Questions about the amount of data collected

[handsomeandbeautiful]

First of all, thank you for your wonderful work. When I reproduced this paper, I had some doubts about the amount of data collected. About 200k of data was collected at the beginning, but the performance gap in town05long was not small. Now we have about 350k of data collected, and I don't know if that's enough. It was mentioned in the interfuser work that there was only a slight drop in performance when 500k of data was used. I would like to know approximately how much data you used in this work, because collecting data is really a time-consuming process. Looking forward to your reply.

[tulerfeng]

Thank you for your interest in our work. As you mentioned, collecting data is indeed a time-consuming process. In our experiments, we used approximately 800k samples for training.

Although we did not conduct ablation study on varying data sizes, my experience suggests that using 500k samples can be sufficient (but more data could be better) . You should train more epochs if your data is less, for example, 15~20 epochs.

Additionally, It's important to ensure that your dataset is uniformly distributed across different towns, weather conditions and route lenth.

Finally, for a more robust evaluation, it is advisable to conduct multiple assessment runs. This helps to smooth out random fluctuations.

[handsomeandbeautiful]

Thank you for your interest in our work. As you mentioned, collecting data is indeed a time-consuming process. In our experiments, we used approximately 800k samples for training.

Although we did not conduct ablation study on varying data sizes, my experience suggests that using 500k samples can be sufficient (but more data could be better) . You should train more epochs if your data is less, for example, 15~20 epochs.

Additionally, It's important to ensure that your dataset is uniformly distributed across different towns, weather conditions and route lenth.

Finally, for a more robust evaluation, it is advisable to conduct multiple assessment runs. This helps to smooth out random fluctuations.

Thank you for your prompt reply. I will try your suggestion