COMPARING TWO OPEN SOURCE DEEP LEARNING PLATFORMS THAT COME IN HANDY WHEN BUILDING DEEP LEARNING MODELS.
DEEP LEARNING IS THE NEW TREND .THE TECH GIANTS ARE INVESTING BIG . THIS HAS LED TO AN UPSURGE IN PEOPLE WANTING TO LEARN DEEP LEARNING TO STAY INDUSTRY RELEVANT . THERE ARE MANY DEEP LEARNING FRAMEWORKS TO WORK WITH . OF ALL , THREE PARTICULAR FRAMEWORKS HAVE GAINED IMMENSE IMPORTANCE SINCE THEIR RELEASE . THEY ARE PYTORCH , TENSORFLOW AND KERAS .IN THIS POST WE COMPARE KERAS AND TENSORFLOW . IF YOU WANT A COMPARISION BETWEEN PYTORCH AND TENSORFLOW CLICK HERE.
LETS START WITH A BRIEF INTRO TO KERAS AND TENSORFLOW .
NEURAL NETWORKS ARE BUILT USING KERAS AND TENSORFLOW FRAMEWORKS. TENSORFLOW IS AN OPEN SOURCE PLATFORM FOR MACHINE LEARNING . ITS WRITTEN IN PYTHON ,C++ AND CUDA . IT WAS DEVELOPED BY GOOGLE BRAIN TEAM AND WAS FIRST RELEASED ON 9TH SEPTEMBER 2015. KERAS IS AN OPEN SOURCE NEURAL NETWORK WRITTEN ON TOP OF TENSORFLOW . IT IS WRITTEN IN PYTHON . NOW LETS COMPARE THEM IN DETAIL!!
THE COMMUNITY
TENSORFLOW GITHUB REPOSITORY HAS GOT HUGE NUMBER OF COMMITS AND FORKS .THIS SHOWS HOW WIDELY IT HAS SPREAD ACROSS THE WORLD AMONG THE DEEP LEARNING ENTHUSIASTS . SO IT IS HIGHLY LIKELY THAT IF YOU FACE ANY PROBLEM WHILE LEARNING BUILDING NETWORKS WITH TENSORFLOW ( OFTEN ABBREVIATED AS TF) YOU WILL FIND AN ANSWER ON THE INTERNET . SUCH HUGE COMMUNITY ENSURES SOMEONE MIGHT HAVE FACED THE SAME PROBLEM AND CHANCES OF FINDING SOLUTIONS ON STACK OVERFLOW OR QUORA ARE HIGH . KERAS IS PRETTY STRAIGHTFORWARD AND SIMPLE TO USE . ITS MORE USER FRIENDLY .
KERAS IS BUILT ON TENSORFLOW
KERAS IS A HIGH LEVEL API THAT IS BUILT ON TENSORFLOW ,INFACT KERAS CAN RUN ON THEANO AND CNTK(MICROSOFT)TOO . THE VERY FACT THAT KERAS IS BUILT ON TENSORFLOW RAISES THE QUESTION THAT HOW THEY ARE DIFFERENT . LETS ANSWER THAT QUESTION .
THE PURPOSE OF KERAS
NOT EVERY ONE WANTS TO DESIGN NEURAL NETWORKS FROM SCRATCH . KERAS PROVIDES A VERY SIMPLE STRAIGHTFORWARD FRAMEWORK FOR BUILDING MODELS . THIS MAKES IT USER FRIENDLY AND HENCE PEOPLE WITH LITTLE OR NO EXPERIENCE IN DEEP LEARNING CAN ALSO DESIGN AND TRAIN MODELS USING KERAS. IT ALSO COMES IN HANDY WHEN YOU ARE NOT CONCERNED WITH THE MINUTE DETAILS OF THE OPERATIONS OF TRAINING A MODEL.
TF ON THE OTHER HAND IS SLIGHTLY MORE SKILL DEMANDING . IN CONTRAST TO KERAS WHERE YOU HAVE LITTLE CONTROL ON THE FLEXIBILITY OF THE MODELS , TENSORFLOW ALLOWS THE ENGINEER TO TO DEVELOP NETWORKS USING BETTER OPERATIONS AND FUNCTIONALITIES , MORE LIKE A CUSTOMISED NETWORK . YOU CAN COMPARE THIS SITUATION WITH SOMETHING LIKE BUYING A CAKE. KERAS REFERS TO A CAKE SHOP WHERE YOU HAVE ALL THESE FANCY CAKES READY TO BE EATEN ,YOU CAN SELECT THESE CAKES BUT CANNOT CHANGE THEIR INGREDIENTS . ON THE OTHER HAND BAKING A CAKE WHERE YOU ADD BASIC INGREDIENTS ACCORDING TO YOUR TASTE, YOU DECIDE THE OVEN TIME .THIS REFERS TO THE TENSORFLOW FRAME WORK.
TF ALLOWS YOU TO USE A TOOL CALLED TENSOBOARD WHERE YOU CAN VISUALISE GRAPHS ,DATA ,GRADIENTS AND OTHER PARAMETERS. TENSORFLOW ALSO COMES WITH A DEBUGGER WHICH KERAS FRAMEWORK LACKS .
SO WHATS THE FINAL WORD? WHAT SHOULD YOU CHOOSE
WHEN YOU LEARN SOMETHING NEW ,OR ARE TRYING TO TACKLE A NEW PROBLEM ITS BETTER T HAVE A THOROUGH UNDERSTANDING OF THE PROBLEM ,TO HAVE THINGS IN CONTROL , PREFER A TENSORFLOW FRAMEWORK , BECAUSE YOU HAVE TO EXPERIMENT AND TRY OUT ALL THE OPTIONS . BUT ONCE YOU FACE A PROBLEM YOU ALREADY HAVE SOLVED YOU KNOW HOW TO HANDLE IT RATHER STRAIGHTFORWARD . LIKE A KERAS FRAMEWORK , EASY!!!!
I WOULD STRONGLY SUGGEST EXPLORING THE DOCUMENTATION PAGES OF TENSORFLOW AND KERAS FRAMEWORKS . THERE YOU CAN FIND TUTORIALS REGARDING SIMPLE MODEL BUILDING PROCESSES AND ALSO REGARDING THE LIBRARIES USED.
