Closed TangAL0203 closed 4 years ago
Thanks for your question.
Actually, I noticed this issue when I was implementing Dynamic R-CNN. From my viewpoint, adopting the statistics across the whole batch (multiple GPUs) will be more robust. However, I had no quota at that time.
I will add this feature when I am available. Contributions are also welcome!
在代码中,计算rcnn_iou_new和rcnn_error_new只考虑到了单个GPU的样本,感觉这会造成一定统计的不稳定性,请问作者有考虑在多GPU之间进行一个同步吗?也就是rcnn_iou_new和rcnn_error_new的更新会考虑到整个batch的样本,而不是统计IMS_PER_GPU的样本