tuofeilun
|
9f01a37ad4
|
Refactor ViTDet backbone and simple feature pyramid (#177)
1. The vitdet backbone implemented by d2 is about 20% faster than the vitdet backbone originally reproduced by easycv.
2. 50.57 -> 50.65
|
2022-09-16 11:03:53 +08:00 |
|
tuofeilun
|
b198c5a81f
|
support DINO algo (#144)
dino_4sc_r50_12e：48.71
dino_4sc_r50_24e：50.53
dino_4sc_r50_36e：50.69
dino_4sc_swinl_12e: 56.86
dino_4sc_swinl_36e: 58.04
dino_5sc_swinl_36e: 58.47
|
2022-08-31 15:18:11 +08:00 |
|
Chen Jiayu
|
5110de7635
|
remove RPNHeadNorm, adapted in mmlab_utils (#90)
* remove RPNHeadNorm, adapted in mmlab_utils
|
2022-06-16 15:53:43 +08:00 |
|
Chen Jiayu
|
e61488d319
|
Support ViTDet algo (#35)
* adapt mmlab modules
* add vitdet
* support conv aggregation
* modify vitdet load pretrained
* support fp16
* modify defaultformatbundle
* modify aug
* bugfix sampler
* bugfix mmresize
* bugfix fp16&nonetype
* bugfix filterannotation
* support dlc
* bugfix dist
* bugfix detsourcecoco
* smodify mmdet_parse_losses
* bugfix nan
* bugfix eval
* bugfix data=nonetype
* modify resize_embed
* support vitdet_conv
* add vitdet_conv init_weight
* add test_vitdet
* uniform rand_another
* uniform use fp16 method
* add test_fp16
Co-authored-by: jiangnana.jnn <jiangnana.jnn@alibaba-inc.com>
|
2022-06-10 21:49:32 +08:00 |