In the first module, we had an overview on how to develop Deep learning application with Intel® distribution of OpenVINO™ toolkit. We need to take a pretrained model, next prepare it for inference, then we perform inferencing using the Inference Engine component of the OpenVINO™ toolkit.
There are many deep learning frameworks like Caffe, Tensorflow, PyTorch, MXnet, ONNX, Kaldi, Paddle Paddle*, which are widely used in industries. We might want to use many of the models from these frameworks for inference. On the other hand, Intel provides a wide range of hardware devices and platforms like CPU, GPU, and VPU. Optimizing and deploying models from these several frameworks to various hardware is a tedious task. Developing optimized code for each model with each of the hardware will be a challenge. Also the model representation in each of the frameworks might be different. Each of the devices might also have different architecture, instruction sets, and programming models. Learning to program for all these software and hardware variants may take time and expertise.
Instead, if we have a common API to do inference across all of the above-mentioned devices and abstract the hardware for you, then Deep Learning applications can easily be developed and run on multiple devices with no changes in the application code.
Model Optimizer is one of the important components from OpenVINO™ Toolkit which will convert models from various supported deep learning frameworks to a single unified representation called Intermediate Representation.
