GPU memory use in the layout and formula detection loop increases without being released. This can cause out of CUDA memory errors in large documents. Added garbage collection after each image prediction.
Also added a batch-size parameter, to give some command line flexibility for GPUs with smaller RAM sizes.
wangbinDL
commented
2 months ago
GPU memory use in the layout and formula detection loop increases without being released. This can cause out of CUDA memory errors in large documents. Added garbage collection after each image prediction.
Also added a batch-size parameter, to give some command line flexibility for GPUs with smaller RAM sizes.