gyhandy / group-supervised-learning Goto Github PK

Thank you for sharing the code. I am trying to reproduce the results with fonts dataset. However, it will take about 1300 hours to train the model on my Tesla P100. I am wondering how long does the training take on your side.
Below is the command to train the model.

python3 main.py --train True --dataset Fonts --dataset_path fonts-v1 --viz_name fonts --viz_on False

Train GZS-Net on datasets used in paper,the command is
python3 main.py --train True --dataset Fonts --dataset_path YOUR_LOCAL_PATH_OF_FONTS --viz_name fonts

in main.py,know the lambda_unsup default 0, so the combine_unsup_loss is not counted in training.

Is this loss useless?

such as
self.gather.insert(images=F.sigmoid(A_recon).data) self.gather.insert(combine_supimages=F.sigmoid(AoC1_2A).data) self.gather.insert(combine_supimages=F.sigmoid(BoC2_2B).data) self.gather.insert(combine_supimages=F.sigmoid(A1Co_2C).data) self.gather.insert(test=F.sigmoid(A2B3D4E5F1_2N).data)

我在Titan X上用ilab_20M训练模型后发现，当我训练约到5w轮的时候模型似乎就已经收敛，但此时效果并不好，重构图像模糊且合成的有些属性也不对，是否该代码与论文中的图片代码有所改变呢？