tfdf 1.9.0 only compatible with tf 2.16.1 which ships Keras 3
Closed this issue · 8 comments
Hi,
i need the fix for the issue described in google/yggdrasil-decision-forests#78 which has been released with tf-df 1.9.0
. However, the compatibility table from known_issues.md
states that this version is only compatible with tensorflow 2.16.1
. But this tensorflow version ships Keras 3 which is incompatible with tfdf.
Force installing an older version of tensorflow to get the legacy keras api will produce this error:
RuntimeError: Op type not registered 'SimpleMLLoadModelFromPathWithHandle' in binary running on hostname. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib (e.g. `tf.contrib.resampler`), accessing should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.
I suspect this is because the incompatibility of the abi between tf versions, but i dont see a way out of here.
I need the newer tfdf version, as pydf didnt get a newer release yet.
Thanks for your help!
Hi,
Unfortunately it's not possible to deploy a TF-DF version that's compatible with multiple TensorFlow versions since TensorFlow's ABI constantly changes.
We are very, very close to releasing a new PYDF version which includes the fix you need. No promises (unexpected things can always happen), but I'd suggest you wait a couple of days and check back then.
If you really can't wait, I'd suggest you give compiling PYDF a try, there are some instructions on how to do this on Github.
Alright, thanks for the info.
I actually tried to build pydf myself, but it always fails at the stage where the manylinux wheel is repaired because my toolchain is too new. And the proper image for building manylinux2014 wheels is CentOS 7 based which doesnt have a recent gcc-9 available (the one provided has a strange bug), so im unable to build the manylinux wheel. Maybe i will just try to build a non-manylinux wheel, but then id have to dig a bit deeper in the provided script.
Ok, I also plan to update the install instructions (and scripts) with the 0.3.0 release, so this will also get easier hopefully :)
Thanks! Then i will just wait a bit i think. But the issue i mentioned in the first comment still remains right?
Currently, there is no way to load any model when using the most recent tfdf version right? Or am i missing something here?
Sorry, I overlooked this question.
No, you can still use old models with TF-DF, but you need to instruct tensorflow / keras to use Keras 2:
- Make sure
tf_keras
is installed (it's a dependency of TF-DF, so this should already be the case). - At the top of your program, set
import os
# Keep using Keras 2
os.environ['TF_USE_LEGACY_KERAS'] = '1'
import tf_keras
and replace tf.keras
with tf_keras
in your pipeline (this might not even be necessary after setting the environment variable)
If there's any issues after this step, we would consider it a bug. Our tutorials https://www.tensorflow.org/decision_forests/tutorials/beginner_colab have been adapted to these changes and seem to work fine
Okay, then i think it is a bug.
Doing
pip install tensorflow-decision-forests==1.9.0 tf_keras
in a fresh environment and then running
import os
# Keep using Keras 2
os.environ['TF_USE_LEGACY_KERAS'] = '1'
import tf_keras
import sys
model = tf_keras.models.load_model(sys.argv[1])
in it yields
RuntimeError: Op type not registered 'SimpleMLLoadModelFromPathWithHandle' in binary running on de2lxl-520977. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib (e.g. `tf.contrib.resampler`), accessing should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.
As described in the first comment.
You also need to import tensorflow decision forests (sorry for not being clear), i.e. the full code is
import os
# Keep using Keras 2
os.environ['TF_USE_LEGACY_KERAS'] = '1'
import tensorflow_decision_forests
import tf_keras
import sys
model = tf_keras.models.load_model(sys.argv[1])
(Technical background: TF-DF defines custom ops in Tensorflow for saving / loading / training Decision Tree models. These ops can only be used by tensorflow after TF-DF has been imported)
Okay i should have figured this out myself 🤦
I was somehow expecting that the custom ops get registered "system-wide" as soon as you install tfdf.
Thank you so much!