TensorFlow Release 18.04
The NVIDIA container image of TensorFlow, release 18.04, is available.
Contents of TensorFlow
This container image contains the complete source of the version of NVIDIA TensorFlow in /opt/tensorflow. It is pre-built and installed as a system Python module.
To achieve optimum TensorFlow performance, for image based training, the container includes a sample script that demonstrates efficient training of convolutional neural networks (CNNs). The sample script may need to be modified to fit your application.
- Ubuntu 16.04
- NVIDIA CUDA 9.0.176 (see Errata section and 2.1) including CUDA® Basic Linear Algebra Subroutines library™ (cuBLAS) 9.0.333 (see section 2.3.1)
- NVIDIA CUDA® Deep Neural Network library™ (cuDNN) 7.1.1
- NCCL 2.1.15 (optimized for NVLink™ )
- Horovod™ 0.11.3
- OpenMPI™ 3.0.0
- TensorBoard 0.4.0-rc1
- MLNX_OFED 3.4
Driver Requirements
Release 18.04 is based on CUDA 9, which requires NVIDIA Driver release 384.xx.
Key Features and Enhancements
- TensorFlow container image version 18.04 is based on TensorFlow 1.7.0.
- Added the Mellanox user-space InfiniBand driver to the container.
- Latest version of MLNX_OFED 3.4
- Added support for TensorRT integration in TensorFlow. For functionality details, see TensorRT Integration Speeds Up TensorFlow Inference and the example in the nvidia-examples/tftrt directory.
- Improved nvidia_examples/nvcnn.py and nvcnn_hvd.py to ensure ResNet-50 model converges correctly out of the box. See Changelog at the top of nvidia_examples/nvcnn.py for more details.
- Enabled Tensor Op math for cuDNN-based RNNs in FP16 precision. This is enabled by default, but can be disabled by setting the environment variable TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH=1.
- Includes integration with TensorRT 3.0.4
- Latest version of NCCL 2.1.15
- Ubuntu 16.04 with March 2018 updates
Announcements
Starting with the next major version of CUDA release, we will no longer provide Python 2 containers and will only maintain Python 3 containers.
Known Issues
There is a degraded performance for graph construction time of grouped convolutions. For more information, see Support for depthwise convolution by groups.