Double Counting Parameters

[TortoiseHam]

When a model takes multiple inputs (ex. a Siamese network) that all go through the same set of layers before the results are compared later on in the network, the model summary double-counts all of the shared parameters. For example, in the network below, if layerA has 10 parameters and layerB has 5, then it will report 25 parameters instead of 15.

x1 -> layerA --v ........................+-> layerB -> y x2 -> layerA --^

[TortoiseHam]

It also duplicates the layers in the summary, in addition to the parameter count. Here's the code for an example problem model:

https://github.com/fastestimator/fastestimator/blob/master/apphub/one_shot_learning/siamese_network/siamese_torch.py

[amarczew]

Hi @TortoiseHam!

Thank you for the reported issue! I understood the scenario but could you send a gists/code with an example that I can run quickly to see the problem?

[TortoiseHam]

import torch
import torch.nn as nn
import torch.nn.functional as fn
import pytorch_model_summary as pms

class SiameseNetwork(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 64, 10)
        self.fc1 = nn.Linear(6400, 1)

    def branch_forward(self, x):
        x = self.conv1(x)
        x = fn.max_pool2d(x, 2)
        x = fn.relu(x)
        x = x.view(x.shape[0], -1)
        return x

    def forward(self, x):
        x1 = self.branch_forward(x[0])
        x2 = self.branch_forward(x[1])
        x = torch.abs(x1 - x2)
        x = self.fc1(x)
        x = torch.sigmoid(x)
        return x

if __name__ == "__main__":
    net = SiameseNetwork()
    img_a = torch.ones(size=(5,3,30,30))
    img_b = torch.zeros(size=(5,3,30,30))
    pms.summary(net, [img_a, img_b], print_summary=True)

[TortoiseHam]

which prints the following:

-----------------------------------------------------------------------
      Layer (type)        Output Shape         Param #     Tr. Param #
=======================================================================
          Conv2d-1     [5, 64, 21, 21]          19,264          19,264
          Conv2d-2     [5, 64, 21, 21]          19,264          19,264
          Linear-3              [5, 1]           6,401           6,401
=======================================================================
Total params: 44,929
Trainable params: 44,929
Non-trainable params: 0
-----------------------------------------------------------------------

[TortoiseHam]

Thanks for the reply by the way. I suspect that this could be fixed by memorizing the ids of the layers that have already been seen while crawling the model, but I haven't had time to study the internals of the repo yet.

[amarczew]

@TortoiseHam you are are right. Internally, I have a dictionary with layers to be printed. I just need to add a flag_print. I will fix this in few days. Thank you so much!

[jonashaag]

I also have this issue.

[amarczew]

hi @jonashaag , unfortunately in the last weeks I couldn't implement a fix for this problem. I will try to fix it next weekend.

Thank you for your report

[amarczew]

@jonashaag and @TortoiseHam issue resolved.

@TortoiseHam for your example, current output is:

-----------------------------------------------------------------------
      Layer (type)        Output Shape         Param #     Tr. Param #
=======================================================================
          Conv2d-1     [5, 64, 21, 21]          19,264          19,264
          Linear-2              [5, 1]           6,401           6,401
=======================================================================
Total params: 25,665
Trainable params: 25,665
Non-trainable params: 0
-----------------------------------------------------------------------

[TortoiseHam]

Perfect! Thanks for resolving this