Open Lujinfu1999 opened 1 year ago
It seems about 4000M+ memory under Tesla V100 32G GPU if I remember correctly. You can store the image feats of training set to reduce the CUDA memory usage.
Thank you very much! I use ruotian luo's code [ImageCaptioning.pytorch(https://github.com/ruotianluo/ImageCaptioning.pytorch) and use swin-transformer instade of bottom-up feature when train, and it can run about 9G memory for SCST. But i use PureT for SCST it takes more than 10G CUDA memory,does it has some diffences when SCST traning?Could you tell me if you know it!Thank you for your patience!
Sorry I am not sure why. If you just train the model using pre-extracted Swin feats, the CUDA memory should not be too high. And I remember wrong, 4000M+ is the training under XE. I just tried it again, the SCST training is about 10000M+. If deleting the backbone of Swin-Transformer model and directly training using the image feats as input, the CUDA memory should be lesser enough, just about 5000M.
Thanks a lot! I will try your advice for training.Thank you very much for your patience again!
Dear Author! Sorry to bother you! I have tried your suggestion and used swin-transformer to extract image features, but it got 2-3 CIDER points lower than use image just in XE training. What's more strange is that it got same score in several experiments. I modificated [coco_dataset.py] and [data_loader.py] when reading features and delete the backbone, I not sure where the error occurred.Could you share your code about using pre-extracted Swin-feats(mainly coco_dataset.py and data_loader.py) if it's convenient for you. Thank you very much!
Hello!Author! I want to reproduce your experimental code,but when i run code for SCST, it occures 'cuda out of memory'. I want to know how much CUDA memory is needed for SCST,could you tell me? Thank you!