Industrial ML five items
- End-to-end industrial monitoring solutions including sensors, gateways, machine learning and more: Monitron
- Hardware-free AWS machine learning detection anomaly tool: Lookout for Equipment
3. Equipped with a CV program in the camera to monitor the factory environment: AWS Panorama Appliance
4. AWS Panorama Appliance software development kit: AWS Panorama (SDK)
5. CV model for detecting industrial product detection: Amazon Lookout for Vision
AUTOML-NNI
Automatic Feature Engineering
Feature engineering is the pre-step of applying classic machine learning algorithms. Through feature engineering, the machine learning process can be faster and better results can be obtained.
The hyperparameter tuning function of NNI described above can be directly applied to various subfields of feature engineering such as feature enhancement and automatic feature selection. To make it easy for novices, NNI also has built-in automatic feature selection algorithms based on gradients and decision trees, and also provides interfaces to extend other algorithms.
The NNI team also compared the effect of automatic feature engineering, and tested it on popular colon-cancer, gisette, rcv1, neews20.binary, real-sim and other data sets. We found that if only the top 20 of the thousands of features are selected, the decision tree can basically achieve better results, and if more features are selected, better results will be obtained.
Neural Architecture Search (NAS)
Neural Architecture Search (NAS for short) achieves better performance by automatically searching the network structure, and a large number of research results have emerged this year. There are various NAS algorithms and implementations vary.
In order to promote the innovation of NAS, we have explored the abstraction and implementation of NAS algorithms, so that users can not only directly apply the algorithm on their own data sets, but also easily compare the effects of different NAS algorithms horizontally.
NNI realizes ENAS, DARTS, P-DARTS algorithms, and provides the interface of one-shot algorithm. In addition, classic search methods such as Network Morphism are also supported.
model compression
With the development of deep learning, the model is getting bigger and bigger. Although the accuracy has been greatly improved, the larger model size not only affects the inference speed, but also requires higher hardware requirements for deployment. Therefore, model compression is also a hot topic.
The main model compression algorithms can be divided into two categories, one is the pruning algorithm to compress the network structure, and the other is the quantization algorithm to reduce the model accuracy. NNI currently provides 9 model compression algorithms such as AGP, L1Filter, Slim, Lottery Ticket, FPGM, QAT, and DoReFa. Users can also implement their own compression algorithms through the model compression interface of NNI according to their needs.