holyseven
commented
4 years ago I got the reported performance using the ImageNet-pretrained parameters. As you already noticed, train_image_size and test_image_size should be the same. For both, I used 864. Sorry that the comment about test_image_size was not correct (updated now).
One more thing, in your training setting, the batch size is 6 (three GPUs). Probably the final result will be different from 78.6 because I used 8.
Hello, thank you for your implementation, it is being used in a comparative experiment in a new paper, but encountered low reproduction accuracy. We used three GPUs to train Cityspace data. The modified settings are as follows:
database = Cityscapes, batch_size = 2, train_max_iter = 50000, snapshot = 10000,
Other parameters have not been modified except the data directory. The training data only includes Cityspace-gtFine. The res_net_101 model is used. After 10 hours of training: [IoUs]: 94.42 & 69.79 & 85.69 & 38.53 & 35.00 & 43.51 & 31.63 & 53.34 & 88.13 & 51.69 & 89.96 & 64.17 & 33.47 & 88.56 & 36.27 & 50.31 & 39.89 & 26.69 & 60.54 & [mIoU]: 56.93 \ This is a big gap from 78.6 in Numerical Results. We really want to know what the training parameters are for 78.6? Does the train_image_size and test_image_size corresponding to Cityspace need to be adjusted?
Looking forward to your reply, thank you!