import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from torch.nn.parameter import Parameter
self.embed = nn.Embedding(100000, 256, sparse=False)
def forward(self, idx):
toy = Toy().cuda()
optimizer = optim.SGD(toy.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss().cuda()
x = Variable(
torch.from_numpy(numpy.random.randint(0, 100000, 256)).cuda(),
t = Variable(
torch.LongTensor(numpy.random.randint(0, 2, 256)).cuda(),
start_time = time.time()
for _ in xrange(2000):
y = toy(x)
cost = criterion(y, t)
print time.time() - start_time
I tried this piece of code, the elapsed time is 18.3s (sparse=True) vs 3.84s (sparse=False) in a Tesla P100. I have Ubuntu 16 and CUDA8. Could someone please elaborate the some possible reasons?
Thank you in advance!
We recently fixed a performance issue with SGD on sparse gradients! If you checkout the master code, it should be vastly faster (and use much less memory!.
The fix is at https://github.com/pytorch/pytorch/pull/3139 if you are interested.
Thanks for your response.
By default, momentum is zero, change in this PR is not reached. May I know when the binary release is built? Is this fix or relate fix is included?
The source code doesn’t build on our server. I am guessing it’s due to out-of-date CUDA/driver/dependency. I don’t have root permission, which may take a lot time to verify who is the trouble maker. Could you please kindly update the binary release at your convenience?
You are correct about the momentum option.
Upon looking at the code, I realized that the problem is actually elsewhere. You should call
optimizer.zero_grad() at beginning of each loop Otherwise, the gradient will just accumulate each loop (so its actually optimizing the wrong gradient values), and without calling
coalesce(), sparse gradient can grow very large (each
x index will have multiple values), causing the slow down.
After adding the
zero_grad call, on current master, sparse version gives me
3.0683s and dense version gives me
About updating binary. we will do so once we release a new version. Currently, if you need the new features/fixes, you can build from source by yourself. It’s very very simple to do so.
Could you please elaborate at which point I should call coalesce() on which variable?
I hit the following when building from source. May I have some advice? I have access to server with Ubuntu16.4, CUDA8.0, 6GB memory and a Tesla P100, but no root permission.
No need to call coalesce. Just add zero_grad.
The current code accumulates gradients at each loop, which shouldn’t be the intended behavior.
Is the change integrated in the version 0.3.1?
I forgot the context of this question… But if it is about SGD being slow with sparse gradients, then it is fixed indeed.
I’ve always been having a hard time reading Pytorch source code, sad.
Could you please tell me the underlying difference between
Embedding being sparse or not?
My guess is that, with
sparse=True, the forward/backward will only collect the rows from the whole huge Embedding matrix and compute grad for these ones; whereas with
sparse=False, forward/backward will do the one-hot multiplication on the whole huge Embedding matrix, which involves much more computation than
With sparse=True, the backward gives sparse gradients instead of dense gradients. So it would be more efficient with large embedding matrix.