InceptionV3 inference using DALI

This is the example, that visualizes InceptionV3 inference, where DALI is used for data preprocessing.

Our Inception model accepts images as an input. These images have to be resized to precisely [299, 299] resolution and they have to be normalized. Thanks to DALI you can implement these operations easily (using DALI's python API) and GPU-accelerate them at the same time.

The ensemble

Triton uses model ensembling to put together multiple models into one bulk inferring. This example is such an ensemble, consisting of two models: DALI model for image preprocessing and InceptionV3 model for actual inference. For more info on ensembling, refer to Triton docs.

Autoserialization

This example present the autoserialization feature in DALI Backend. The dali.py file contains a definition of DALI Pipeline. In general, user would need to serialize this pipeline to model.dali file. However, leveraging the autoserialization, DALI Backend will serialize the model itself. For more details about autoserialization, please refer to @autoserialize documentation.

Run the example

To run the example, you have to download InceptionV3 model into the model repo. Please refer to setup_inception_example.sh for details, how to do this.

Remember

As always in DALI Backend case, remember that dali.fn.external_source's name parameter must match with input name provided in config.pbtxt file.