TensorFlow-Intel

2.18.0 · active · verified Thu Apr 16

TensorFlow-Intel is an optimized distribution of the open-source machine learning framework TensorFlow, specifically tailored for Intel CPUs. It leverages Intel's oneAPI Deep Neural Network Library (oneDNN) primitives to enhance performance for deep learning workloads. This package is compatible with existing TensorFlow code and is maintained by Intel in collaboration with Google, aiming to provide maximal performance on Intel architectures. It is actively developed with regular updates to support newer TensorFlow versions and Intel hardware.

Common errors

```
Could not find CUDA drivers on your machine, GPU will not be used.
```
cause This message is often benign when running `tensorflow-intel` on a CPU-only system or an Intel GPU system without NVIDIA CUDA. TensorFlow attempts to detect CUDA devices by default.

fix If you intend to use only the CPU or an Intel GPU (and have `Intel® Extension for TensorFlow*` installed and configured correctly), this message can usually be ignored. If you have an NVIDIA GPU, ensure CUDA and cuDNN are correctly installed and configured for TensorFlow.
```
ModuleNotFoundError: No module named 'tensorflow' (or 'keras', 'absl-py')
```
cause The package or one of its core dependencies is not installed, or the Python environment is not correctly activated.

fix Ensure you have activated the correct virtual environment and run `pip install tensorflow-intel`. If `absl-py` is specifically mentioned, `pip install absl-py` might be needed, though it's typically a transitive dependency.
```
Could not find a version that satisfies the requirement tensorflow-intel==X.Y.Z
```
cause The specified version of `tensorflow-intel` is not available for your Python version, operating system, or architecture.

fix Check the available versions on PyPI for your specific Python version and OS (e.g., `pip install 'tensorflow-intel<2.11'` for older Windows GPU versions, or simply `pip install tensorflow-intel` for the latest compatible version). Consider upgrading your Python interpreter if needed, as `tensorflow-intel` requires Python >=3.9.

Warnings

deprecated The separate `Intel® Optimization for TensorFlow*` package (`intel-tensorflow`) has been discontinued as of Q1 2024. Users should transition to `tensorflow-intel` or stock TensorFlow with `Intel® Extension for TensorFlow*` for Intel CPU optimizations.
Fix: Use `pip install tensorflow-intel` or `pip install intel-extension-for-tensorflow[cpu]` alongside `tensorflow`.
gotcha Installing `tensorflow` or `tensorflow-cpu` on Windows might automatically install `tensorflow-intel` to provide Intel optimizations. While generally beneficial, this can be unexpected for some users.
Fix: Be aware of this automatic substitution; if a pure upstream TensorFlow build is desired, specific build-from-source methods or alternative distributions might be needed.
gotcha Enabling oneAPI Deep Neural Network Library (oneDNN) optimizations required setting the environment variable `TF_ENABLE_ONEDNN_OPTS=1` for TensorFlow versions 2.5 through 2.8. Starting with TensorFlow v2.9 (and thus in `tensorflow-intel`), these optimizations are enabled by default.
Fix: For versions <2.9, ensure `export TF_ENABLE_ONEDNN_OPTS=1` is set before running your script. For >=2.9, no action is needed as it's enabled by default.
gotcha Attempting to install both `tensorflow-intel` and `intel-extension-for-tensorflow` in the same Python environment, or mixing `tensorflow-intel` with the old `intel-tensorflow` (Intel® Optimization for TensorFlow*), can lead to conflicts and undefined behavior.
Fix: Choose one package for Intel optimizations. `tensorflow-intel` is for Intel CPU optimizations, while `intel-extension-for-tensorflow` is a plugin for stock TensorFlow to enable both CPU and GPU acceleration. Do not install both simultaneously.

Install

pip install tensorflow-intel Install latest stable version

Imports

tensorflow
```
import tensorflow as tf
```
The primary import is the standard TensorFlow module; optimizations are applied under the hood.
tensorflow.keras
```
from tensorflow import keras
```
Keras is integrated into TensorFlow and imported directly from the tensorflow package.

Quickstart

This quickstart code verifies the TensorFlow-Intel installation by performing a simple tensor operation and demonstrates a basic Keras model setup. A successful installation should print a tensor sum and the model summary.

import tensorflow as tf

# Verify TensorFlow installation and CPU setup
print(tf.reduce_sum(tf.random.normal([1000, 1000])))

# Basic Keras example
model = tf.keras.Sequential([
    tf.keras.layers.Dense(10, activation='relu', input_shape=(784,)),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
model.summary()

view raw JSON →