Can you provide some output files？
aaaaxxc opened this issue · comments
Can you provide the output files of these two code segments, including VOC and COCO?
python tools/zero_shot_utils.py configs/faster_rcnn_r101_fpn_1x.py --classes seen --load_from ./work_dirs/coco2014/epoch_12.pth --save_dir ./data --data_split train
python tools/zero_shot_utils.py configs/faster_rcnn_r101_fpn_1x.py --classes unseen --load_from ./work_dirs/coco2014/epoch_12.pth --save_dir ./data --data_split test
Hi @aaaaxxc ,
As far as I remember, the extracted features are stored as numpy files. But I'm not currently at the university, so I don't have access to those output files. I can upload them by Monday. Meanwhile, can you tell me what's the issue you are facing in running zero_shot_utils.py?
Hello!
I saw that in another issue you implemented the code in a higher version of torch, but mmdet and mmcv are still giving the error "undefined symbol: _ZN6caffe26detail36_typeMetaDataInstance_preallocated_7E". This seems to be a problem with the CUDA version being too high (CUDA 11.1, which is the minimum version my graphics card can support). [https://github.com/yrcong/STTran/issues/51] I tried to solve this issue by upgrading the versions of mmdet (2.28.2) and mmcv (1.7.2), but mmdet 2.x is not compatible with the models and pre-trained models of mmdet 1.x.
Moreover, the machines that I can use which support lower versions of CUDA have smaller memory, and the memory is insufficient when running COCO code. Therefore, my machine cannot simultaneously meet the requirements of supporting a lower version of CUDA and having high memory.
At the same time, I am also unsure whether different features will affect the final results. In several generative codes I have run, the VOC results seem to differ slightly from those in the paper.
Ah, I see. Yeah, these version incompatibilities can be seriously frustrating. Well then I can upload the files on Monday.
@aaaaxxc Sorry for the delay. Due to some unforeseen circumstances, I reached my institute today. Here are the extracted train and test features for the PASCAL VOC dataset. Unfortunately, I do not have enough cloud storage for uploading the features of COCO dataset, since it is 35 GB in size.
See if this will be enough for you. You're welcome to suggest any alternate way of sending large files over the internet if you want the COCO features as well.