Introduction

The error OpKernel ('op: "BestSplits" device_type: "CPU"') for unknown op: BestSplits occurs when TensorFlow cannot find the kernel implementation for the BestSplits operation. This operation is part of TensorFlow's tensor_forest contrib module (used for random forests). The error typically appears on TensorFlow 1.x on Windows because the contrib op was not properly registered in the Windows build, or the required contrib module was not imported. The fix depends on your TensorFlow version — upgrade to TF 2.x with tensorflow-decision-forests, or explicitly import the contrib module in TF 1.x.

The Error

1import tensorflow as tf
2
3# Using tensor_forest in TF 1.x
4hparams = tf.contrib.tensor_forest.python.tensor_forest.ForestHParams(
5    num_classes=2,
6    num_features=10,
7    num_trees=50
8)
9# OpKernel ('op: "BestSplits" device_type: "CPU"') for unknown op: BestSplits

The BestSplits op is a custom C++ kernel in tensor_forest. On Windows builds of TF 1.0.0-rc2, this kernel was not always compiled or registered correctly.

Fix 1: Import the Contrib Module Explicitly

In TF 1.x, some contrib ops need explicit imports to register their kernels:

1import tensorflow as tf
2
3# Force-load the tensor_forest ops
4from tensorflow.contrib.tensor_forest.python import tensor_forest
5from tensorflow.contrib.tensor_forest.python.ops import tensor_forest_ops
6
7# Now the BestSplits op should be registered
8hparams = tensor_forest.ForestHParams(
9    num_classes=2,
10    num_features=10,
11    num_trees=50
12)

Importing tensor_forest_ops triggers the shared library load that registers the custom ops including BestSplits.

Fix 2: Upgrade TensorFlow

TF 1.0.0-rc2 is extremely outdated. Many Windows-specific build issues were fixed in later releases:

1# Upgrade to latest TF 1.x (if you must stay on 1.x)
2pip install tensorflow==1.15.5
3
4# Better: upgrade to TF 2.x
5pip install tensorflow>=2.10

In TF 2.x, tf.contrib was removed entirely. Use tensorflow-decision-forests (TF-DF) instead:

pip install tensorflow-decision-forests

1import tensorflow_decision_forests as tfdf
2
3# TF-DF random forest — modern replacement for tensor_forest
4model = tfdf.keras.RandomForestModel(num_trees=50)
5model.fit(tf_dataset)

Fix 3: Load the Op Library Manually

If the op is compiled but not auto-registered:

1import tensorflow as tf
2import os
3
4# Find and load the shared library containing BestSplits
5contrib_dir = os.path.join(os.path.dirname(tf.__file__), 'contrib')
6forest_ops_path = os.path.join(
7    contrib_dir,
8    'tensor_forest',
9    'python',
10    'ops',
11    '_tensor_forest_ops.so'  # .pyd on Windows
12)
13
14if os.path.exists(forest_ops_path):
15    tf.load_op_library(forest_ops_path)
16else:
17    print(f"Op library not found at {forest_ops_path}")

On Windows, the file extension is .pyd or .dll instead of .so.

Fix 4: Build from Source

If the pre-built binary lacks the op, build TensorFlow from source with contrib support:

1# Clone TF 1.x source
2git clone -b v1.15.5 https://github.com/tensorflow/tensorflow.git
3cd tensorflow
4
5# Configure for Windows
6python configure.py
7
8# Build with contrib
9bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package

Building from source ensures all contrib ops including BestSplits are compiled for your platform.

Modern Alternative: TensorFlow Decision Forests

For new projects, use TF-DF instead of the deprecated tensor_forest:

1import tensorflow_decision_forests as tfdf
2import pandas as pd
3
4# Load data
5train_df = pd.DataFrame({
6    'feature1': [1.0, 2.0, 3.0, 4.0, 5.0],
7    'feature2': [5.0, 4.0, 3.0, 2.0, 1.0],
8    'label': [0, 0, 1, 1, 1]
9})
10
11train_ds = tfdf.keras.pd_dataframe_to_tf_dataset(train_df, label='label')
12
13# Train a random forest
14model = tfdf.keras.RandomForestModel(
15    num_trees=100,
16    max_depth=16
17)
18model.fit(train_ds)
19
20# Evaluate
21model.compile(metrics=["accuracy"])
22evaluation = model.evaluate(train_ds)
23
24# Inspect the model
25model.summary()

TF-DF has its own optimized C++ ops that are properly built for all platforms including Windows.

Checking Op Registration

1import tensorflow as tf
2
3# List all registered ops (TF 1.x)
4all_ops = tf.contrib.framework.get_registered_ops() if hasattr(tf, 'contrib') else []
5
6# Check if BestSplits is registered
7ops_list = [op.name for op in tf.get_default_graph().as_graph_def().node]
8
9# In TF 2.x, check with:
10# tf.raw_ops has all registered ops
11print(hasattr(tf.raw_ops, 'BestSplits'))  # True if registered

Version Compatibility

Common Pitfalls

• Using TF 1.0.0-rc2 in production: This is a release candidate, not a stable release. Upgrade to at least TF 1.15.5 for the last stable 1.x build with contrib support.
• Mixing TF versions: Installing tensorflow-decision-forests alongside TF 1.x causes conflicts. TF-DF requires TF 2.x.
• Windows .pyd vs .so: TensorFlow op libraries use .so on Linux/macOS and .pyd on Windows. Path lookups must use the correct extension.
• GPU vs CPU ops: Some custom ops are only registered for CPU. Running on GPU with tensor_forest may produce a different error: No OpKernel for device_type GPU.
• tf.contrib removal in TF 2.x: All tf.contrib modules were removed in TensorFlow 2.0. Code using tf.contrib.tensor_forest must be migrated to tensorflow-decision-forests or another library.

Summary

• The BestSplits unknown op error is caused by an unregistered kernel in TF 1.x Windows builds
• Import tensor_forest_ops explicitly to register the op in TF 1.x
• Upgrade to TF 1.15.5 for the most stable 1.x experience on Windows
• For new projects, use tensorflow-decision-forests (TF-DF) which replaces tensor_forest entirely
• tf.contrib was removed in TF 2.0 — all tensor_forest code must be migrated