z-zheng / simplecv Goto Github PK

Hello, you have really done a good job! Congratulation! And would you please open the code about the paper 'Foreground-Aware Relation Network for Geospatial Object Segmentation in High Spatial Resolution Remote Sensing Imagery'? I am quite interested about your work. Thank you very much!

学长好！请问下
def th_intersection_over_union_per_class(y_true: torch.Tensor, y_pred: torch.Tensor, num_classes=None): cm_th = th_confusion_matrix(y_true, y_pred, num_classes) sum_over_row = cm_th.sum(dim=0) sum_over_col = cm_th.sum(dim=0) diag = cm_th.diag() denominator = sum_over_row + sum_over_col - diag iou_per_class = diag / denominator return iou_per_class
这一段的
sum_over_col = cm_th.sum(dim=0)
是不是有问题呀？
应该是
dim=1吧~

学长好！请问下
def th_intersection_over_union_per_class(y_true: torch.Tensor, y_pred: torch.Tensor, num_classes=None): cm_th = th_confusion_matrix(y_true, y_pred, num_classes) sum_over_row = cm_th.sum(dim=0) sum_over_col = cm_th.sum(dim=0) diag = cm_th.diag() denominator = sum_over_row + sum_over_col - diag iou_per_class = diag / denominator return iou_per_class
这一段的
sum_over_col = cm_th.sum(dim=0)
是不是有问题呀？
应该是
dim=1吧
不好意思刚不小心关闭了提问，重新发一下~

data=dict( train=dict( type='NewPaviaLoader', params=dict( training=True, num_workers=0, image_mat_path='PU_gyh.mat', gt_mat_path='PU_gt.mat', num_train_samples_per_class=30, sub_minibatch=20 ) ), test=dict( type='NewPaviaLoader', params=dict( training=False, num_workers=0, image_mat_path='PU_gyh.mat', gt_mat_path='PU_gt.mat', num_train_samples_per_class=30, sub_minibatch=20 ) ) ),
When I try to adjust the sub_minibatch parameter to test the effect of "GS2 sampling strategy", I found that Houston can also converge and get 83% OA without GS2 sampling strategy.
And I found that when I adjust the "sub_minibatch" in train dict, the effect on accuracy is small, but the "sub_minibatch" in test dict have great effect on accuracy.
Could you please tell me what's the function of "sub_minibatch" in test dict? Cuase I think there is no need to set minibatch in the test process.
Thank you very much!!!

Excellent work from RS-IDEA 😁 Actually, considering all the well maintained codes and carefully designed data structures, I've noticed that the document appears no exhaustive, though concise enough. Hence, is it possible to add Wiki or some more detailed manuals for the convenience of users?

BTW, a dead link reported here:
The link to USAGE.md in README.md actually directs to a Not Found page.

Can you open your source code? I had doubted it for some concern.