trained model provided by baidu drive

[lyxlynn]

Hi, thanks for open-soursing this work.
I used the trained models "LIP_CE2P_trainVal_473.pth" "LIP_CE2P_trainVal_321_681.pth" "LIP_CE2P_train.pth" from baidu drive to predict test images in LIP , but the results are not so good.
So then I used the "LIP_CE2P_train.pth" to evaluate LIP_val images and use the code 'job_evaluat.sh ' you provided . The results are : Pixel accuracy: 0.735310 Mean accuracy: 0.384664 Mean IU: 0.264362

Is the results normal or there is something wrong ?
I wonder whether he three trained models from baidu drive are the well-trained models that can achive the best performance as your paper shows ?

[liutinglt]

@Liuyixuan95
The version of PyTorch we used is 0.3.1. Someone using pytorch 0.4.1 can not reproduce the results. The reason is that the align_corners of upsample function is set to default to False. Maybe you can try to replace all the Upsample(size=output_size, mode='bilinear')" with "Upsample(size=output_size, mode='bilinear',align_corners=True)".

[lyxlynn]

@liutinglt Thanks ! got it