Can not use multiply gpus for bottom-up feature extraction about lxmert HOT 6 OPEN

Please refer to here. Thanks.

from lxmert.

Currently, the original CAFFE implementation (originated from Ross's py-faster-rcnn repo) could not do batch-wise inference. Because they kept the width-height ratio during inference.

A possible solution is to use detectron2/mmdetection as the backbone extractor. These libraries add the support of batch-wise inference. However, I do find that D2's batch-wise mode would decrease the score of downstream tasks because the features and RoI's are not perfectly aligned.

from lxmert.

Hi, Thanks for authors great doker environment for bu feature extraction.
However, I find it cannot use the multiply GPUs by using CUDA_VISIBLE_DEVICES=0,1,2,3 in docker？
And I check the extract.py file, it seemed that it does not support the multiply gpu extracting?

Or am I missing anything?
Thanks~

Hi @Wangt-CN, can you share me /workspace/features/extract_nlvr2_image.py? I follow this issue#79.

from lxmert.

Hi, Thanks for authors great doker environment for bu feature extraction.
However, I find it cannot use the multiply GPUs by using CUDA_VISIBLE_DEVICES=0,1,2,3 in docker？
And I check the extract.py file, it seemed that it does not support the multiply gpu extracting?

Or am I missing anything?
Thanks~

Hi @Wangt-CN ，I am also interested in installing Detectron (used by VisualBERT) or Detectron2.

from lxmert.

Currently, the original CAFFE implementation (originated from Ross's py-faster-rcnn repo) could not do batch-wise inference. Because they kept the width-height ratio during inference.

A possible solution is to use detectron2/mmdetection as the backbone extractor. These libraries add the support of batch-wise inference. However, I do find that D2's batch-wise mode would decrease the score of downstream tasks because the features and RoI's are not perfectly aligned.

Hi @airsplay thanks! Seems results of these feature extraction tools are comparable and consequentially, other feature extraction tools.

By the way, Any method no need of caffe/caffe2? although I personally can install caffe under this case of LXMERT.

from lxmert.

Currently, the original CAFFE implementation (originated from Ross's py-faster-rcnn repo) could not do batch-wise inference. Because they kept the width-height ratio during inference.

A possible solution is to use detectron2/mmdetection as the backbone extractor. These libraries add the support of batch-wise inference. However, I do find that D2's batch-wise mode would decrease the score of downstream tasks because the features and RoI's are not perfectly aligned.

OK, I think batch-wise inference can be resolved by following detectron2_mscoco_proposal_maxnms.py and I writes down codes for customized images here. I also follow issue#7 in airsplay/py-bottom-up-attention.

from lxmert.