Prefer YDF for new projects
YDF is Google's new library to train Decision Forests.
YDF extends the power of TF-DF, offering new features, a simplified API, faster training times, updated documentation, and enhanced compatibility with popular ML libraries.
Some of the issues mentioned below are fixed in YDF.
Windows Pip package is not available
TensorFlow Decision Forests is not yet available as a Windows Pip package.
Workarounds:
- Solution #1: Install Windows Subsystem for Linux (WSL) on your Windows machine and follow the Linux instructions.
Incompatibility with Keras 3
Compatibility with Keras 3 is not yet implemented. Use tf_keras or a TensorFlow version before 2.16. Alternatively, use ydf.
Untested for conda
While TF-DF might work with Conda, this is not tested and we currently do not maintain packages on conda-forge.
Incompatibility with old or nightly versions of TensorFlow
TensorFlow's ABI is not compatible in between releases. Because TF-DF relies on custom TensorFlow C++ ops, each version of TF-DF is tied to a specific version of TensorFlow. The last released version of TF-DF is always tied to the last released version of TensorFlow.
For these reasons, the current version of TF-DF might not be compatible with older versions or with the nightly build of TensorFlow.
If using incompatible versions of TF and TF-DF, you will see cryptic errors such as:
tensorflow_decision_forests/tensorflow/ops/training/training.so: undefined symbol: _ZN10tensorflow11GetNodeAttrERKNS_9AttrSliceEN4absl14lts_2020_09_2311string_viewEPSs
- Use the version of TF-DF that is compatible with your version of TensorFlow.
Compatibility table
The following table shows the compatibility between
tensorflow_decision_forests and its dependencies:
| tensorflow_decision_forests | tensorflow |
|---|---|
| 1.11.0 | 2.18.0 |
| 1.10.0 | 2.17.0 |
| 1.9.2 | 2.16.2 |
| 1.9.1 | 2.16.1 |
| 1.9.0 | 2.16.1 |
| 1.8.0 - 1.8.1 | 2.15.0 |
| 1.6.0 - 1.7.0 | 2.14.0 |
| 1.5.0 | 2.13.0 |
| 1.3.0 - 1.4.0 | 2.12.0 |
| 1.1.0 - 1.2.0 | 2.11.0 |
| 1.0.0 - 1.0.1 | 2.10.0 - 2.10.1 |
| 0.2.6 - 0.2.7 | 2.9.1 |
| 0.2.5 | 2.9 |
| 0.2.4 | 2.8 |
| 0.2.1 - 0.2.3 | 2.7 |
| 0.1.9 - 0.2.0 | 2.6 |
| 0.1.1 - 0.1.8 | 2.5 |
| 0.1.0 | 2.4 |
- Solution #2: Wrap your preprocessing function into another function that squeezes its inputs.
Not all models support distributed training and distribute strategies
Unless specified, models are trained on a single machine and are not compatible
with distribution strategies. For example the GradientBoostedTreesModel does
not support distributed training while DistributedGradientBoostedTreesModel
does.
Workarounds:
- Use a model that supports distribution strategies (e.g.
DistributedGradientBoostedTreesModel), or downsample your dataset so that it fits on a single machine.
No support for GPU / TPU.
TF-DF does not supports GPU or TPU training. Compiling with AVX instructions, however, may speed up serving.
No support for model_to_estimator
TF-DF does not implement the APIs required to convert a trained/untrained model to the estimator format.
Loaded models behave differently than Python models.
While abstracted by the Keras API, a model instantiated in Python (e.g., with
tfdf.keras.RandomForestModel()) and a model loaded from disk (e.g., with
tf_keras.models.load_model()) can behave differently. Notably, a Python
instantiated model automatically applies necessary type conversions. For
example, if a float64 feature is fed to a model expecting a float32 feature,
this conversion is performed implicitly. However, such a conversion is not
possible for models loaded from disk. It is therefore important that the
training data and the inference data always have the exact same type.
Tensorflow feature name sanitization
Tensorflow sanitizes feature names and might, for instance, convert them to lowercase.