search

aimerykong / Recurrent-Scene-Parsing-with-Perspective-Understanding-in-the-loop

CVPR2018 - scene parsing network regulated by geometric prior
https://www.cs.cmu.edu/~shuk/recurrentDepthSeg.html
37 stars 7 forks source link

Can I ask you how you train the base model? #7

Closed stephenjia closed 7 years ago

stephenjia commented 7 years ago

Hi Shu,

Can I ask you how you train the base model? What are the hyperparameters for training from deeplab resnet model file to get a good base model as mentioned in the paper? I check the given checkpoint and find that you set learning rate to 0 for layers in base model and also pyramidpooling layers.

Thanks a lot! Best, Xu

aimerykong commented 7 years ago

Hi, Xu, For the model you see, I use pspnet as a part of base model. Hope this helps!

Regards, Shu

stephenjia commented 7 years ago

Hi Shu,

Thank you very much for the quick reply. I am not sure whether my understanding is correct. Does that mean in base model you use the same ResNet backbone as in PSPNet, and only update the parameters of two newly added conv layers but not the ones of ResNet backbone? I find in the model file learning rates of the two newly added conv layers are also set to 0. I have another question. I notice that the two newly added conv layers are also followed by batch normalization and relu layers. How do you get the parameters and statistics for these two bn layers, especially when batch size is set to 1?

Thanks a lot.

Best, Xu

aimerykong commented 7 years ago

hi, Xu,

As for the model I release, I copy the resnet-like backbone of pspnet, and build new layers which are randomly initialized. Of course, other pre-trained backbones can also be used like DeepLab. In the new layers, I also build batch normalization. To train it, I adopt a stage-wise procedure. First, I freeze the backbone and only train new layers. It's interesting that even I set batch size as one, feeding one image, I have no problem in training the batch normalization layer. Once the new layers are trained to a reasonable point, I include the backbone and fine-tune. The stage-wise training procedure is also adopted when training the depth-aware gating module and recurrent module.

If you have further questions and findings, I'm happy to discuss.

Regards, Shu

On Thu, Aug 24, 2017 at 1:22 AM, stephenjia notifications@github.com wrote:

Hi Shu,

Thank you very much for the quick reply. I am not sure whether my understanding is correct. Does that mean in base model you use the same ResNet backbone as in PSPNet, and only update the parameters of two newly added conv layers but not the ones of ResNet backbone? I find in the model file learning rates of the two newly added conv layers are also set to 0. I have another question. I notice that the two newly added conv layers are also followed by batch normalization and relu layers. How do you get the parameters and statistics for these two bn layers?

Thanks a lot.

Best, Xu

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/aimerykong/Recurrent-Scene-Parsing-with-Perspective-Understanding-in-the-loop/issues/7#issuecomment-324567939, or mute the thread https://github.com/notifications/unsubscribe-auth/AGKZJA51fECdv6mQMe6g60uZrwj0sbiqks5sbTLFgaJpZM4PAin6 .

stephenjia commented 7 years ago

Hi Shu,

Thank you so much for your detailed reply. Now I understand better how you trained the model.

Best, Xu