performance problem when training ADE20k about fastfcn HOT 9 CLOSED

@HolmesShuan In my experiment, more scales get bad results

from fastfcn.

The scales in FastFCN is consistent with PyTorch-Encoding under this commit.

from fastfcn.

This is caused by the different distribution between the training image and testing image. With larger crop size in the training phase, the size in the testing phase should also be enlarged.

from fastfcn.

@wuhuikai checkpoint have record the crop size of training phase,which are equal to the crop-size of testing phase.

from fastfcn.

In testing phase, ms and flip are employed. The scales in ms should be manually tuned.

from fastfcn.

So, if i want to train ADE20k on crop-size=520 for better performance, how should i modify the code?

from fastfcn.

Tuning this line on the val set by yourself.

from fastfcn.

@Anikily @wuhuikai
Hi, I wonder why the multiscale testing setting in FastFCN is different from PyTorch-Encoding.

FastFCN:

scales = [0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 2.25] if args.dataset == 'citys' else \
        [0.5, 0.75, 1.0, 1.25, 1.5, 1.75]

Enocding:

scales = [0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 2.25] if args.dataset == 'citys' else \
        [0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0]

Thanks.

from fastfcn.

@Anikily @wuhuikai Thanks for the quick responses.

from fastfcn.