NVIDIA/DALI v1.18.0 on GitHub

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Unified batch representation in the GPU and CPU stages of the pipeline (effort towards conditional execution) (#4253, #4236, #4220, #4189).
Added support to specify the fill_value argument for each sample in the fn.erase operator (#4182).
Added support for the memory video file in FramesDecoder (#4184).
Moved the audio_resample operator out of experimental module (#4194).

Fixed Issues

The following issues were fixed in this release:

Fixed an unnecessary synchronization in MakeContiguous. (#4248).
Fixed the Python tool to create the webdataset index (#4226).
Added a fix to prevent DALI from allocating GPU memory when constructing CPU TensorList (#4203).
Fixed a PyTorch example to comply with the new PyTroch (#4213).

Improvements

GPU image data conversion (#4208)
Fix libtiff and libtar vulnerabilities (#4245)
Update third party dependencies (#4233)
Reduce batch size in the WebDataset integration using External Source example (#4240)
Rename the set and copy sample APIs in TensorList (#4236)
Move nvjpeg decoder files to imgcodec/decoders/nvjpeg/ (#4235)
Add Nvjpeg decoder (#4178)
Rename TensorVector to TensorList (#4220)
Make JPEG HW decoder test to fully use HW and not hybrid approach (#4222)
Add bulk parameter passing to decoders and factories. (#4212)
Support any bitdepth in TIFF (#4180)
Remove TensorList and use only TensorVector (#4189)
[imgcodec] API adjustments (#4205)
ROI support for nvjpeg2k decoder (#4175)
Use deprecated PIL resampling import for Python 3.6, due to lack of availability of a newer version of PIL (#4200)
Add arithmetic expression broadcasting utils (#4188)
Support higher TIFF bitdepths (#4174)
Enable per-sample fill_value argument in Erase operator (#4182)
Fix python linter errors for the qa/ directory (#4117)
Fix usage of deprecated np.float in tests (#4192)
Adjust PIL interpolation types to module PIL.Image.Resampling (#4195)
Move audio_resample out of experimental module (#4194)
Support different layouts in imgcodec's Convert (#4157)
Fix typos in iterator last_batch_policy argument documentation (#4170)
Fix synchronization in external source tests (#4153)
Add support for memory video file in FramesDecoder (#4184)
Support outputting YCbCr in libjpeg-turbo decoder (#4156)
Use std::exchange in move operator for Tensors (#4183)

Bug Fixes

Unify buffers caching in CPU/GPU external source (#4253)
Fix builds without nvJPEG (#4252)
Separate nvjpeg lib wrapper and stub from the decoder (#4249)
Prevent unnecessary synchronization in MakeContiguous. (#4248)
Do not leak DecodeParams (#4242)
Fix AssertClose bug in Imgcodec tests (#4243)
Fix bug in CPU Convert (#4237)
Fix webdataset python index creation script (#4226)
Fix In memory video decoding tests (#4216)
Fix UnpackBits (#4227)
Fix issues detected by Coverity. (#4221)
Make TensorList constructor for CPU not using GPU memory (#4203)
Fix the indexing for newer PyTorch (#4213)
Fix possibly incorrect parallel write access to vector (#4211)
Fix Layout propagation in TensorVector (#4202)

Breaking API changes

There are no breaking changes in this DALI release.

Deprecated features

There are no deprecated features in this DALI release.

Known issues:

The video loader operator requires that the key frames occur, at a minimum, every 10 to 15 frames of the video stream.
If the key frames occur at a frequency that is less than 10-15 frames, the returned frames might be out of sync.
Experimental VideoReaderDecoder does not support open GOP.
It will not report an error and might produce invalid frames. VideoReader uses a heuristic approach to detect open GOP and should work in most common cases.
The DALI TensorFlow plugin might not be compatible with TensorFlow versions 1.15.0 and later.
To use DALI with the TensorFlow version that does not have a prebuilt plugin binary shipped with DALI, make sure that the compiler that is used to build TensorFlow exists on the system during the plugin installation. (Depending on the particular version, you can use GCC 4.8.4, GCC 4.8.5, or GCC 5.4.)
In experimental debug and eager modes, the GPU external source is not properly synchronized with DALI internal streams.
As a workaround, you can manually synchronize the device before returning the data from the callback.
Due to some known issues with meltdown/spectra mitigations and DALI, DALI shows best performance when running in Docker with escalated privileges, for example:
- privileged=yes in Extra Settings for AWS data points
- --privileged or --security-opt seccomp=unconfined for bare Docker.

Binary builds

Install via pip for CUDA 10.2:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda102==1.18.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda102==1.18.0

or for CUDA 11:

CUDA 11.0 build uses CUDA toolkit enhanced compatibility. It is built with the latest CUDA 11.x toolkit
while it can run on the latest, stable CUDA 11.0 capable drivers (450.80 or later). 
Using the latest driver may enable additional functionality. 
More details can be found in enhanced CUDA compatibility guide.

pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda110==1.18.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda110==1.18.0

Or use direct download links (CUDA 10.2):

Or use direct download links (CUDA 11.0):

FFmpeg source code:

This software uses code of FFmpeg licensed under the LGPLv2.1 and its source can be downloaded here

Libsndfile source code:

https://developer.download.nvidia.com/compute/redist/nvidia-dali/libsndfile-1.1.0.tar.gz

NVIDIA/DALI v1.18.0 DALI v1.18.0 on GitHub