Open shiyanshilaoshu opened 1 week ago
Hello,
Thank you for your attention to detail. Both approaches are valid and can be used based on your preference. The final outcomes are similar, but our final models were trained using the configuration in the code. In the GCE module described in the paper, we mentioned setting the stride of the convolutional layer to match the kernel size (e.g., k=[3,3,5], s=[3,3,5]). However, in the code, we used slightly smaller strides (e.g., k=[3,3,5], s=[2,3,4]) to increase the overlap between features.
Thank you very much for your answer. I'm going to start the experiment now
Hello, I truly appreciate your work. However, I noticed that in the GCE module described in the paper, the stride of the convolutional layer matches the kernel size (for instance, k=[3,3,5], s=[3,3,5]), whereas in the code, it does not (k=[3,3,5], s=[2,3,4]). Could you please clarify which should be followed?![截图 2024-06-26 11-00-40](https://github.com/Ascend-Research/CascadedGaze/assets/164864409/d088feee-8ae9-4f5f-9736-754888f19063)