I've run the code with pretrained model and got mIoU>0.8 suceessfully. But I trained it on the VOC2012 dataset with 21 classes without pretrained model and got bad results. I set the following parameters :
ba=8 wd=5e-4 max_iter=50k
I only got the final mIoU of 0.46 ,px_accurancy of 0.82 .I visualized it by tensor board and there is no trend of increase from mIoU in long time .
My environments were :
GPU=Titan Xp
input size=512 ,512 ,3(and I've done nothing with the parameters in the python file except two comment line in deep_model.py)
ba= 8 (I've tried ba=10 with the error of run out of GPU memory ,then I changed ba =8 without the error and trained successfully)
I also changed the wd = 2e-4 ,1e-4 ,5e-5,and I trained with them separately. I've never got a better result than mIoU=0.46 .
I think that different GPUs can only influence the speed of computing .Can you give me some suggestion about my situation? @rishizek .if you need ,I can show you the tensor board and the last evalution personally.
I've run the code with pretrained model and got mIoU>0.8 suceessfully. But I trained it on the VOC2012 dataset with 21 classes without pretrained model and got bad results. I set the following parameters : ba=8 wd=5e-4 max_iter=50k I only got the final mIoU of 0.46 ,px_accurancy of 0.82 .I visualized it by tensor board and there is no trend of increase from mIoU in long time . My environments were : GPU=Titan Xp input size=512 ,512 ,3(and I've done nothing with the parameters in the python file except two comment line in deep_model.py) ba= 8 (I've tried ba=10 with the error of run out of GPU memory ,then I changed ba =8 without the error and trained successfully) I also changed the wd = 2e-4 ,1e-4 ,5e-5,and I trained with them separately. I've never got a better result than mIoU=0.46 . I think that different GPUs can only influence the speed of computing .Can you give me some suggestion about my situation? @rishizek .if you need ,I can show you the tensor board and the last evalution personally.