flops are counted multiple times if a module is shared by other modules

[CocytusDuo]

If a module is passed to a sub-module, for example:

import torch.nn as nn
import ptflops

class Block(nn.Module):
    def __init__(self, linear_layer) -> None:
        super().__init__()
        self.linear_layer = linear_layer
    
    def forward(self, x):
        return self.linear_layer(x)
    
class Test_model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.linear_layer = nn.Linear(1000, 1000)
        self.block = Block(self.linear_layer)

    def forward(self, x):
        out = self.linear_layer(x)
        out = self.block(out)
        return out

net = Test_model()
print(ptflops.get_model_complexity_info(net, (20, 1000)))

then, the flops of module nn.Linear(1000, 1000) will be counted twice in Test_model and Block:

Warning: variables __flops__ or __params__ are already defined for the moduleLinear ptflops can affect your code!
Test_model(
  2.0 M, 200.000% Params, 80.0 MMac, 100.000% MACs, 
  (linear_layer): Linear(1.0 M, 100.000% Params, 40.0 MMac, 50.000% MACs, in_features=1000, out_features=1000, bias=True)
  (block): Block(
    1.0 M, 100.000% Params, 40.0 MMac, 50.000% MACs, 
    (linear_layer): Linear(1.0 M, 100.000% Params, 40.0 MMac, 50.000% MACs, in_features=1000, out_features=1000, bias=True)
  )
)
('80.0 MMac', '1.0 M')

[sovrasov]

Hi! Yeah, that's a drawback of the tracing process. Since it doesn't affect the output value, I'd consider it as a minor one. I'll have a look if a simple fix is possible.

[sovrasov]

I added more meaningful warning for that case in #109
Manipulations with module totals will also lead to contradictive results.

[erwangccc]

Since it doesn't affect the output value, I'd consider it as a minor one

Hi, @sovrasov
Why you said it doesn't affect the output value, if the linear_layer just use one time, we calculate the op's MACs twice. I I think it's wrong. Please correct me if there is misunderstanding.

Thanks!

[sovrasov]

By the output value I assume the return value of get_model_complexity_info, and it is not affected. For calculating extended statistics during model printing, I use a different mechanism, which can not handle shared submodules, and therefore counts parameters of shared submodules several times.
At the same time, inference of those shared submodules is conducted several times, so they contribute to the final MACs several times as well.