A Quick Comparison Of The Two High-Level Frameworks For Deep Learning
fastai
- Built on the top of PyTorch
- Emphasis on:
a) Rapid initial results for beginners — SOTA results with a handful of lines of code.
b) Straightforward customisation as your skills increase
c) Smooth integration with PyTorch
3. Backbone of popular deep learning courses
Keras
- Since TensorFlow 2.0, integrated with TensorFlow
- Default platform for learning about deep learning
- Taught in dozens of courses and used by thousands of developers
fastai : pros and cons
Pro :
a) Complete solution in 5 lines
b) Intelligent defaults & clever built-in optimizations = SOTA / close to SOTA results “out of the box”
c) Layered architecture = room to grow from begineer to expert
d) Huge suite of convenience functions, e.g. for data preparation
e) Big set of curated datasets
Cons :
a) Smaller community than Keras
b) Limited examples of implementations in commercial production
c) Inconsistent documentation
Keras : pros and cons
Pro :
a) Huge user community
b ) Many examples of deployments in commercial production
c) Consistent, orderly evolution
d) Good documentation
Cons:
a) Harder to get started
b) More knobs to manage
c) Step function between Keras and TF in terms of complexity