sleepychord / cogdata Goto Github PK

[background]
I want to use my own text-image datasets to generate binary format dataset for CogView training in 'https://github.com/THUDM/CogView'. It has been mentioned in that repo the author use this cogdate toolkit to preprocess data.

[question]
Would you please tell me how to organize my raw text-image dataset, and then how to use the cogdata toolkit to generate the target bin file? for example, whether i should name the a text-image pair the same, such as 'a dog sits on the ground.txt' and 'a dog sits on the ground.png', or i should take other forms?

when I run cogdata process --task_id test_task --nproc 2 --dataloader_num_workers 1 --ratio 1 test_ds in the terminal, I got a different result from the guide.

And when I try to view the log via tail -f /mnt/disk/reself/projects/CogView2/cogdata_task_test_task/main_pid_46834.log, it appears as follow:

how can I fix it?

我按照cogview的README,使用vqvae_hard_biggerset_011.pt处理一个1300多张的小型图片数据集,分辨率为256*256,但我使用单卡P100显示

RuntimeError: CUDA out of memory. Tried to allocate 16.00 GiB (GPU 0; 15.90 GiB total capacity; 856.64 MiB already allocated; 14.07 GiB free; 874.00 MiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

所以我比较好奇处理这个小数据集需要多大的显存