Graph visualisation in tensorflow is much better.

Frequent updates as its backed by Google. Review collected by and hosted on G2.com.

Tensorflow lacks behind in terms of computation and speed.

It's only supported in NVIDIA GPUs. Review collected by and hosted on G2.com.

Easy to train images and create models which you can use in multiple plateforms like Windows, embedded devices and Mobile applications. Review collected by and hosted on G2.com.

Sometime it is hang on medium config system otherwise it is ok. Review collected by and hosted on G2.com.

This is having will defined and organized Classes for NN Review collected by and hosted on G2.com.

Sometimes dependencies give error during run time Review collected by and hosted on G2.com.

The Keras interface is super cool for beginners. Once you get comfortable, there are ample option to customize the data access during training and export the model. Review collected by and hosted on G2.com.

Its lerning curve is a little steep once you cross the beginners threshold. But can say its the worst :). Review collected by and hosted on G2.com.

Useful in all stages of development and production as well as most types of research work. It also gathers extensive pre-built and pre-trained systems and allows high level and low level access to most if not all components of the model. Review collected by and hosted on G2.com.

The team behind tensorflow is doing a great job and there is very little to dislike about it. Of course like every library, first comers will take a bit of time to get acquainted to it but it is getting easier to use with every version update.For new comers, watch out of the big shifts between some versions, which might require a bit of more work for compatibility. Review collected by and hosted on G2.com.

The Keras API: it makes building, training, working with DNNs, CNNs very easy. It requires only a very small knowlegde of (python) coding. Line by line one can add layers to the network to construct it, and the built-in optimizers do the training. No need to go through the pain if implementing backprop or optimizations steps on your own.

Also nice: the support for GPUs which really speeds up all computions. Review collected by and hosted on G2.com.

The (remants of) older TF versions using weird and sometimes incomprehensible stuff like placeholder, logits, etc. Review collected by and hosted on G2.com.

Neural network application has become so easy. User now no need to focus on the development and optimising the models, and rather could stay more focused with the application part of it. Review collected by and hosted on G2.com.

Required Memory and ram size could be optimized. Review collected by and hosted on G2.com.

It is open-source and can use for different platforms. There are lots of tutorials available that can make help students when that stuck in some part. Also, I really like the visualization tools of that through TensorBoar. Review collected by and hosted on G2.com.

I have a windows machine and there is no support for Windows Review collected by and hosted on G2.com.

Rapid prototyping. Especially in building a neural network. Review collected by and hosted on G2.com.

The documentation some times don't have plenty of examples for different scenarios Review collected by and hosted on G2.com.