imag2shape about o-cnn

[wang-ps]

According to your description, it seems that the normal regression is not good enough.
Please send the one of the generated octree file and the trained caffe model to me ([email protected]) so that I can figure out the reason.

[FredFloopie]

Thank you very much for looking! I have sent the caffe model and octrees to your email.

[wang-ps]

I saw your results. According to your caffemodel, the network is only trained for 22000 iterations. Please follow the solver we provided (https://github.com/Microsoft/O-CNN/blob/master/caffe/examples/ao-cnn/image2shape.solver.prototxt), and train the network for 350000 iterations.

[FredFloopie]

yes however, i changed batchsize to 384, so therefore i reduced iterations from 350,000/10 = 35,000.    Then i also trained on multi-gpu (2 gpus)  with the --gpu=0,1 commandline option,  so i reduced the iterations in half, to 22,000 iterations.

[wang-ps]

If the hyper-parameters are changed, the results may be quite different.
Please follow the solver and paramters we provided to reproduce our results.

[FredFloopie]

Ok, running it now without batch_size increase + iteration decrease. I.e., i'm using the exact solver parameters unchanged.

However, I notice in the solver parameters, the net source is image2shape_resnet.train.prototxt, while in the repository, there is only image2shape.train.prototxt. So I changed it to image2shape.train.prototxt.

Will this be ok?

[wang-ps]

Yes, it is a typo, and I have fixed it. Thank you!

[FredFloopie]

Hi, so I ran at batchsize 32 with all your default hyper-parameters and it works great! It's a duplicate of your paper.

However if I try different batch sizes, the "normal regression" as you say seems to fail and I get the problem I mentioned, only horizontal and 45 degree patches.

Is this to do with the the learning rate and step value hyper-parameters? e.g. do the following hyperparameters need to be refined to increase batchsize and have the normal regression proceed properly:
base_lr: 0.1
momentum: 0.9
weight_decay: 0.0005
lr_policy: "multistep"
gamma: 0.1
stepvalue: 150000
stepvalue: 300000
stepvalue: 350000

or is it something more fundamental about the batchsize?
Thanks again.

[wang-ps]

As far as I know, it seems that there is on solid theory about the batchsize.
If the batch size is changed, the learning rate and step value should also be properly tuned, and perhaps better results can be acheived.

For our network, you can try to remove the caffe layers whose type are "Normalize". The "Normalize" layers are used to normalize the length of normal to be 1, I have observed that after removing the "Normalize" layers, the normal regression converges faster. You can add back these "Normalize" layers in the testing stage.

If you want to use multi-GPUs, this paper (https://arxiv.org/pdf/1706.02677.pdf) decribes some guidelines to tune the batch size and learning rate based the single-GPU paramters.