search

svishwa / crowdcount-cascaded-mtl

Single Image Crowd Counting (CNN-based Cascaded Multi-task Learning of High-level Prior and Density Estimation for Crowd Counting)
MIT License
197 stars 90 forks source link

et_cnt 显示为0 #18

Open zhangsong1213 opened 5 years ago

zhangsong1213 commented 5 years ago

我用自己的数据集 进行训练 当训练到第4轮的时候 et_cnt一直显示为0 BEST MAE: 3.6, BEST MSE: 4.2, BEST MODEL: mtl_fish_2.h5 epoch: 5, step 0, Time: 0.0046s, gt_cnt: 0.0, et_cnt: 0.0 epoch: 5, step 500, Time: 0.0133s, gt_cnt: 11.0, et_cnt: 0.0 epoch: 5, step 1000, Time: 0.0125s, gt_cnt: 4.8, et_cnt: 0.0 epoch: 5, step 1500, Time: 0.0122s, gt_cnt: 8.9, et_cnt: 0.0 epoch: 6, step 0, Time: 0.0030s, gt_cnt: 4.0, et_cnt: 0.0 epoch: 6, step 500, Time: 0.0112s, gt_cnt: 1.0, et_cnt: 0.0 epoch: 6, step 1000, Time: 0.0106s, gt_cnt: 10.7, et_cnt: 0.0 epoch: 6, step 1500, Time: 0.0101s, gt_cnt: 1.6, et_cnt: 0.0

Veefas commented 3 years ago

@zhangsong1213 你好，我用MCNN训练自己的数据集没有问题，切换到cascaded就碰到和你一样的问题，et_cnt一直为0，我觉得应该和分组类别有关，作者默认分组类别为10，我修改成4还是不行。请问你后面解决了嘛？

Veefas commented 3 years ago

Hi @svishwa , when I train my own datasets(about grape counting), the et_cnt value always 0.0. I think it might be caused by the number of groups. I've changed the number of groups in both the CrowdCounter and loss function, which are failed. Could you give me some guidance?

By the way，my preprocessed dataset can be trained successfully on MCNN.

pasquale90 commented 3 years ago

Hello! I am facing the same issue when I try to train my own labeled dataset. Could there be any potential clues on what exactly is happening? producing zero estimating count in my point of view indicates that the model isn't able to identify the count in the image and while stuck there forever may indicate that the model isn't able to learn the pattern in the new data.

At least if any raw assumptions or directions on what changes should we look into would may prove very beneficial..