search

zychen423 / KE-VIST

The code and output of our AAAI paper "Knowledge-Enriched Visual Storytelling"
GNU General Public License v3.0
40 stars 9 forks source link

Extracted faster rcnn features of VIST #5

Closed FeatherZhong closed 3 years ago

FeatherZhong commented 3 years ago

Thank you again for such an excellent work, When I used the code to extract VIST image features at stage 1, I found that my GPU 2080Ti did not have enough memory which need 12GB to run the code. I tried to use CPU for calculation, but it took too long, about half a year. I would really appreciate if you could share the extracted features of VIST.

joe32140 commented 3 years ago

Hi,

Have you tried running on your gpu? The original repo said you need 12G to train the model. I remembered it can run the model with less memory.

FeatherZhong commented 3 years ago

Thank you for the response, Yes, I tried running code on my GPU, and I got the this error message Check failed: error == cudaSuccess (2 vs. 0) out of memory I found a issue in the original repository that mentioned the need for 12GB of GPU memory was also required for eval.

FeatherZhong commented 3 years ago

I try to modify the NCCL setting in cafffe and finnally extract the features using around 10GB memory. Thank you very much for your help!