Peize Sun comments

Results 64 comments of


                                            Peize Sun

Resume started from zero again!

Hi~ Can you provide your command line?

Resume started from zero again!

`python projects/SparseRCNN/train_net.py --num-gpus 1 --config-file projects/SparseRCNN/configs/sparsercnn.res50.300pro.3x.yaml --resume` works well in my case. Do you delete last_checkpoint in output folder?

How the feature re-use is performed

Hi~ The feature re-use means that the feature of previous stage is used in current stage. [code](https://github.com/PeizeSun/SparseR-CNN/blob/main/projects/SparseRCNN/sparsercnn/head.py#L104)

pos_cost_class and neg_cost_class

Hi~ I guess the basic idea comes from Discriminative Loss Function: the same instance lie close together while different instances are repulsed.

About the loss format

Hi~ Actually (cx, cy, w, h) and (x1, y1, x2, y2) can be converted by box_ops.box_cxcywh_to_xyxy and box_ops.box_xyxy_to_cxcywh.

About the loss format

yep~

Question about the Iterative architecture

Hi~ You are right, in this table, we not use the dynamic head and proposal features. For the code of feature reuse, please refer to [issue78](https://github.com/PeizeSun/SparseR-CNN/issues/78#issuecomment-805446458)

DLA34 Backbone Result

Hi~ We will add the results based on modified DLA34 backbone later.

Why the network needs proposal features/boxes?

Hi~ I think your idea is quite creative. It looks even simpler than Sparse R-CNN. Both proposal features and boxes are transformed in more elegant way. Have you carried out...

Hi,Have you verified the model to ONNX, i meet some bugs to onnx.

Hi~ We have not verified yet. But we will !