Created by: pkuyym

Survey and conclusion

Here, I surveyed several popular dl frameworks including tensorflow, caffe2 and pytorch to check their ops' documentation. I select fully connected operator as a typical example.

Component

• Feature summarization: Summary function of this op (including equation and detailed description)
• Usage example: Tell how to use this op (refer PyTorch)
• Python api definition: Show definition of the python api
• Python wrapper location: Link to wrapper code
• Parameters description: Describe each parameters in python api
• CPP code location: Link to cpp source code
• Highlight note: Something should pay attention to
• Other

At least

• Feature summarization
• Python api definition
• Python wrapper location
• Parameter description

More

• Usage example
• Highlight note
• CPP code location
• Other

Documentation for PaddlePaddle Ops

• Python api definition [Python code snippet]
• Python wrapper location [URL]
• Feature summarization [Function summarization] [Equation and description] [Tips]
• Parameters description [Description] [Data type] [Shape]
• Usage example [Python code snippet]

An example

FC [python/paddle/v2/fluid/layers.py#fc]

fc(input,
   size,
   num_flatten_dims=1,
   param_attr=None,
   bias_attr=None,
   act=None,
   name=None,
   main_program=None,
   startup_program=None)

Applies linear transformation to the input data. The equation is:

$$Y = Act(W^T * X + b)$$

In the above equation:

• X: input value, a tensor with rank at least 2.
• W: weight, a 2D tensor with shape [M, N].
• b: bias, a float scalar.
• Act: function to apply non-linearity activation.

All the input variables of this function are passed in as local variables to the LayerHelper constructor.

Args

input (Variable): The input vaule, a tensor with rank at least 2.
size (int): The output size, an interge value.
num_flatten_dims (int): Column number of the input.
param_attr: The parameters/weights.
param_initializer: Initializer used to initialize transoformation weights. If None, XavierInitializer is used.
bias_attr: The bias parameter.
bias_initializer: Initializer used to initialize bias. If None, ConstantInitializer is used.
act (str): Activation type.
name (str): Name/alias of the layer.
main_program (Program): The main program calling this.
startup_program (Program): The startup program.

Returns

Variable: the tensor variable storing the transformation and non-linearity activation result

Exceptions

ValueError: if rank of input less than 2

Usage Examples

```python
data = fluid.layers.data(name='data', shape=[1], dtype='float32')
fc = fluid.layers.fc(input=data, size=10, act="tanh")
```