Caffe2 Release 17.06
The NVIDIA container image of Caffe2, release 17.06, is available.
Caffe2 container image version 17.06 is based on Caffe2 0.7.0.
Contents of Caffe2
This container image contains the complete source of the version of Caffe2 in
/opt/caffe2. It is pre-built and installed into the
/opt/caffe2/[binaries,lib] directories in the container image.
The container also includes the following:
- Ubuntu 16.04
- NVIDIA CUDA 8.0.61
- NVIDIA CUDA® Deep Neural Network library™ (cuDNN) 6.0.21
- NCCL 1.6.1 (optimized for NVLink™ )
- OpenMPI™ 1.10.3
Key Features and Enhancements
This Caffe2 release includes the following key features and enhancements.
- Ubuntu 16.04 with May 2017 updates
The NCCL library version 1.6.1 included in this image, modifies the output buffers on all GPUs during in-place ncclReduce() operations, whereas normally only the "root" (target) device's output buffer should be modified. This is fixed in later versions of NCCL, as will be packaged in later versions of this image. As a workaround, either use ncclAllReduce(), which correctly modifies output buffers of all GPUs to the same values, or use out-of-place ncclReduce(), wherein the output buffer is distinct from the input buffer.
In some cases, the use of GPU transforms in the ImageInput operator can cause out-of-bound memory accesses. To avoid this, set
use_gpu_transform=False in the ImageInput operator of your network definition.