TorchSemiSeg
TorchSemiSeg copied to clipboard
charlesCXK
Performance on PascalVOC with default config in voc8.res50v3+.CPS
Hello, I am using 1 Telsa P100 16gb with batchsize 4, with your default config. I only reach about 66% mIOU, different with your report in paper and your log: 73.28
Hi, if you only use 1 GPU, the learning rate should be adjusted according to the batch size (e.g. use the linear scaling rule for lr - bs). However, we recommend using 4 GPUs since the batch size is very important in the semantic segmentation task.
Thank you. I change learning rate decrease to about 0.001 (1/5 your default learning rate) with batchsize 4. But I only got 71% mIOU. I decrease learning rate to 0.0005 (1/10 your default learning rate) with batchsize 4, but result is only stable between 70.5%-71% mIOU.
Hello, I am using 1 Telsa P100 16gb with batchsize 4, with your default config. I only reach about 66% mIOU, different with your report in paper and your log: 73.28
Hello, the log file after my training is red, which cannot be called in the verification process. Have you ever encountered such a situation? (I also train with a single GPU )
@lijiwen97 Hi, I think this issue has nothing to do with the number of GPUs. It seems that something was wrong when saving the checkpoint, e.g., the disk was full at that time.
Hello, I am using 4gpu, and your configuration lr=0.0025, batchsize is 8 for training. But only reached 67% mIOU.
