why should fix parameters in training maskrcnn?

[gaosanyuan]

In mx-maskrcnn, there are FIXED_PARAMS and FIXED_PARAMS_SHARED in the config.
It it obvious that FIXED_PARAMS_SHARED is needed, but why should need FIXED_PARAMS?
I tried in faster rcnn, the result would be better if FIXED_PARAMS is set empty.
In default, the FIXED_PARAMS is: ['conv0', 'stage1', 'gamma', 'beta'].

[hdjsjyl]

@gaosanyuan , I met similar problem to you. I want to ask you where this parameter 'FIXED_PARAMS_SHARED' is used? What's the meaning of this part? Any advice will be appreciated. Thank you very much.

[gaosanyuan]

Generally, once a parameter is fixed, it will not be updated in the training process. In mxnet framework, the fixed parameter works as a parameter of the Module class. Please refer to mxnet.module.Module. @hdjsjyl

[hdjsjyl]

@gaosanyuan , I get it. Thanks

[zl1994]

@gaosanyuan In the last train_maskrcnn stage, it was initialized with rpn2+rcnn1. But I found it actually be initialized with rcnn1 for the rpn2's feature extraction part was initialized with rcnn1 and fixed during the training. Is that right?