NotImplementedError when attempting to run network forwards

ProGamerGov · March 14, 2018, 11:42pm

When I try to run my net forward with: net.forward(content_image), I get the following error:

Traceback (most recent call last):
  File "st.py", line 150, in <module>
    net.forward(content_image)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/container.py", line 67, in forward
    input = module(input)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 71, in forward
    raise NotImplementedError
NotImplementedError

Running net(content_image) results in:

  File "st.py", line 150, in <module>
    net(content_image)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/container.py", line 67, in forward
    input = module(input)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 71, in forward
    raise NotImplementedError
NotImplementedError

And running net.updateOutput(content_image) results in:

Traceback (most recent call last):
  File "st.py", line 150, in <module>
    net.updateOutput(content_image)
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 398, in __getattr__
    type(self).__name__, name))
AttributeError: 'Sequential' object has no attribute 'updateOutput'

My loss modules are setup like this:

class ContentLoss(nn.Module):
    def __init__(self, strength, normalize):
        super(ContentLoss, self).__init__()

    def updateOutput(self, input):
        return self.output
  
    def updateGradInput(self, input, gradOutput):
        return self.gradInput

When setting up my sequential network, I first add a layer from a pre-trained model from torchvision:

net = nn.Sequential()
net.add_module(layer_name, layer)

Then I add my loss modules:

loss_module = ContentLoss(content_weight, norm)
net.add_module(layer_name, loss_module)

My full code can be found here:

gist.github.com

https://gist.github.com/ProGamerGov/2066ec2c09117931f69d226b9115d21f

LossModules.py

import torch
import torch.nn as nn
import torchvision

# Define an nn Module to compute content loss in-place
class ContentLoss(nn.Module):
    def __init__(self, strength, normalize):
      super(ContentLoss, self).__init__()
      self.strength = strength
      self.target = torch.Tensor()

This file has been truncated. show original

st.py

import torch
import torch.nn as nn
import torchvision
import torchvision.models as models
import torchvision.transforms as transforms
from torch.autograd import Variable
import torch.optim as optim

from PIL import Image
from LossModules import ContentLoss

This file has been truncated. show original

So what am I doing wrong here? And how do I fix the issue? Is it an issue with how I’ve setup my custom loss modules? Or how I am trying to run the network forward?

Edit

I found the solution here: NotimplementedError while densenet implementation

Unlike torch.legacy.nn, nn.Module requires the sub module name forward instead of updateOutput.

Second Edit:

Though my self.output variable is not being passed correctly to my GramMatrix loss function, when it should have been created in my ContentLoss function:

AttributeError: 'GramMatrix' object has no attribute 'output'

albanD · March 15, 2018, 1:16pm

Hi,

Be careful that the legacy module follows exactly the torch7 semantic and most of it’s api.
The regular pytorch module do not, nn.Module should be called with their input, and the function that should be implemented is forward, not updateOutput. Moreover, no updateGradInput nor updateParameters is needed because this is done by the autograd engine automatically.