ModuleAttributeError: 'DataParallel' object has no attribute 'custom_function'

Hi everyone,

I have a class:

import torch
import torch.nn as nn
import torch.nn.functional as F

def create_model():
    return Net()

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, 1)
        self.conv2 = nn.Conv2d(32, 64, 3, 1)
        self.dropout1 = nn.Dropout2d(0.25)
        self.fc1 = nn.Linear(9216, 128)
        self.dropout2 = nn.Dropout2d(0.25)
        self.fc2 = nn.Linear(128, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.relu(self.conv2(x))
        x = F.max_pool2d(x, 2)
        x = torch.flatten(self.dropout1(x), 1)
        x = F.relu(self.fc1(x))
        x = self.dropout2(x)
        x = self.fc2(x)
        output = F.log_softmax(x, dim=1)

        return output

    def log_weights(self, step, writer):
        writer.add_histogram('weights/conv1/weight',, step)
        writer.add_histogram('weights/conv1/bias',, step)
        writer.add_histogram('weights/conv2/weight',, step)
        writer.add_histogram('weights/conv2/bias',, step)
        writer.add_histogram('weights/fc1/weight',, step)
        writer.add_histogram('weights/fc1/bias',, step)
        writer.add_histogram('weights/fc2/weight',, step)
        writer.add_histogram('weights/fc2/bias',, step)

With the training function:

def train(use_cuda, model, epoch, optimizer, log_interval, train_loader, writer):
    for batch_idx, (data, target) in enumerate(train_loader):
        if use_cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data), Variable(target)
        output = model(data)
        loss = F.nll_loss(output, target)
        if batch_idx % log_interval == 0:
            print(f'Train Epoch: {epoch} [{batch_idx * len(data)}/{len(train_loader.dataset)}'
                  f'({100. * batch_idx / len(train_loader):.0f}%)]\tLoss: {loss.item():.6f}')
            step = epoch * len(train_loader) + batch_idx
            log_scalar('train_loss',, step, writer)
            model.log_weights(step, writer)

And define the model like this:

    # Define model, device and optimizer
    model = create_model()
    if torch.cuda.device_count() > 1:
        model = nn.DataParallel(model)
    optimizer = optim.Adam(model.parameters())

However, I keep running into:
ModuleAttributeError: 'DataParallel' object has no attribute 'log_weights'


This only happens when MULTIPLE GPUs are used.
It does NOT happen for the CPU or a single GPU.

I expect the attribute to be available, especially since the wrapper in Pytorch ensures that all attributes of the wrapped model are accessible.


  • PyTorch Version (e.g., 1.0): 1.6
  • OS (e.g., Linux): Ubuntu 18
  • How you installed PyTorch (conda, pip, source): conda
  • Python version: 3.7
  • CUDA/cuDNN version: 10.1
  • GPU models and configuration: 2 x V100
  • Any other relevant information:

Again, only happens with multiple GPUs.

I am happy to share the full code. However, it is a mlflow project and you need docker with the nvidia-container thingy to run it.

Just tell me if desired.

Can you replace

            model.log_weights(step, writer)


            model.module.log_weights(step, writer)

This works. However, I expected this not to happen anymore, due to

Apparently this is not working?

That is a feature request I think ?

Since they mentioned pull requests several times I thought that they merged a PR for that recently.


I consider this solved then.