Closed Belval closed 5 years ago
Note: Still need to fix the multiprocessing part so CUDA memory is fully freed on exit.
I am not able to reproduce the initial CUDA memory not free error so I will merge and handle it if someone creates an issue with a reproducible example.
Using the new generators from TextRecognitionDataGenerator, I skip the load-from-disk part of the data manager to use the generators directly, this is both faster and requires less RAM.