# Is there any good alternative for Keras library?

I am currently using keras as a deep learning library on top of tensorflow. I just want to know is there any other library which is more efficient/easier than keras.

• I believe Keras's interface is most intuitive among all of the library (my personal opinion). You may want to look at pytorch if it meets your nees. – apsdehal Nov 12 '17 at 0:50
• When you say efficient / easier, it would be better if you could elaborate on what you are actually looking for. – Nischal Hp Nov 12 '17 at 2:37
• @apsdehal Thank you for your suggestion. I will definitely take a look at pytorch. – Chinmay Kulkarni Nov 12 '17 at 4:43
• @NischalHp I just want to know the alternative. It would be great if you could suggest me any library for implementing neural networks(specifically deep feedforward nets) other than keras. – Chinmay Kulkarni Nov 12 '17 at 4:48
• There are good uses of Lasagne (it works on top of Theano) – David Dale Nov 12 '17 at 9:46

I can't say which is more efficient or easier but MXNet exists, has interfaces to most languages (python, scala, R, C++, perl, Julia), and the code looks fairly similar to keras.

https://mxnet.incubator.apache.org/

Another one that I've come across is caffe, mostly because people seem to share a lot of pre-trained nets in their model zoo.

http://caffe.berkeleyvision.org/

Update: I didn't see caffe 2 but I think this is quite separate to the original caffe.

Finally I've not fully had time to check out H2O's offering in deep water which is an alternative to Keras but still sits on top of TF/MXNet/Caffe

I reccomend pytorch. You can find good tutorials here.

• Why do you recommend PyTorch? – nbro Jul 10 '18 at 20:58

Keras is a high-level API that can be used on top of TensorFlow, CNTK and Theano. You can use each of the low-level APIs but the problem of those is that you can get complicated if you design very deep nets whilst dealing with Keras is much easier. Consequently, Keras is designed for accelerating deep nets' designing. Keras is opensource like the underlying libraries it comes for and I guess its project is not for Google. TensorFlow itself has a high-level API, namely TFLearn. I cannot say which is better but the point is that try to master one of them perfectly.

It seems that plaidML Keras backend is also available which enables training on AMD graphics.

• Nowadays there is also the plaidML keras backend which allows training with amd cards. – SleepProgger Mar 16 '19 at 21:38