Created by: wangguibao

bias_attr is a function parameter to convolution layers like sequence_conv, conv2d, conv3d etc.

def conv2d(input,
           num_filters,
           filter_size,
           stride=1,
           padding=0,
           dilation=1,
           groups=None,
           param_attr=None,
           bias_attr=None,
           use_cudnn=True,
           use_mkldnn=False,
           act=None,
           name=None)

According to this issue its acceptable to set bias_attr in a lot of ways, especially when set to True/False it will turn on/off using bias entirely.

Today when one user asked if setting bias_attr to None (not True or False) will turn off using bias, I checked the source code and find that this will actually not, because the LayerHelper.bias_attr() method will return a ParamAttr object even if the required key doesn't exist:

attr = ParamAttr._to_attr(self.kwargs.get('bias_attr', None))

The problem is in ParamAttr._to_attr(), which will create a new ParamAttr object when the argument is None:

if arg is None:
    return ParamAttr()

Following this logic, when bias_attr of conv2d() layer is passed in None, fluid will add an bias.

I'd ask if this is intended behavior, and if so, please make better documentation on sequence_conv, conv2d, conv3d and other alike layers which use bias_attr as a parameter. Thanks.