search

ZhengyaoJiang / PGPortfolio

PGPortfolio: Policy Gradient Portfolio, the source code of "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem"(https://arxiv.org/pdf/1706.10059.pdf).
GNU General Public License v3.0
1.74k stars 750 forks source link

Low GPU-Usage while training #100

Closed ghost closed 6 years ago

ghost commented 6 years ago

i'm modifying the code for my personal usage for a while now, but encountered very low GPU-usage for training since i cloned the repository. Training uses 2-5% at max. of the GPU (shown in nvidia-smi). It occurs when using a single or multiply processes. Increasing the hard coded values in network.py above 0.2 also has no impact.

Correct drivers are installed, using a gtx 1080 on Ubuntu 18.04. I'm using Tensorflow v. 1.8.0 and NVIDIA UNIX x86_64 Kernel Module 390.67

Has anybody ran into the same problem, or has a solution for it? Thanks

dexhunter commented 6 years ago

what about the memory usage?

ghost commented 6 years ago

Memory Usage was 0% - i figured out that my IDE destroyed my pip-installation of tensorflow. A native installation of tensorflow-gpu solved the problem for me.

GPU usage is now at 55%, Memory at 100% - I'm almost sure the bottleneck was produced by modifications. Thanks for your help, @DexHunter