search

TUI-NICR / ESANet

ESANet: Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis
Other
239 stars 52 forks source link

The low mIoU of validation dataset #17

Open Huhaowen0130 opened 3 years ago

Huhaowen0130 commented 3 years ago

I have trained the model for 100 epochs on SUNRGBD with pretrained ResNet34 on Imagenet, and the best mIoU is still 10%. Is it normal?

My batchsize is 16, and the other hyper-parameters are set as default.

mona0809 commented 3 years ago

Have you reduced the maximum epochs with the args or have you just stopped the training after 100 epochs? Our learning rate scheduler reaches the highest learning rate only after 10% of the total epochs. Since we use 500 epochs, the highest learning rate is reached after 50 epochs and then only slowly decreases. So maybe the learning rate is too high. Still, the best mIoU should be higher after training for 100 epochs. Due to limited GPU memory, we never trained with batchsize 16. Do you have the same problems when training with batchsize 8?