powerful and hence very costly. For purpose of learning one can rent GPU from Amazon web services (AWS).
I also use GPU from AWS to provide you with rich experimental content on Convnet and LSTM. AWS GPU-based instances provide access to NVIDIA GPUs with several thousands of compute cores (generally 1000 to 8000 cores each GPU)
Following GPU instances are available on AWS, that on can go for to run deeper/Bigger LSTM or ConvNet netwerks.
1) P2 Instances
Having NVIDIA Tesla K80 GPUs and supports General Purpose GPU computing using CUDA or OpenCL programming models.
Single and double precison floating point capabilities
12 GB RAM per GPU
Enhanced Network with Elastic Network Adapter.
EBS-optimized
idead choice for graph databases, High performance databases,deep learning and computational molecular and fluid dynamics.
2) G2 Instances
NVIDIA GRID K520 GPUs
support NVIDIA’s fast capture and encode API operations
ideal for 3D visualizations, streaming graphics-intensive applications
3) CG1 Instances
NVIDIA Tesla M2050 GPUs
Single and double precison floating point capabilities
Enhanced Network with Elastic Network Adapter.
EBS-optimized
error-correcting code (ECC) memory
ideal for high performance computing (HPC)
For experimentation I always use G2.8xlarge instance, It comes with Intel Xeon E5-2670 Processors and NVIDIA GPUs with 1,500+ CUDA cores and 4GB of video memory to each card.
Tricks with AWS (obviously tom save time & money!!)
Torch [AMI Id: ami-b36981d8] and Theano [AMI Id: ami-8827efe0] AMIs are present in AWS AMI page, use these Amazon Machine Images (AMI) for quicker turnaround.
Use GPU spot instances - they’re much cheaper.
you may have look at spot instance pricing at AWS console.
$0.12/hr often gets you a g2.xlarge instance, which you may use continuously for days.
If you use spot instance, program in a resume-able manner.
You may setup your custom ubuntu/amazon AMI for machine learning by following instructions given in below link : https://medium.com/towards-data-science/how-to-set-up-a-deep-learning-environment-on-aws-with-keras-theano-b0f39e3d861c.
I am Not going much deeper as hundred of nice tutorial for setting up GPU instance on AWS are already available. I don't post for things which are already available in best form. You may go to this link for further details regarding this topic.
