Questions about training detail

[qiminchen]

Hi,

I am wondering when training the marrnet2 with ShapeNet synthetic images, did you train the Marrnet2 only with ground truth normal, depth or did you train the Marrnet2 with ground truth normal, depth for some epochs and then keep training Marrnet2 with predictions from Marrnet1.

And did you use canon_sup when training Marrnet2? What are the batch_size and epoch_batches (how many data used in training per epoch?) For example, when training chair class, did you use 4 x 2500 training points or 6778 x 20 all training points (ignore invalid point)? Because if I use all the training point per epoch, the Eval loss fluctuate a lot and the reconstruction results are not good.

And are these setups the same for GenRe?

Thanks

[xiumingzhang]

Hi. You should be able to find all these training details in the invocation scripts in https://github.com/xiumingzhang/GenRe-ShapeHD/tree/master/scripts. If there's anything unanswered, please feel free to reopen this issue.

[qiminchen]

Thanks, one more question: did you train the Marrnet2 ONLY with ground truth normal, depth or did you train the Marrnet2 with ground truth normal, depth for some epochs and then resume training Marrnet2 with the predictions from Marrnet1 instead?

[xiumingzhang]

Yes, we trained our Net 2 with ground-truth maps for some epochs and switched to predicted maps for some finetuning. Adding @ztzhang in case I misremember something.

[qiminchen]

Just to make sure that for ShapeNet dataset, the synthetic one, canon_sup=True when training Marrnet2 right? That's what I found on the training scripts. And how many epochs do you switch to predicted maps?

When switched to predicted maps for some finetuning on synthetic ShapeNet dataset, I noticed that https://github.com/xiumingzhang/GenRe-ShapeHD/blob/62fc39b6a4b7c90d3df39b2949f1b0f7a9b3fcb3/scripts/finetune_marrnet.sh#L22-L27 canon_sup=False which does not make too much sense to me.

[xiumingzhang]

canon_sup is the flag indicating whether the supervision is in the canonical pose or not.

If you are talking about MarrNet or ShapeHD, we had this flag on, as these two models predict voxels in the canonical pose.

if you are talking about GenRe, we had this flag off, since GenRe predicts view-aligned voxels.