The best part is if offers broad adoption, support for a wide range of production deployment options, integration with at so our best framed back-end engines - super easy to deploy Review collected by and hosted on G2.com.

We think Keras should do its own operations - not rely on back-end services. I know its too much to ask for but it would be great to have it's own. Just that too much reliability on low level operations makes us think - did they do that due to infrastructure or was it a choice? That being said - the feature extraction is something to die for - with our favorite language Python in use Review collected by and hosted on G2.com.

I like that tensorflow is moving away from tensorflow extended, slim, keras, etc. formats to having everything be done in keras. It is see hardware that touts support for tensorflow models only to learn that the associated software only supports object detection api models or only supports tensorflow extended models. I want things to just work across hardware and software libraries. Review collected by and hosted on G2.com.

Although Keras is an improvement towards easy customization of existing models it doesn't go far enough in making it easy to try out a variety of custom architectures. Simple things like being able to view weights on a neuron and surgically remove it if it doesn't meet some custom criteria, then save and reload are not supported. I'm using third party projects to try to make this easier (https://github.com/Raukk/tf-keras-surgeon/), but I feel like the third party project is a bit of a hack and could be improved by having native library support.

It is hard to take a new an improved model that someone else has written and tweak it with the method from a recent paper of replacing all the X layers with Y, for which they saw some improvement. I have about twenty papers with improvements over a baseline model and a lot of these aren't independent, but putting them all into one model is unnecessarily difficult. Review collected by and hosted on G2.com.

1. I love the fact that with very few lines of code and little knowledge about deep learning and convolutional neural networks, one can easily build an image classifier using by just reading its documentation.

2. Keras community and developer support is very high and active, so there are a lot of tutorials on Keras available.

3. I love the fact that I can easily integrate Keras and Scikit-learn, and apply functions on scikit-learn

such as ensembles, cross-validation and stacking. Review collected by and hosted on G2.com.

Keras is a high-level deep learning framework, which makes customization and tweaking hard. Review collected by and hosted on G2.com.

Keras is an open source deep learning framework with lots and lots of features it allows you to do so many things like creating multi later neural networks etc. Keras is performs computations quickly and it is built upon Tensorflow which is one of the best frameworks out there. The fact I can build a full neural network in less than 9 lines is because of keras. It also has all the required deep learning functions killer Dropout,Dense, softmax amongst others Review collected by and hosted on G2.com.

Keras is structured as an API to simply make function calls to Tensorflow. Making it hard to implement features that aren't out of the box Review collected by and hosted on G2.com.

Keras is very flexible abd user-friendly when it comes to training models Review collected by and hosted on G2.com.

Not compatible with Tensorflow. Unable to convert some trained TF models to Keras Review collected by and hosted on G2.com.

I believe Keras is one of the main pillars that raised deep learning to where it is now. It is easy to use and powerful. The seamless integration with Tensorflow is fantastic. Keras is one of the best if not the best. Review collected by and hosted on G2.com.

The only thing I did not like during my journey with Keras is how they did not do a good job of keeping backward compatibility. Sometimes, I had to make major changes to accommodate newer versions. Review collected by and hosted on G2.com.

Its the frontend every ML library should use. The api is quite understandable, building up layers makes sense, the estimator api of tensorflow is too lost in the details to abstract such ease of use. Loss functions, activations, optimizers, lstm you got it all Review collected by and hosted on G2.com.

Number of open issues on github, this project needs more directed support. It can rock the AI world. Tensorflow realized its importance and optimized it to intergrate it into the roots in 2.0 but still 2.0 is not released on stable channel. A few months more maybe Review collected by and hosted on G2.com.

I love the ease that Keras provides, offering a brilliant abstracted wrapper around TensorFlow, allowing you to build most customizations for neural networks. Review collected by and hosted on G2.com.

The fact that Keras is dropping functionalities for other frameworks one after the other and proving to be a strong abstraction of TensorFlow. What is the difference between TensorFlow Keras then? Review collected by and hosted on G2.com.

easy to deploy models, fine tune them, and export them Review collected by and hosted on G2.com.

the documentation is lacking; only the most basic of networks are clearly explained; the concepts behind the hyperparameters are not explained at all (what is an optimizer? what is the difference between adagrad and adam?); how do you build more complicated networks like recurrent NN ? Review collected by and hosted on G2.com.

Keras is a high level API. It is easy to use. Better than tensorflow (low level api). Review collected by and hosted on G2.com.

There is nothing in keras which you don't like. It is perfect. Review collected by and hosted on G2.com.