What is layer's interface standard in Python API ?

[qingqing01]

1. About the name:

Some interfaces have the name, some not.

Paddle/python/paddle/v2/fluid/layers/nn.py

Lines 1432 to 1439 in 279aa62

	def pool2d(input,
	pool_size,
	pool_type,
	pool_stride=None,
	pool_padding=None,
	global_pooling=False,
	use_cudnn=True,
	name=None):

Paddle/python/paddle/v2/fluid/layers/nn.py

Lines 1136 to 1145 in 279aa62

	def conv2d(input,
	num_filters,
	filter_size,
	stride=None,
	padding=None,
	groups=None,
	param_attr=None,
	bias_attr=None,
	use_cudnn=True,
	act=None):

2. About kwargs

Some interfaces have the **kwargs, some not.

Paddle/python/paddle/v2/fluid/layers/nn.py

Lines 832 to 836 in 279aa62

	def cos_sim(X, Y, **kwargs):
	"""
	This function performs the cosine similarity between two tensors
	X and Y and returns that as the output.
	"""

Paddle/python/paddle/v2/fluid/layers/nn.py

Lines 818 to 819 in 274f4e9

	def crf_decoding(input, param_attr, label=None):
	helper = LayerHelper('crf_decoding', **locals())

[JiayiFeng]

I think we shall avoid using **kwargs in user interfaces. It makes users confused about what arguments are required exactly.

Can anyone show me any examples the **kwargs is necessary?

[qingqing01]

I think we shall avoid using **kwargs in user interfaces. It makes users confused about what arguments are required exactly.

I agree with you.

[reyoung]

name

I was wishing the layer with the same name shared the same parameters. However, it seems that we have not implemented this feature.

We can give an easier way for two layers share the same parameters, and remove the name parameter.

**kwargs

I think we shall avoid using **kwargs in user interfaces. It makes users confused about what arguments are required exactly.
Can anyone show me any examples the **kwargs is necessary?

It because we pass the program to each layer function before. Now we change program to global variables. It is no need to use **kwargs.

[JiayiFeng]

Please notice that **kwargss in layer interfaces has been removed by #8405 .

[abhinavarora]

@reyoung I was thinking that we should keep the name everywhere. What I see in the layer_helper.py that we prefix the names of the parameters with the name that is provided. I think this is very useful for debugging deep learning algorithms. For example, your model might have lots of fc layers. By giving name to layer, it is very easy to only concentrate on the parameter that you want to debug. Similarly, this will also be very useful once PaddleBoard for visualization is ready. You can visualize params from any particular layer easily if you know the name. What do you think?