shuffle(mbq) is extremly slow

1 次查看(过去 30 天)
I run the Object Detection with yolov3 Deep Learning with a custom dataset with only 126 Images. The epochs themself are running in a reasonable speed but the function calls
reset(mbqtrain)
shuffle(mbqtrain)
takes unreasonable 25 sec for the reset and 111 sec for the shuffle. (both tested with a simple tic toc). This slows the trainingtime extremly down. The System runs on a 6 Core i5-10600, a Nvidia 1050 Ti and 16 GB RAM.
I didn't yet test the times with the VehicleSet but I realized also a pause between epochs.
Any Ideas how adress this issue?
  3 个评论
Matti Kaupenjohann
编辑:Matti Kaupenjohann 2021-9-2
which('shuffle')
says: datastoreio\+matlab\+io\+datastore\@ImageDatastore\shuffle.m
and
doc shuffle
shows the documentation for datastores.
But I know that their is a documentation which shows the syntax
shuffle(mbq)
which seems to be a different shuffle function. If they are actually different functions thats kinda weird.
I also tested it now in the other direction:
mbq.reset();
mbq.shuffle();
which runs faster (both around 5 seconds) but then the next(mbq) or the mbq.next() method calls takes much longer for the first time called and then the reset and shuffle method again takes much longer (25 sec and 115 sec) in the second epoch.
Matti Kaupenjohann
In the meantime I realized, that the size of image matters a lot, which seems kind of weird for methods, which handle the indexing (Or IMHO it should only handle the indexing). Also the loading of my small images (399 x 244) should not take 25 seconds...

请先登录,再进行评论。

回答(0 个)

类别

Help CenterFile Exchange 中查找有关 Image Data Workflows 的更多信息

标签

产品


版本

R2021a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by