Training PSPNet

[DonghyunK]

Hi,

I am trying to train PSPNet50.

With a Titan X, I could only train it with the batch size of 1. I cannot train it with the batch size of 2 even using 2 Titan GPUs.

Could please let me know how many gpus you used to train PSPNet and how many gpus are needed to train PSPNet successfully?

Thank you so much.

[wistone]

The experiments with ResNet101 are trained with 4 GPUs. Actually our training code contains many memory optimization so it needs less memory.

[DonghyunK]

@wistone Thank you for the reply.

Could you please let me know how do you optimize memory??

Thank you.

[wistone]

https://github.com/dmlc/mxnet-memonger

You can refer to the memory optimization in MXNet

[DonghyunK]

@wistone

Did you train a model using MXNet and then convert the MXNet model into Caffe model??

Thank you.

[wistone]

We implement this method in our training platform in caffe.

[DonghyunK]

@wistone

Is it publicly available?

If not, could you please let me know how you implement this method in Caffe?

Thank you

[wistone]

No it is not public. You can read the paper for the theory, and refer the code in MXNet. It is doable.

[LearnerInGithub]

@hszhao @wistone Could you explain what's mean of the three accuracy output on training phase? There only one SegAccuracy, so I think there should one accuracy output, why three ones occurs? example output: I0605 10:56:09.200160 16167 solver.cpp:245] Train net output #0: accuracy = 0.971667 I0605 10:56:09.200170 16167 solver.cpp:245] Train net output #1: accuracy = 0.873694 I0605 10:56:09.200176 16167 solver.cpp:245] Train net output #2: accuracy = 0.803811

[balloch]

Wait, @LearnerInGithub , where did you get training files?!

[ThienAnh]

@DonghyunK Can you public script for training?