How can I get a multi-scale proposal

[ZhongkaiZ]

Ask a question again.
Through experiments, the model makes proposals in different positions by sliding the window. The length of the proposal is basically the same. This phenomenon is due to the setting of positional embedding or other reasons. I added the start and end score convolution to the overall model to train together. @tony2016uestc

[JackyTown]

'The length of the proposal is basically the same.'
I've never been in this situation before, and I think it has nothing to do with positional embedding.

[ZhongkaiZ]

id	f-init	f-end	score	video-name	duration
59272	8941.171587	8977.378497	0.617192626	youtu_fps_5	36.2069108
31496	261.2841475	297.6285645	0.605475843	youtu_fps_5	36.344417
59144	8901.180378	8937.332202	0.603523493	youtu_fps_5	36.151824
57480	8381.318635	8417.45	0.59566766	youtu_fps_5	36.1313652
31880	381.2441874	417.521347	0.592478871	youtu_fps_5	36.27715969
57736	8461.184256	8497.324283	0.592164576	youtu_fps_5	36.14002676
59240	8931.141665	8967.450015	0.590808988	youtu_fps_5	36.30835028
57448	8371.069671	8407.376627	0.582358301	youtu_fps_5	36.30695572
5672	8932.751745	9112.894943	0.581983328	youtu_fps_1	180.143198
5320	8305.016166	8486.745621	0.580924988	youtu_fps_1	181.729455
31752	341.2028193	377.4039001	0.580215931	youtu_fps_5	36.2010808
59112	8891.13182	8927.524298	0.57909888	youtu_fps_5	36.39247856
51720	6581.16676	6617.600854	0.578297973	youtu_fps_5	36.43409481
58280	8631.045631	8667.365689	0.577133536	youtu_fps_5	36.32005787

[JackyTown]

Please check the log and make sure that loss is dropping.

[ZhongkaiZ]

'The length of the proposal is basically the same.'
I've never been in this situation before, andI think it has nothing to do with positional embedding.

Thank you for you reply,I found that the parameters of 'pred_boxes' in a batch are equally distributed.

[ZhongkaiZ]

Please check the log and make sure that loss is dropping.
Train loss is dropping
such as :
"train_class_error": 1.6918702916952688, "train_loss": 1.4451521763076676,
"train_class_error": 1.5234429799695726, "train_loss": 1.3731449029435672
"train_class_error": 1.2739102990058675, "train_loss": 1.3295617185044655,

[JackyTown]

Maybe you can check outputs of starting and ending score conv.

[ZhongkaiZ]

Maybe you can check outputs of starting and ending score conv.

tanks,I will try to transfer the weight of BMN and freeze the update.