Exploding gradients when using both backward() and DataParallel


I am building a model that needs the gradients computed during a backward pass, in the forward of my network. To speed up experimentation I am using nn.DataParallel to spread the workload across multiple GPUs. However, I’m running into a problem of exploding gradients (and losses) when I compute the backward() from the forward() function when I use multiple GPUs.

The following code is a copy of the CIFAR-10 tutorial, except that the forward() function now also does a backward to collect gradients (I would usually do this using backward hooks, but have left out that piece of code for clarity), I have also wrapped the network inside nn.DataParallel

import torch
import torchvision
import torchvision.transforms as transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

transform = transforms.Compose(
     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
                                          shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
                                         shuffle=False, num_workers=2)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)

        predictions = F.softmax(x).argmax(dim=1)
        ohe = torch.zeros_like(x)
        for i, label in enumerate(predictions):
            ohe[i, label] = 1

        ohe = torch.autograd.Variable(ohe)
        x.backward(gradient=ohe, retain_graph=True)

        return x

net = Net()
if torch.cuda.device_count() > 1:
    net = nn.DataParallel(net)

criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

for epoch in range(2):  # loop over the dataset multiple times

    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        # get the inputs; data is a list of [inputs, labels]
        inputs, labels = data
        inputs, labels = inputs.cuda(), labels.cuda()

        # zero the parameter gradients

        # forward + backward + optimize
        outputs = net(inputs)
        loss = criterion(outputs, labels)

        # print statistics
        running_loss += loss.item()
        if i % 2000 == 1999:    # print every 2000 mini-batches
            print('[%d, %5d] loss: %.3f' %
                  (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0

print('Finished Training')

When I run this code on a single GPU, it works fine (i.e. losses stay in normal range), when I run it on multiple GPUs, the losses explode.

What am I doing wrong?

Does anybody have an idea to fix this issue?