Keras is supposed to be simple?

[Russ Hall]

The two object detection examples on the Keras.io site are complicated scripts. One uses transformers and the other retina. If Keras is supposed to be simple and easy, why is there not a simple object detection model listed here?

[Dennis S]

Russ - good morning. I'm afraid that we might have a case of "selection bias" here :) Keras being "simple" is a relative term to any number of comparisons. I think that many of the veterans on this list have experience standing up a neural network from scratch (like with the raw math) and Keras does indeed make it simple. Same goes for the general fit functions and the collection of metrics, the management of all the epochs, etc, etc. Keras is indeed simple for those operations. 

Does Keras make everything simple? No, of course not - very few frameworks (of any kind, not just nn) can master all scenarios to make everything simple. But I think most of us here appreciate Keras for the things that it does do well. 

Perhaps you can take this on as a challenge and wrap your own code around Keras for a simple object detection model and add it to the codebase. Keras makes contributing to the codebase simple too :) 

Good Luck!

Dennis

On Tue, Jul 5, 2022 at 8:47 AM Russ Hall <rll...@gmail.com> wrote:

The two object detection examples on the Keras.io site are complicated scripts. One uses transformers and the other retina. If Keras is supposed to be simple and easy, why is there not a simple object detection model listed here?

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/5c439de6-3e03-4762-82e0-0d8f3256f788n%40googlegroups.com.

--

Thanks,

Dennis

[Matias Valdenegro]

This is for a very simple reason, the task of object detection with CNNs is itself a complex task, with many pre/post-processing steps required to obtain decent detections, so maybe select a much simpler task to make a good comparison.

[Russ Hall]

What I mean is, could a simple object detection script be included in the examples? This would be a script without the "latest and greatest" features, just a simple one that works. I did get a Tensorflow Lite Model Maker to train yesterday on my own data. That wasn't simple, but it was gratifying.

[Matias Valdenegro]

Then people would complain that the detection in keras works pretty bad, there is a trade-off between implementation complexity and performance, so you get one or the other, and you can complain in both cases.

On Wednesday, 6 July 2022 17:10:15 CEST Russ Hall wrote:

| What I mean is, could a simple object detection script be included in the

| examples? This would be a script without the "latest and greatest"

| features, just a simple one that works. I did get a Tensorflow Lite Model

| Maker to train yesterday on *my own data*. That wasn't simple, but it was

[Lance Norskog]

Tensorflow Hub might be a better hunting ground for you. It attempts to deliver packaged running models.

https://tfhub.dev/s?module-type=image-object-detection&tf-version=tf2

There is a class for using a TF-Hub model as a Keras layer, I don't remember where.

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/9cb39ed6-ea5a-46c7-baf4-03da5671a498n%40googlegroups.com.

--

Lance Norskog
lance....@gmail.com
Redwood City, CA