A machine learning model retains the mapping from the parameter name and its value. It is a type of key-value structure in general. Technically, it can be written in any kind of format that can express structured data, but it is important to make the model portable, so that we can reuse it somewhere different to where the model is trained. Here are the characteristics the portable model format should have:
- Lightweight: Small enough to be stored in limited memory capacity
- Serializable: Sharable through the disk or network I/O
- Compatible: Usable by multiple platforms
Nowadays, the range of platforms where machine learning can run is diverse. A machine learning algorithm is expected to run not only on a typical server-side machine but also on edge devices such as mobile or embedded systems. Even with the limited memory capacity of edge devices, the model...
