boncey / Flickr4Java

Java API For Flickr. Fork of FlickrJ
BSD 2-Clause "Simplified" License
176 stars 154 forks source link

Keras loop slow down issue #548

Closed slowkimo closed 3 years ago

slowkimo commented 3 years ago

Subject of the issue

I have download a keras sample program to test my computer capability with tensorflow-GPU, but it seems that the program loop will be slow down by hidden memory issue.

Your environment

tensorflow-GPU ver: v2.4 GPU: RTX-3090 CUDA: v11.4.2

Steps to reproduce

The program I used: https://keras.io/examples/rl/deep_q_network_breakout/

I have add time recorder to record the time used in each loop.

from time import time

. . .

start = time()

. . .

if frame_count % update_target_network == 0:

elapse = time()-start

print("Loop time is {}!!!".format(elapse))

start = time()

Expected behaviour

The time used for each loop should not increase as frame count increase.

Actual behaviour

I have star program at 16:41 and it reach 100000 frame count at 22:30, time used is at below: 10000 frame count: 381.39 20000 frame count: 701.10 30000 frame count: 1090.84 40000 frame count: 1490.53 50000 frame count: 1916.36 60000 frame count: 2280.29 70000 frame count: 2674.47 80000 frame count: 3248.64 90000 frame count: 3646.62 100000 frame count: 4091.70

Log

I don't know how to get the log, please tell me if you really need it.