some code problems

[zhouzhen0720]

Firstly, congratulations on being accepted to the NIPS2023 conference!
Here are some issues I encountered while running the code:

train_net.py line 113: self.model.module.criterion.matcher.iter = self.iter
AttributeError: 'RankDETR' object has no attribute 'module'

It seems that the right code is: self.model.criterion.matcher.iter = self.iter

rank_transformer.py line 279: outputs_class_tmp = self.class_embedlayer_idx
TypeError: 'NoneType' object is not subscriptable

The reason for the error is likely to be line 167: "self.class_embed = None"

Looking forward to your reply.

[yifanpu001]

Hi Zhouzhen, I appreciate your interest in our work.

For problem 1, I believe you are running this code using only one GPU. When we create the model in model = create_ddp_model(model, **cfg.train.ddp), the model will be wrapped with DistributedDataParallel if you are using 8 GPUs. The issue occurs when you are only using 1 GPU (see ./detectron2/detectron2/engine/defaults.py, line 70 for more details). To fix this issue, I add a condition to determine whether to use ".module".

Regarding problem 2, I haven't encountered the same situation. Although in line 167 of rank_transformer.py, there is a "self.class_embed = None" in the init method of RankDetrTransformerDecoder, it will be rewrite by some Linear head in rank_detr.py line 148. Could you please provide more details to reproduce this error?

[zhouzhen0720]

@yifanpu001 Thanks for your reply.
I think problem 2 is because I used the configuration file named h_deformable_detr_r50_50ep.py, where as two stage is set to False. As a result, the class_embed is not passed to the decoder. I noticed that this error does not occur in other configuration files that use two-stage settings.

Additionally, I have another question. In line 280 of rank_transformer.py, there is no inclusion of Sl as a learnable embedding. However, in lines 264-269 of rank_detr.py, it is included. Is this difference in the code intentional, or is it designed that way?

[yifanpu001]

Hi, we designed it this way. Inside the QRL, we obtain the confidence score from the classification head without adding any additional learnable logit bias. We have clarified it on page 4 of our paper.

[zhouzhen0720]

thanks a lot!

[yhj-1]

Hi Zhouzhen, I appreciate your interest in our work.

For problem 1, I believe you are running this code using only one GPU. When we create the model in model = create_ddp_model(model, **cfg.train.ddp), the model will be wrapped with DistributedDataParallel if you are using 8 GPUs. The issue occurs when you are only using 1 GPU (see ./detectron2/detectron2/engine/defaults.py, line 70 for more details). To fix this issue, I add a condition to determine whether to use ".module".

Regarding problem 2, I haven't encountered the same situation. Although in line 167 of rank_transformer.py, there is a "self.class_embed = None" in the init method of RankDetrTransformerDecoder, it will be rewrite by some Linear head in rank_detr.py line 148. Could you please provide more details to reproduce this error?

Although "self. class_imbed" will be overridden by a Linear header in line 148 of rank_detr. py, its condition is that it must be "as two stages=True". I am training the "rank_detr_r50_50ep. py" configuration file and do not need "as_two_stage". This problem is still unresolved