OBJECT DETECTION VIA ALTERNATIVE TRANSFORMER
Neychev Radoslav, Arman Stepanyan · COLLECTION OF SCIENTIFIC ARTICLES / ԳԻՏԱԿԱՆ ՀՈԴՎԱԾՆԵՐԻ ԺՈՂՈՎԱԾՈՒ · 2023
Transformer is a great building block for state-of-the-art models for every direction in artificial intelligence (AI). In natural language processing, GPT3 is one of the leading language models, and its fine-tuning leads to the creation of automation products in various fields (chatting, analysis of data, content generation, etc.). In computer vision, it is DALLEE-2 with its fantastic capability to generate realistic art images with the desired style. In reinforcement learning (RL) decision transformers [1] are widely used and achieve great results in such RL baseline games as ATARI, Key-To-Door tasks, etc. Even though modern transformer blocks for end-to-end object detection tasks converge very slowly, which makes the training process computationally hard. We introduce alternative transformers improving architecture and training process, reducing convergence and training time with achieving same results in object detection tasks. Training process is parallelized, and a loss function is modified to increase model’s capability in multiple tasks. Finally, this architecture can be used as a building block in other models, improving their performance