The Training result is higher than the test?

[liuzhidemaomao]

Hello, when I trained the PSPNet50 on the cityscapes dataset, I get 0.8266 miou in the training after 200 epochs, and 0.7764 miou in the test, and 0.6971 miou in the valuation. My environment is pytorch 1.6, 8 Tesla V100 GPUs. Thanks!

[XRodriguez10]

Have you solved it? I'm facing the same situation here.

[XRodriguez10]

The evluation in train.py gives different result compare to test.py. I don't understand why

[hanzhy-code]

The evluation in train.py gives different result compare to test.py. I don't understand why

I meet the same question.

[AndreyStille]

The evluation in train.py gives different result compare to test.py. I don't understand why

I meet the same question.

1. Test. For some unknow reason author decided to f __k up everybody demostating better result than they really are. In test inference is launched per patches and than combined together. So results on standart datasets are higher than in other implementation, and he doesn't give a s__t on perfomance.
2. Train. For some unknow reason author decided to add random crop (instead resize) to transformations in dataset though it can dramatically change results for many applications.

Such is it vanity of github contributors and we need to use their code because have tasks on our bad jobs when we aren't provided with enough time to write it from scratch

[hanzhy-code]

The evluation in train.py gives different result compare to test.py. I don't understand why

I meet the same question.

1. Test. For some unknow reason author decided to f __k up everybody demostating better result than they really are. In test inference is launched per patches and than combined together. So results on standart datasets are higher than in other implementation, and he doesn't give a s__t on perfomance.
2. Train. For some unknow reason author decided to add random crop (instead resize) to transformations in dataset though it can dramatically change results for many applications.

Such is it vanity of github contributors and we need to use their code because have tasks on our bad jobs when we aren't provided with enough time to write it from scratch

I have solved this problem. There is an error in the saving path of the picture in test.py. You should modify it in your own code.
I change 'image_name = image_path.split('/')[-1].split('.')[0]' to 'image_name = image_path.split('/')[-1][:-4]' in test function and cal_acc function. After my modification, I find the calculation result is right.

[LM0223]

您好，我想询问您，您是否可以将test.py中的batch size 调高，而不是只能设置为1，谢谢