search

weii41392 / AQT

AQT: Adversarial Query Transformers for Domain Adaptive Object Detection
Apache License 2.0
13 stars 1 forks source link

Backbone Alignment #2

Open michaelku1 opened 2 years ago

michaelku1 commented 2 years ago

Hello, Thanks for your amazing work. I was wondering if you had performed ablation studies on using backbone alignment? I assume the performance did not improve after using all four alignment strategies? Thanks.

helq2612 commented 2 years ago

Hi @michaelku1 , I have tried to enable four alignments on cityscapes to foggy cityscapes. You are right, I did not see much difference comparing to the model trained without the backbone alignment. But all the AP50s (w/ or w/o backbone alignment) I got are around 36. Can you reproduce this work?

michaelku1 commented 2 years ago

Hi @michaelku1 , I have tried to enable four alignments on cityscapes to foggy cityscapes. You are right, I did not see much difference comparing to the model trained without the backbone alignment. But all the AP50s (w/ or w/o backbone alignment) I got are around 36. Can you reproduce this work?

I was able to reproduce a mAP of 45.3 on city2foggy with the default config.

helq2612 commented 2 years ago

Thank you @michaelku1 ! How many GPUs are you used in the training?

michaelku1 commented 2 years ago

ow many GPUs are you used in the traini

I was training on 4 GPUs.

weii41392 commented 2 years ago

Sorry for my late response!

Thanks for your interest in AQT, @michaelku1! Frankly, we didn't assume AQT is exempt from backbone alignment. After all, AQT focuses on the transformer instead of the CNN backbone. Their objectives are subtly different.

Unfortunately, as the project has been finished for quite some time, I can neither recall nor retrieve the experiments not specified in the paper.

peterstratton commented 1 year ago

@helq2612 @michaelku1 I'm also trying to reproduce this, but I get a mAP of 44.3. Do both of you have the exact same environment setup as described in the README?

yuguoliua commented 2 months ago

ow many GPUs are you used in the traini

I was training on 4 GPUs.

Hello, senior. I am a beginner, and I have been trying to reproduce this algorithm recently. I have four GPUs, but I am unable to perform multi-GPU training using the instructions in the readme file. Could you please let me know how you set it up when you worked on it? I would really appreciate your response. Thank you!