Google's recently launched NasNet is the best model in the current image recognition field. Recently, this model has been reproduced, and I have a general understanding of its principle. This model is not designed by humans, but is automatically trained by AutoML, which Google launched a long time ago. The purpose of the project is to achieve "automated machine learning", that is, to train machine learning software to create machine learning software, develop the code layer of the new system by itself, and it is also a neural architecture search technology (Neural Architecture Search technology). However, while AutoML is able to design small neural networks with performance comparable to those designed by human experts, it is still limited to small academic datasets such as CIFAR-10.
1. Thesis principle
The model in this paper is based on AutoML to first perform a neural network architecture search on a dataset such as CIFAR-10, so that AutoML can find the best layer and flexibly stack multiple times to create the final network, and transfer the best learned architecture. into ImageNet image classification and COCO object detection. This is the origin of NasNet.
The composition of NasNet is composed of two network units.
The stacking scheme of the two units is as follows:
2. Thesis practice
The model has been added to the tensorflow slim module, which includes the mobile version and the large version. This article mainly tests for large.
Practical analysis: The effect comparison of NasNet has not been tested yet. It is introduced in tensorflow slim that its accuracy is currently the highest.
However, from a practical point of view, the recognition speed of NasNet is far less than that of other models.