not suport two input image network?

[wikiwen]

No description provided.

[sovrasov]

Try to use the input_constructor argument:

class Siamese(nn.Module):
    def __init__(self):
        super(Siamese, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, 3, 1)
        self.conv2 = nn.Conv2d(1, 10, 3, 1)

    def forward(self, x):
        # assume x is a list
        return self.conv1(x[0]) + self.conv2(x[1])

def prepare_input(resolution):
    x1 = torch.FloatTensor(1, *resolution)
    x2 = torch.FloatTensor(1, *resolution)
    return dict(x = [x1, x2])

if __name__ == '__main__':
    model = Siamese()
    flops, params = get_model_complexity_info(model, input_res=(1, 224, 224), 
                                              input_constructor=prepare_input,
                                              as_strings=True, print_per_layer_stat=False)
    print('      - Flops:  ' + flops)
    print('      - Params: ' + params)

[wikiwen]

Try to use the input_constructor argument:

class Siamese(nn.Module):
    def __init__(self):
        super(Siamese, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, 3, 1)
        self.conv2 = nn.Conv2d(1, 10, 3, 1)

    def forward(self, x):
        # assume x is a list
        return self.conv1(x[0]) + self.conv2(x[1])

def prepare_input(resolution):
    x1 = torch.FloatTensor(1, *resolution)
    x2 = torch.FloatTensor(1, *resolution)
    return dict(x = [x1, x2])

if __name__ == '__main__':
    model = Siamese()
    flops, params = get_model_complexity_info(model, input_res=(1, 224, 224), 
                                              input_constructor=prepare_input,
                                              as_strings=True, print_per_layer_stat=False)
    print('      - Flops:  ' + flops)
    print('      - Params: ' + params)

Thanks~~it works！！

[chyohoo]

what if two inputs have different sizes?

[sovrasov]

@chyohoo in that case you can ignore the resolution parameter and use custom shapes:

def prepare_input(resolution):
    x1 = torch.FloatTensor(1, 3, 224, 224)
    x2 = torch.FloatTensor(1, 3, 128, 128)
    return dict(x = [x1, x2])

[JamesLee789]

@chyohoo in that case you can ignore the resolution parameter and use custom shapes:

def prepare_input(resolution):
    x1 = torch.FloatTensor(1, 3, 224, 224)
    x2 = torch.FloatTensor(1, 3, 128, 128)
    return dict(x = [x1, x2])

Hi. I tried to implement the calculation following your advice,

def prepare_input(resolution):
x = torch.FloatTensor(1, 3, 224, 224)
depth = torch.FloatTensor(1, 1, 224, 224)
return dict(x=[x, depth])

...

flops, macs, params = get_model_complexity_info(model, input_res=((1, 3, 224, 224),(1, 1, 224, 224)),input_constructor=prepare_input,as_strings=True,print_per_layer_stat=True, verbose=True)

However,I got the error.

Warning: module Softmax is treated as a zero-op.
Warning: module TestNet is treated as a zero-op.
Traceback (most recent call last):
File "get_model_complexity_info.py", line 18, in
flops, macs, params = get_model_complexity_info(model, input_res=((1, 3, 384, 384),(1, 1, 384, 384)),input_constructor=prepare_input,
File "/home/user/software/anaconda3/lib/python3.8/site-packages/ptflops/flops_counter.py", line 34, in get_model_complexity_info
_ = flops_model(**input)
File "/home/user/software/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'depth'

How could I fix it? Many thanks.

[sovrasov]

Hi! Take a look at the forward method of my test model: it assumes x to be a list, whereas your model may require a different input format as well as different arguments besides x. General suggestion here is to preserve input args and their layout in prepare_input.
My guess is your forward method looks like this: def forward(self, x, depth). If so, you have to add depth argument also:

def prepare_input(resolution):
  x = torch.FloatTensor(1, 3, 224, 224)
  depth = torch.FloatTensor(1, 1, 224, 224)
  return dict(x=x, depth=depth)

[JamesLee789]

@chyohoo in that case you can ignore the resolution parameter and use custom shapes:

def prepare_input(resolution):
    x1 = torch.FloatTensor(1, 3, 224, 224)
    x2 = torch.FloatTensor(1, 3, 128, 128)
    return dict(x = [x1, x2])

Hi. I tried to implement the calculation following your advice,

def prepare_input(resolution):
x = torch.FloatTensor(1, 3, 224, 224)
depth = torch.FloatTensor(1, 1, 224, 224)
return dict(x=[x, depth])

...

flops, macs, params = get_model_complexity_info(model, input_res=((1, 3, 224, 224),(1, 1, 224, 224)),input_constructor=prepare_input,as_strings=True,print_per_layer_stat=True, verbose=True)

However,I got the error.

Warning: module Softmax is treated as a zero-op.
Warning: module TestNet is treated as a zero-op.
Traceback (most recent call last):
File "get_model_complexity_info.py", line 18, in
flops, macs, params = get_model_complexity_info(model, input_res=((1, 3, 384, 384),(1, 1, 384, 384)),input_constructor=prepare_input,
File "/home/user/software/anaconda3/lib/python3.8/site-packages/ptflops/flops_counter.py", line 34, in get_model_complexity_info
_ = flops_model(**input)
File "/home/user/software/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'depth'

How could I fix it? Many thanks.

Thanks for your prompt reply! It works now.

[hj611]

Does batchsize have to be 1? Can you customize the batch size? I tried your prepare_input and got the following error:
input = input_constructor(input_res)
TypeError: 'dict' object is not callable

[sovrasov]

Hi! If you have an ordinary model that consumes only one input tensor x, the following would work for you:

bs = 2
input_constructor = lambda _: {"x":  torch.FloatTensor(bs, 3, 224, 224)}
macs, params = get_model_complexity_info(net, (3, 224, 224),
                                             as_strings=True,
                                             input_constructor=input_constructor,
                                             print_per_layer_stat=True,
                                             ost=ost)