Performance issue in /deeppavlov/core/common (by P3)

[DLPerf]

Hello! I've found a performance issue in /deeppavlov/core/common/check_gpu.py: with tf.Session(config=sess_config)(here) is defined in the function check_gpu_existence(here) which is called in the function cudnn_gru_wrapper(here) which is called in the function cudnn_bi_gru(here) which is repeatedly called in the loop for n in range(n_stacks)(here).

tf.Session being defined repeatedly could lead to incremental overhead. If you define tf.Session out of the loop and pass tf.Session as a parameter to the loop, your program would be much more efficient. Here is the Stack Overflow post to support it.

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

[DLPerf]

Hello, I'm looking forward to your reply~

[vaskonov]

Hey @DLPerf sorry for the late response. Unfortunately, deprecated tf 1.x models have very limited support nowadays. I suggest you to describe what are developing exactly, maybe we can find a modern replacement in our DeepPavlov.

[Rushour0]

Hello! I've found a performance issue in /deeppavlov/core/common/check_gpu.py: with tf.Session(config=sess_config)(here) is defined in the function check_gpu_existence(here) which is called in the function cudnn_gru_wrapper(here) which is called in the function cudnn_bi_gru(here) which is repeatedly called in the loop for n in range(n_stacks)(here).

tf.Session being defined repeatedly could lead to incremental overhead. If you define tf.Session out of the loop and pass tf.Session as a parameter to the loop, your program would be much more efficient. Here is the Stack Overflow post to support it.

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

Hey, I think the tf.Session is not called every single time, the if condition if _gpu_available is None: here helps in this. global _gpu_available is used to keep track of any previous results of existence of gpu.

Correct me if I am wrong. Thank you

[IgnatovFedor]

Hello, @DLPerf! We are now using PyTorch instead of TensorFlow, so there should be no problem with this now. If you are still having difficulties with our library, please let us know.