fit with split produces error about tensorflow-directml HOT 11 CLOSED

Right, these are the more dangerous changes in the APIs since they don't throw errors and will just silently give you the wrong answer. This can make it hard to backport a script from 2.0 to 1.15. Sometimes, some 2.0 APIs just don't exist in 1.15 (or require a lot more work to make it work).

from tensorflow-directml.

Hi @TashaSkyUp,

Have you tried running this script with the official tensorflow 1.15 package? pip install tensorflow==1.15 --force-reinstall

tensorflow-directml doesn't really modify the API surface since we want scripts to be compatible as much as possible with the official package, and this particular error looks like an issue at the python level. It could still be an error on our side, but I want to make sure that you're not getting the same error with the official tensorflow 1.15 package before.

from tensorflow-directml.

Thank you no I have not I had it working in TF 2.x I did not realize this project was only for TF 1.15? Its probably that i'm not using standard conventions as far as how to organize my input ("x") . But im self taught so right now how i do things is an amalgamation of 100 tutorials 4 udemy classes and 10000 stack overflow posts.

from tensorflow-directml.

Yes, this project is based on TF 1.15.

TF 1.15 and TF 2.x have quite a few differences when it comes to the API. The biggest one is probably that TF 2.x uses eager execution by default, but you need to manually enable it for TF 1.15. I think this could be the cause of the error here.

from tensorflow-directml.

I enabled it already actually, that was one of many things I've done to make this work again. also replacing line 1871 with line 1873 seems to have worked for getting training going.. but now my output when i call the model is.. wrong.. Direct ml with my 5700x is about 2.5x faster than my cpu (3900x) But i dont know if i want to start over completely. I've been working on this sooo long.

from tensorflow-directml.

The easiest way to know whether the wrong output is caused by tensorflow-directml is to run your script with the tensorflow 1.15 package and compare the results between the two.

Otherwise, if it's possible for you to share the script (e.g. if it's on a github or somewhere else), we could give it a look.

from tensorflow-directml.

That sounds like a great Idea I should be able to share the code.. ill make a repo...

https://github.com/TashaSkyUp/UnShiney/blob/main/UnShineyMask.ipynb

from tensorflow-directml.

tested after re installing tf 1.15. Having both of the same problems.

from tensorflow-directml.

So for my output problem. It looks like tf 1.15 handles

img= tf.keras.preprocessing.image.load_img(img,color_mode="grayscale")

differently then my implementation I had when using 2.0
1.15 ends up with an array of floats between 0 and 255
2.0 ended up with floats between 0 and 1

from tensorflow-directml.

I created a pull request to fix this.

from tensorflow-directml.

#75

from tensorflow-directml.