Training MGAN?

[ShaneGilroy]

Is it possible to train MGAN using Pedestron?

[hasanirtiza]

In theory, it should be. However, currently we have not added the script but it should be doable. As official MGAN repo does not provide configs for the training part. Based on the paper, my impression is that they are quite similar/ or exactly are that of Adaptive Faster R-CNN, of course they use the visible box/mask additionally.

[ShaneGilroy]

Great thanks.

[123dddd]

Hi, I also found that the config file of the MGAN doesn't provide the training script until now, if it means until now we still cannot train the MGAN using Pedestron? @hasanirtiza

[hasanirtiza]

True, unfortunately the authors never released the training configs for MGAN. You can check on the official repo. However, I did stumble upon this config, it does provide training configs, I have never tried it myself, may be you can see if it works.

[123dddd]

Thanks! Because I saw the pre-trained MAGN model in the benchmark table listed in the README is available, but the config for the training MGAN doesn't exist so I was little confused. How did you generate the pre-trained MGAN model?