One import function does not exist.

[yexiguafuqihao]

from dqrf.ops.functions.ms_deform_attn import SamplingAttention_RA, SamplingEncAttention, SamplingAttention_dec
the function 'SamplingAttention_dec' seems missed in dqrf/ops/functions/ms_deform_attn.py, would you please kindly add this function to the script?

[Hatmm]

Hi,

You can comment out SamplingAttention_dec, it is the vanilla deformable attention from https://github.com/fundamentalvision/Deformable-DETR

[ljpadam]

Hi,

After comment out SamplingAttention_dec, the below error occured:

Traceback (most recent call last):
File "train_net.py", line 338, in
args=(args,),
File "/home/jl2/miniconda3/envs/fastreid/lib/python3.7/site-packages/detectron2/engine/launch.py", line 62, in launch
main_func(*args)
File "train_net.py", line 318, in main
trainer = Trainer(cfg)
File "train_net.py", line 61, in init
super().init(cfg)
File "/home/jl2/miniconda3/envs/fastreid/lib/python3.7/site-packages/detectron2/engine/defaults.py", line 310, in init
model = self.build_model(cfg)
File "/home/jl2/miniconda3/envs/fastreid/lib/python3.7/site-packages/detectron2/engine/defaults.py", line 452, in build_model
model = build_model(cfg)
File "/home/jl2/miniconda3/envs/fastreid/lib/python3.7/site-packages/detectron2/modeling/meta_arch/build.py", line 21, in build_model
model = META_ARCH_REGISTRY.get(meta_arch)(cfg)
File "/home/jl2/code/PED-DETR-for-Pedestrian-Detection/dqrf/dqrf_detr.py", line 36, in init
self.transformer = build_transformer(cfg) # 11 464 220
File "/home/jl2/code/PED-DETR-for-Pedestrian-Detection/dqrf/transformer.py", line 378, in build_transformer
return Transformer(cfg)
File "/home/jl2/code/PED-DETR-for-Pedestrian-Detection/dqrf/transformer.py", line 55, in init
checkpoint_ffn=checkpoint_dec_ffn, rectified_attention=rectified_attention)
File "/home/jl2/code/PED-DETR-for-Pedestrian-Detection/dqrf/transformer.py", line 317, in init
self.multihead_attn = SamplingAttention_dec(d_model, dec_sampling_points=dec_sampling_points,
NameError: name 'SamplingAttention_dec' is not defined.

Could you kindly provide the commands to train the model on coco dataset?

[yexiguafuqihao]

And also, we try to repeat this paper, unfortunately, we can't achieve the performance published in the paper with this open-sourced codes.

[Hatmm]

Hi,

As I said, you can remove SamplingAttention_dec in the transformer architecture. In configs you can use set RECTIFIED_ATTENTION to true.

Here are some results on COCO dataset for reference:

Model	AP
Baseline	46.2
RA	46.4
Baseline with 1000 queries	46.6
Baseline with DQ	46.3

Our DQ module does not work well on COCO