tf.keras instead of keras about keras-pandas

[bjherger]

Hey, great question. I started this project before tf had built in Keras support, but that’s not a great reason for continuing in this way. My only thought is that using the Keras module separately could better support other (non-tf) backends. If there’s a good reason to switch to tf.keras, I’d be happy to evaluate. BH

On Wed, Nov 7, 2018 at 01:55 Rudolf Hohn ***@***.***> wrote: Hi, Thanks for the time you are putting to make life easier between pandas and keras! I was checking out the code and the current issues and I saw that you are using the original keras module and not the version inside TensorFlow core. Is there any reason in particular to this? If not, what would be your thoughts about changing this? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#82>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHXyreBNwpA2UubRm0X5lSL1w8q-ZIyyks5usq4igaJpZM4YSHAp> .

-- Thanks, Brendan Herger 415.326.3513

[rudolfhohn]

By using tf.keras, it might be easier for people that uses both Keras and TF in the same project to use the module. But, actually I have never had a situation where I needed TF in my Keras code, so there might not be that many people in this particular situation.

About the other Keras backends, Theano is no longer maintained by its creators and the Keras support of CNTK is still in Beta for about a year. I guess nowadays most people uses either Keras and TF, or PyTorch (more an opinion that a fact, I don't have the numbers).

My main argument is that I don't have Keras installed on my machine, just TensorFlow. It means that in order to use your module I would need to install the right Keras version that my version of TF embeds. It is more a convenience rather than a real need I would say.

Happy to hear your thoughts.

[bjherger]

Ah, this is a great thought. Yeah, I won't take a side on Keras's support for other backends. I'm in the process of a 3.0 release which greatly simplifies supporting new data types, so it wouldn't make sense to update the existing code. A simple fix might be to check if keras is installed, and if not fall back to tf.keras. If this is a convenience issue though, it might difficult to prioritize unless there is a reason that users can't install keras. I'll keep this issue open, and recommend the actions below. In the mean time, I recommend installing keras-pandas via pip, which will install the necessary pre-reqs. **Current state** - Keras is installed as a pre-req package **Recommendation** - Have backend check if keras is installed; if it is, use it. If not use tf.keras. BH

On Wed, Nov 7, 2018 at 7:18 AM Rudolf Hohn ***@***.***> wrote: By using tf.keras, it might be easier for people that uses both Keras and TF in the same project to use the module. But, actually I have never had a situation where I needed TF in my Keras code, so there might not be that many people in this particular situation. About the other Keras backends, Theano is no longer maintained by its creators and the Keras support of CNTK is still in Beta for about a year. I guess nowadays most people uses either Keras and TF, or PyTorch (more an opinion that a fact, I don't have the numbers). My main argument is that I don't have Keras installed on my machine, just TensorFlow. It means that in order to use your module I would need to install the right Keras version that my version of TF embeds. It is more a convenience rather than a real need I would say. Happy to hear your thoughts. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#82 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHXyrXJD7SKpFfXdq2tS2ecBwhIlE6Yiks5usvnfgaJpZM4YSHAp> .

-- Thanks, Brendan Herger 415.326.3513