different with torchsummary

[ltm920716]

hello， I use craft from https://github.com/clovaai/CRAFT-pytorch，

the torchsummary output is：

the torchinfo output is：

there is a big different， please help

[TylerYep]

torchinfo should be more correct than torchsummary; torchsummary is not currently maintained. Feel free to reopen if you have any additional concerns.

[ltm920716]

hi, @TylerYep this is a demo

################################# `import torch import torch.nn as nn import torch.nn.functional as F from torchinfo import summary as sm

device = torch.device('cpu') num_classes = 10

class LeNet(nn.Module): def init(self, in_channels, num_classes): super(LeNet, self).init() self.conv1 = nn.Conv2d(in_channels, 20, kernel_size=5, stride=1) # 20x24x24 self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2) # 20x12x12 self.conv2 = nn.Conv2d(20, 50, kernel_size=5, stride=1) # 50x8x8 self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2) # 50x4x4 self.fc1 = nn.Linear(50 4 4, 500) # 500 self.fc2 = nn.Linear(500, num_classes) # 10

def forward(self, input):
    out = self.conv1(input)
    out = self.pool1(out)
    out = self.conv2(out)
    out = self.pool2(out)
    out = out.reshape(out.size(0), -1) 
    out = F.relu(self.fc1(out))
    out = self.fc2(out)
    return out

model = LeNet(1, num_classes).to(device)

sm(model, input_size=(1, 1, 28, 28))` ################################

here is the output：

and here is the ground-thurh：

• Params: 520 + 25050 + 400500 + 5010 = 431080
• Params_mem：431080 * (32/8) / 1024 / 1024 = 1.64 MB
• Input_mem：1*28*28 * (32/8) / 1024 / 1024 = 0.003 MB - forwardmem：(20*24*24 + 20*12*12 + 50*8*8 + 50*4*4 + 500 + 10)* (32/8) / 1024 / 1024 = 0.072 MB

So what the Forward/backward pass size (MB): 0.12 means? If I want to know the forward inference memory， could torchinfo show？

Thanks！

[TylerYep]

Note that memory calculation should be slightly more for a tensor, because torch.Tensor stores other information too, not just the 20x24x24 shape. We use sys.getsizeof(tensor) to get the most accurate size in MB, which works for all types of layers.

0.12 and 0.072 are close enough to explain the difference, but if you see an error in the calculation, feel free to open a PR!

[ltm920716]

@TylerYep Great，thanks！

[Yuxinyi-Qiyu]

hey, sorry to bother you. I want to know how the "Params size" calculate. the params are 431080, so the Params_mem：431080 * (32/8) / 1024 / 1024 = 1.64 MB, why the "Params size" is 1.72 instead of 1.64?

[mert-kurttutan]

hey, sorry to bother you. I want to know how the "Params size" calculate. the params are 431080, so the Params_mem：431080 * (32/8) / 1024 / 1024 = 1.64 MB, why the "Params size" is 1.72 instead of 1.64?

Hi @Yuxinyi-Qiyu What you are calculating is MiB, not MB MB is calculated in terms of power of 10: 1MB= 10*6 B. So, the calculation done in source code: 431080 (32/8) / 1000 / 1000 = 1.72

[Yuxinyi-Qiyu]

hey, sorry to bother you. I want to know how the "Params size" calculate. the params are 431080, so the Params_mem：431080 * (32/8) / 1024 / 1024 = 1.64 MB, why the "Params size" is 1.72 instead of 1.64?

Hi @Yuxinyi-Qiyu What you are calculating is MiB, not MB MB is calculated in terms of power of 10: 1MB= 10*6 B. So, the calculation done in source code: 431080 (32/8) / 1000 / 1000 = 1.72

I got it! thanks!!! 🥰