In pytorch function__ call__ And forward functions

Beginners nn.Module, I can't understand all kinds of calls , Later I understood , I guess I'll forget , So write a note

init & call

Code ：

class A():
    def __init__(self):
        print('init function ')        
        
    def __call__(self, param):
        print('call  function ', param)
a = A()

Output

analysis ：A Instantiate the class , Generating objects a, This process Automatically call _init_(), There is no call _call_()

---

Add a line to the above code

class A():
    def __init__(self):
        print('init function ') 
        
    def __call__(self, param):
        print('call  function ', param)
a = A()
a(1)

Output

analysis ：a It's the object ,python Let the object have parentheses like a function （ Parameters ） The function of , When using this function , Automatically call _call_（）

---

_ call_() Other functions can be called in , Such as forward function

class A():
    def __init__(self):
        print('init function ')
        
    def __call__(self, param):
        print('call  function ', param)
        res = self.forward(param)
        return res + 2
        
    def forward(self, input_): 
        print('forward  function ', input_)
        return input_
    
a = A()
b = a(1)
print(' result b =',b)

analysis ：_call _() Successfully called forward（）, And the return value is given to b

---

Besides, I had a misunderstand , Think that the value of this class can only be used when the parameter is declared , This is wrong

class A():
    def __init__(self):
        print('init function ')
        self.a = 100  #  Declare parameters a
        
    def __call__(self, param):
        print('call  function ', param)
        res = self.forward(param)
        return res + 2
        
    def forward(self, input_): 
        print('forward  function ', input_, self.a)
        return input_
    
a = A()
b = a(1)
print(' result b =',b)
print(a.a)

---

nn.Module

Look at the example above , You know the _call _() The role of , Let's see more below CNN Example

from torch import nn
import torch

class Ding(nn.Module):
    def __init__(self):
        print('init')
        super().__init__()
    
    def forward(self, input):
        output = input + 1
        print("forward")
        return output

dzy = Ding()
x = torch.tensor(1.0)
out = dzy(x)
print(out)

result ：

analysis ：
There is no call here _call_() and forward(), But it still shows forward, as a result of ：Ding This subclass inherits the parent class nn.Module Inside call function , Next, go to the source code

Find out _call_ Called _call_impl This function , It's equivalent to a nickname , Then go to this function to see

There are many parameters here , See reference for details 2. Found here forward_call Or _slow_forward, Or self.forward(), And this _slow_forward（） Can also use self.forward()

therefore ： _call _（） It was used forward, And the parent class forward Override... In subclasses （ Simple code ）

---

Of course , You can also rewrite __call__(), For example, we don't let it use forward（）

from torch import nn
import torch

class Ding(nn.Module):
    def __init__(self):
        print('init')
        super().__init__()
        
    def __call__(self, input_):
        print(' rewrite call,  no need forward')
        return 'hhh'
        
    def forward(self, input):
        output = input + 1
        print("forward")
        return output

dzy = Ding()
x = torch.tensor(1.0)
out = dzy(x)
print(out)

summary

Use object dzy(x) when , Parent class used nn.Module Of call function , Called forward, And this forward It was rewritten by us in subclasses .

Reference resources

https://blog.csdn.net/dss_dssssd/article/details/83750838
https://zhuanlan.zhihu.com/p/366461413