Nvidia provides the DALI library (Data Loading Library) to help loading data on GPUs. DALI can speedup deep learning using a mix of GPUs and CPUs for data loading process. DALI is capable of overlapping the training and preprocessing stages; thus, reducing the latency and training time. Another advantage is that it provides a unified data loading interface to Deep learning frameworks. Different frameworks have different data loading classes, and it is not possible to use a data loading class from one framework with another one. DALI provides a data loading class that can be used with all well-known Deep learning frameworks.
Below, we present a Jupyter notebook that loads and process data with DALI. Follow the steps below to launch the notebook in the NCI ARE interactive environment.
Launch ARE notebook
- go to the url: https://are.nci.org.au/ and login.
- Open the web form to launch a new Jupyter Notebook.
- Choose "gpuvolta" queue and compute size "4gpu". Also add the necessary storage option: gdata/dk92+gdata/wb00 (You can also append extra storage space as shown below)
- Go to the "Advanced section..." and choose the following:
Modules directories: /g/data/dk92/apps/Modules/modulefiles
Modules: NCI-ai-ml/22.11
- Wait for the JupyterLab to start.
Launch the Notebook
- Navigating to the directory "/g/data/dk92/apps/NCI-ai-ml/22.11/examples/dali/" and open the Jupyter notebook "dataloader.ipynb".
At this stage, interactively run and experiment with DALI.
Note: you are running the notebook in the read-only mode. If you want to revise the notebook, please copy the notebook from "/g/data/dk92/apps/NCI-ai-ml/22.11/examples/dali/dataloader.ipynb" to a location that you have write permission.Troubleshooting
If you have problem running the notebook, then, make sure that you have access to all necessary files.
If you are still having problem, then, please let us know.