The ComboDataloader takes advantage of the CPU optimization of the existing
PyTorch video dataloader and the GPU optimization of NVIDIA's DALI video dataloader
to create a more efficient dataloader for video datasets.
It runs a separate producer subprocess for each base dataloader, allowing better parallelism between the CPU and GPU. The pytorch dataloader also uses pushdown of video resize to the decoding step with the decord decoder, using this fork of pytorchvideo to be able to pass arguments to the decoder used in the torch video dataloader.
First, install the combo_dataloader module: python setup.py
Several (quite similar) sample notebooks with example usage are included:
speedup_demo.ipynb: demonstrates the speedup of applying different optimizations to the ComboDataloader, and how to set up an optimal configuration.
train_demo.ipynb: demonstrates how to use the ComboDataloader in a video labeling train-test pipeline.
train_speedup.ipynb: demonstrates the speedup with the same configurations as in speedup_demo.ipynb, but in running a training pipeline rather than just dataloading.
model_zoo.ipynb: demonstrates using the ComboDataloader in pipelines with a few different models.
To build the Docker image from the provided Dockerfile, run docker build -t <image name> .. Note that the Dockerfile includes dependencies needed for running the test notebooks that are not included in requirements.txt, since the requirements.txt includes combo_dataloader module dependencies only.
To use a GPU in the Docker container, make sure to pass --runtime=nvidia, --ipc=host,
and --gpus '"capabilities=compute,utility,video"' to docker create.
Creating the container:
docker create --name="container name"
-v path/to/video_processing:/path/to/video_processing/in/container
-v path/to/dataset:path/to/dataset/in/container:ro
-i -t
--runtime=nvidia --ipc=host --gpus'"capabilities=compute,utility,video"'
<image name>
Starting the container:
docker start <container name>
Note that DALI does not support Python 3.10.