Performance issues with ddrnet
mhmd-mst opened this issue · comments
Hello, I am using ddrnet , i tried several other models with the same dataset i am working on and they improve over time except ddrnet after 10 epoch the miou starts decreasing , Does that mean ddrnet perform bad or is something wrong with the code i should fix?
That is strange. I trained all light-weight models in my repo on face parsing datasets and among them, DDRNet has the best accuracy. Wait for longer epochs and if it is still lower than other models, you can change align_corners
to True
in 'ddrnet.py'. And set aux_weights
to [1, 0.5]
. Furthermore, try to tune hyperparameters.
I tried 50 epochs so maybe it is overfitting right?
Observe the other models' behavior and see if it is similar. It also depends on your dataset size.
I tried bisenetv1 with resnet18 and sfnet with resnetd18 both keep increasing, my dataset is 16k
Your dataset size is already large. Try to train more epochs and do the above tricks I mentioned.