The mIoU is low

[jackieyung]

Thanks for your sharing @dong-x16 .

I have an issue. When I train my model without the prior channel, the mIoU of the model can reach 94%, but when I add prior channel, the mIoU of the model is much lower than without the prior channel, which even cannot reach 90%.

Could you give me a piece of advice about how to set the parameters of every loss?

Thanks a lot.

[ccly1996]

hello，excuse me, could you display your requirements details?thank you

[jackieyung]

hello，excuse me, could you display your requirements details?thank you

Hi, I just need to improve the mIoU of the model whose input has the prior channel. But I find when I add the fourth prior channel, the mIoU of the model drops a lot.

[ccly1996]

emmm, I am deploying this code, but wrong happened when i install requirements, can you help me?The author may write something wrong in requirements, so i need requirements list.😂

[jackieyung]

emmm, I am deploying this code, but wrong happened when i install requirements, can you help me?The author may write something wrong in requirements, so i need requirements list.😂

python 2.7
PyTorch 0.3.0.post4
Jupyter Notebook
pip install easydict matplotlib tqdm opencv-python scipy pyyaml numpy

[ccly1996]

thanks,i install these requirements ,but something wrong.

[jackieyung]

thanks,i install these requirements ,but something wrong.

Could you paste your detailed wrong message?

[ccly1996]

the wrong is from torchvision, when i import torchvision.models.i think this wrong from the version of torchvision, but the author did not note that.

[decajcd]

Thanks for your sharing @dong-x16 .

I have an issue. When I train my model without the prior channel, the mIoU of the model can reach 94%, but when I add prior channel, the mIoU of the model is much lower than without the prior channel, which even cannot reach 90%.

Could you give me a piece of advice about how to set the parameters of every loss?

Thanks a lot.

Have you resolved it?

[shivSD]

@jackieyung @dong-x16
When you are training with Prior are you using Images (EG1800 or Supervisely) or video, Correct me if i'm wrong while training with Prior channel network will be receiving previous mask information, So during training, there should be some correction between the current frame and previous frame right, we need to train the prior with video.

[dpcq]

Thanks for your sharing @dong-x16 .

I have an issue. When I train my model without the prior channel, the mIoU of the model can reach 94%, but when I add prior channel, the mIoU of the model is much lower than without the prior channel, which even cannot reach 90%.

Could you give me a piece of advice about how to set the parameters of every loss?

Thanks a lot.

hello when you train you model ,how many epoch do you choose? i have to train 2000 epoch to get miou 95%, do you have the question