OPTIMASI MODEL DEEP LEARNING UNTUK RASPBERRY MENGGUNAKAN PRUNING DAN KUANTISASI

Abstract

Artificial intelligence with deep learning. Has been widely used at many field especially in computer vision technology. It can  be found in many purpose and variety of device for it’s deplayment. Deep learning is a neural network that is arranged in very deep layers in order to extract information in more detail. However, with the high level of performance can be achieved, in other hand, another problem arises; the need for very large computing resources in the process. This study has objective to find method and strategies for optimizing edge device in order to continue implementing deep learning with good performance even with limited computing resources. This study uses pruning and quantization methods. After the conventional training process, additional treatment is carried out by reducing the layers and quantizing the weights to reduce the fractional numbers on each weight. At the final step, the conversion of the optimized model into a format suitable for Raspberry Pi 4 is carried out. The results of these experiment showed a significant increase in prediction latency of 2.8x faster and a decrease in file size to 13.74% smaller. This is very beneficial in the implementation of deep learning on Raspberry Pi which has minimal memory and computing capacity.