search

aigzhusmart / Slim-UNETR

An Efficient, High-Quality 3D Segmentation for Medical Image Analysis with Constrained Computational Resources
MIT License
35 stars 4 forks source link

infenrence problem #5

Open sawadikakawayi opened 2 months ago

sawadikakawayi commented 2 months ago

Why is my Dice score only 78% and HD95 over 30 when running infenrence?

laaaa123 commented 2 months ago

I conducted experiments on MSD2019 Task-01 Brain Tumor dataset, and the same problem appeared, and good data results could not be obtained according to the inference. Here are the results of the train and inference data. image image Do you have the same problem as me？

sawadikakawayi commented 2 months ago

我是在Brats2021跑的，按照实验配置，根本就跑不起来。我把batch size设成4，结果能到91。但是那个inference确实有问题，最高只能到78，有一个办法是，你把hd数和dice系数同时放到main中计算，确实能得到挺好的效果，但hd数计算又会大幅度拖慢速度，你可以把训练好的模型加进去只跑一轮，也能出结果

At 2024-09-06 17:21:17, "laaaa123" @.***> wrote:

I conducted experiments on MSD2019 Task-01 Brain Tumor dataset, and the same problem appeared, and good data results could not be obtained according to the inference. Here are the results of the train and inference data. image.png (view on web) image.png (view on web) Do you have the same problem as me？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

laaaa123 commented 2 months ago

请问您在训练模型时、cpu占用率会很高吗。 因为我在训练时、会出现gpu忽高忽低的情况（有时甚至只有百分之10左右），我在网上查找问题、说可能与cpu有关。

sawadikakawayi commented 2 months ago

不会

---- 回复的原邮件 ---- | 发件人 | @.> | | 日期 | 2024年09月06日 20:07 | | 收件人 | @.> | | 抄送至 | @.>@.> | | 主题 | Re: [aigzhusmart/Slim-UNETR] infenrence problem (Issue #5) |

请问您在训练模型时、cpu占用率会很高吗。 因为我在训练时、会出现gpu忽高忽低的情况（有时甚至只有百分之10左右），我在网上查找问题、说可能与cpu有关。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

bhaswara commented 3 weeks ago

Any solution on this? @sawadikakawayi @yunhaoli24