Recently, I used Keras to do some language modeling with TESLA K40M GPU.
I set up an LSTM-LM, and it works well for a small data set (~1M in txt format), and GPU memory is only consumed for approximately ~300M.
But when i tried to train on a larger data set (~150M in txt format), the program can not be executed, and report "Memory Allocation Error".
it indicated that the program wants to allocate >30GB GPU memory space, which is impossible for any regular GPUs.
The hyper-parameters for small and large datasets are identical(hiddenSize=100,vocabulary=30k).
I don't know whether the memory allocation is also related to the input data size or not.
Hi,
Recently, I used Keras to do some language modeling with TESLA K40M GPU. I set up an LSTM-LM, and it works well for a small data set (~1M in txt format), and GPU memory is only consumed for approximately ~300M. But when i tried to train on a larger data set (~150M in txt format), the program can not be executed, and report "Memory Allocation Error". it indicated that the program wants to allocate >30GB GPU memory space, which is impossible for any regular GPUs. The hyper-parameters for small and large datasets are identical(hiddenSize=100,vocabulary=30k). I don't know whether the memory allocation is also related to the input data size or not.
Thanks in advance.