Reposting of an old question I never got a reply to Multi GPU Hook not correctly filling buffer. I know runnable minimal examples are the way to go so I now I have a full runnable minimal example of the code written. Please take a look and help me get to the bottom of this. This is what it looks like with two runs:
> CUDA_VISIBLE_DEVICES=1 python mainMinimalExampleMultiGPU.py
Conv Average Gradient:
[0.00044749726757895453, 0.0014000369330415242, -0.0008686411516918384]
fc Average Gradient:
[-0.004141018711068057, 0.0015833583892040112, 0.0011787552821185693, 0.0010372935249398085, -0.004048425233274684, 0.0006052607123126522, 0.0013055124756185216, 0.0007034393619838467, 0.0007521140892023609, 0.0010237101089629694]
> CUDA_VISIBLE_DEVICES=0,1 python mainMinimalExampleMultiGPU.py
Conv Average Gradient:
[0.0, 0.0, 0.0]
fc Average Gradient:
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
The following is the entire program which was started with the mnist example script.
from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.optim.lr_scheduler import StepLR
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
def saveAverageD(grad_out, Values):
with torch.no_grad():
if(len(grad_out.shape) == 2):
Values[0].average = Values[0].average * 0.99 + grad_out.sum((0)) * 0.01
else:
Values[0].average = Values[0].average * 0.99 + grad_out.sum((0,2,3)) * 0.01
class valueTracker(nn.Module):
def __init__(self, out_channels):
super(valueTracker, self).__init__()
self.register_buffer('average', torch.zeros(out_channels, device=device, dtype=torch.double))
class averageSaveConv(nn.Module):
def __init__(self, startLayer, out_channels):
super(averageSaveConv, self).__init__()
self.values = nn.ModuleList([])
self.values.append(valueTracker(out_channels))
self.layer = startLayer.double()
def forward(self,x):
out = self.layer(x)
out.register_hook(lambda grad: saveAverageD(grad, self.values))
return out
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv = averageSaveConv(nn.Conv2d(1, 3, 5, 1),3)
self.fc = averageSaveConv(nn.Linear(432, 10),10)
def forward(self, x):
x = self.conv(x)
x = F.relu(x)
x = F.max_pool2d(x, 2)
x = torch.flatten(x, 1)
x = self.fc(x)
output = F.log_softmax(x, dim=1)
return output
def train(args, model, device, train_loader, optimizer, epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device).double(), target.to(device).long()
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
print('Conv Average Gradient:')
print(model.module.conv.values[0].average.tolist())
print('fc Average Gradient:')
print(model.module.fc.values[0].average.tolist())
exit()#just here for debugging
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
if args.dry_run:
break
def test(model, device, test_loader):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device).double(), target.to(device).long()
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=2, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=14, metavar='N',
help='number of epochs to train (default: 14)')
parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
help='learning rate (default: 1.0)')
args = parser.parse_args()
torch.manual_seed(1)
kwargs = {'batch_size': args.batch_size}
if use_cuda:
kwargs.update({'num_workers': 1,
'pin_memory': True,
'shuffle': True},
)
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
dataset1 = datasets.MNIST('../data', train=True, download=True,
transform=transform)
dataset2 = datasets.MNIST('../data', train=False,
transform=transform)
train_loader = torch.utils.data.DataLoader(dataset1,**kwargs)
test_loader = torch.utils.data.DataLoader(dataset2, **kwargs)
model = Net()
model = torch.nn.DataParallel(model, device_ids=range(torch.cuda.device_count())).to(device)
optimizer = optim.SGD(model.parameters(), lr=args.lr)
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
test(model, device, test_loader)
if __name__ == '__main__':
main()
