About the result of cityscape

[Aktcob]

I really care about the result of cityscape dataset. Especially the model Bisenet/ Pspnet. Hope u can release it.

[Tramac]

Please give me some time.

[Aktcob]

I train bisenet last night. mIoU = 60.9%...orz..

[Tramac]

Hi, I am very sorry for the late reply. It is my responsibility that i have not noted that these code should be change for bisenet:

if hasattr(self.model, 'exclusive'):
    for module in self.model.exclusive:
        if module == 'context_path' or module == 'spatial_path':
            params_list.append({'params': getattr(self.model, module).parameters(), 'lr': args.lr})
        else:
            params_list.append({'params': getattr(self.model, module).parameters(), 'lr': args.lr * 10})

when config: crop_size=1024, lr=0.01, epochs=80, use_ohem=True, mIoU=68.4%, which is still lower than the result in paper, ~74.8%. I will continue to check better parameters or bugs in the code.

[Aktcob]

thx... lets fix it~~~

[Tramac]

Update: mIoU=70.2% with aux_loss=True

[Aktcob]

I trained again with the config u mentioned above. mIoU=70.5% with aux.

[Tramac]

Good job! However, it is still lower than the result reported in paper, maybe there are better parameters or hidden bugs in code.

[RAYRAYRAYRita]

@Aktcob Hello! I am training bisenet on cityscapes. May I ask your base-size of cityscapes? If you would be sharing more about your config, I would appreciate it!