All this while I have been coding up my Deep Learning projects using Keras. Its needless to explain the scintillating functionality that it provides. A few days back as I was going through the Deep learning courses offered by deeplearning.ai(coursera) revising my concepts , I was reminded of the fact that most of the Deep Learning and Machine models are better understood by the means of computation graphs.These computation graphs are more intuitive in nature.This engendered curiosity,forcing me to go a level lower and look through the abstraction i.e TensorFlow. Whenever I discuss this topic with my peers the response always is “Why reinvent the wheel ?”.To that I say- “If you absolutely love graphs than this is the level of abstraction to be in”.
I started googling out looking for differences in Keras and TensorFlow. This search resulted in the following crisp points-
- Keras is a High level API(wrapper) on back-end frameworks like TensorFlow (TF) and Theano. For those using a Theano back-end its better you switch to TF as the growth of Theano community is at its limit.
- Keras has a faster learning Curve thus making it easier to code and prototype in.
- Efficiency and results are same in both.
- TF has a fleeting edge over Keras with respect to flexibility.I call it fleeting as the gap is dwindling with the use of Custom layers,lambda layers,functional API etc.
- The place where TF dominates is when designing new original algorithms come into play.If you wan’t to design your own cost functions,metrics and layers,If you wish to have a better handle on Threading Queues and graph debugging then TF is the better choice.
We can infer from the above that almost everything can be achieved using Keras and it is soon to overtake TF and become the leading TF-High level API.
Ranking popular deep learning libraries
Article Link- https://www.kdnuggets.com/2017/10/ranking-popular-deep-learning-libraries-data-science.html
So with this Surfing through TensorFlow series of articles I will go about explaining various functionalities and practical aspects of TensorFlow that I have gained and gaining through researching various sources.I thought of this as an idea to review the concepts that I grasp,but I sincerely hope that it helps some of you.
Source: Deep Learning on Medium