Some questions

[kerryhhh]

Hi! I would like to ask if the parameters of the pre-trained model you provided were converted from the original author's tensorflow version? Then can you provide the relevant training code? Thank you very much!

[RonyAbecidan]

Hello, Yes the parameters of the pre-trained model provided here are exactly the same from the original implementation released by the author's in tensorflow. However, I don't have access to the training code since the author's didn't share it and, as you may have understood, it is clearly not so simple to reproduce this training.I just reproduced the architecture in pytorch and plugged the weights given by the authors. I still don't know how training mantranet properly but it is pending. According to the paper you have to train the two core modules separately and concatenate them. 

[kerryhhh]

Thanks! In addition, I found that the code you provided cannot run in a multi-gpu environment, and the parameters in the bayar_conv part will report an error in multi-gpu environments.

[RonyAbecidan]

Thanks for noticing me about that issue. I will try to make the code operational in a multi-gpu environment. In the same time if you find a way to make it compatible with a multi-gpu environment please don't hesitate to do a pull request 😉

[RonyAbecidan]

Hello,

I think I fixed the cause of the bug in the multi-gpu setting. Can you try to run again your code ? If it does not work try also to add in the beginning of your code the following lines.

import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0,3"

replacing 0,3 by the GPUs you want to use.

My patch is in the branch patch-12