Hi ,I was working on a transformer based Ai model which is an time series sequence classifier with also using and encoder and decoder .this time series transformer unlike other models like gpt which are decoder only and basic vision models use encoder only ,my model uses both halves and the encoder ingests the past values and processes 15 frame history matrix of coordinates (x,y,area) to check for the speed,patterns and movement contexts while the decoder takes the multi head attention map outputted by the encoder and cacultes what state it looks like in edge of impact or futute_state also and also uses cross-layer entropy loss or the weigthed croos layer entropy loss in my training code to take and check the diffrence between what the model thinks and waht the label is (collision,warn,safe) and it penalized the model heavily if it is confident in a wrong answer or a prediction. you can also see a picture below of my rc car crashing into the wall and see it flags it as a collision.you can also try out my model by visting my project repository and look at the instructions and in depth detail of my project and the Ai model.