# Model update with "share\_memory" need lock protection **URL:** \ **Category:** reinforcement-learning\ **Created:** [March 11, 2020, 10:24am UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857 "2020-03-11T10:24:31Z")\ **Posts on this page:** 5\ **Page:** 1
**Author:** ![S\_Q](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/s_q/32/15892_2.png) [@S\_Q](https://discuss.pytorch.org/u/S_Q)\ **Post date:** [March 11, 2020, 10:24am UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857/1 "2020-03-11T10:24:32Z")
TL;DR. Maybe a lock is needed for `optimizer.step()` while using multiprocessing. * * * The example provided in [MULTIPROCESSING BEST PRACTICES] ([https://pytorch.org/docs/stable/notes/multiprocessing.html](https://pytorch.org/docs/stable/notes/multiprocessing.html)) is to update parameters without lock protection. ```auto import torch.multiprocessing as mp from model import MyModel def train(model): # Construct data_loader, optimizer, etc. for data, labels in data_loader: optimizer.zero_grad() loss_fn(model(data), labels).backward() optimizer.step() # This will update the shared parameters if __name__ == ' __main__': num_processes = 4 model = MyModel() # NOTE: this is required for the ``fork`` method to work model.share_memory() processes = [] for rank in range(num_processes): p = mp.Process(target=train, args=(model,)) p.start() processes.append(p) for p in processes: p.join() ``` And I checked some A3C code from github, and they do it in the similar way. To double check whether it’s safe to do that, I created a test case: - The global network is init with 0 - one worker will add it 1 for 1000 times - another worker will less it 1 for 1000 times - if it is safe, the final network’s weight shall be 0 - repeat 100 times to check whether it is safe Test results: - With lock protection, it’s safe - Without lock protection (comment the lock, and unindent the `optimizer.step()`), from time to time, the weight can not come back to 0, which means: not safe ```auto import torch import torch.nn as nn import torch.multiprocessing as mp import torch.optim as optim from torch.autograd import Variable import os class Net(nn.Module): def __init__ (self): super(Net, self). __init__ () self.fc1 = nn.Linear(1, 1) self.fc1.weight.data.fill_(0.0) self.fc1.bias.data.fill_(0.0) def forward(self): return 0 def worker1(net, grad, lock): optimizer = optim.SGD(net.parameters(), lr=1) for _ in range(1000): optimizer.zero_grad() for param in net.parameters(): param._grad = Variable(torch.ones_like(param)) with lock: optimizer.step() # print(f'in process {os.getpid()}, the grad of weight is: {net.fc1.weight.grad.item()}, weight is {net.fc1.weight.item()}') def worker2(net, grad, lock): optimizer = optim.SGD(net.parameters(), lr=1) for _ in range(1000): optimizer.zero_grad() for param in net.parameters(): param._grad = -Variable(torch.ones_like(param)) with lock: optimizer.step() # print(f'in process {os.getpid()}, the grad of weight is: {net.fc1.weight.grad.item()}, weight is {net.fc1.weight.item()}') if __name__ == ' __main__': n = Net() n.share_memory() lock = mp.Lock() for idx in range(100): # print(idx) # print(f'in main {os.getpid()}, weight is {n.fc1.weight.item()}') p1 = mp.Process(target = worker1, args=(n, 1.0, lock)) p2 = mp.Process(target = worker2, args=(n, 1.0, lock)) p1.start() p2.start() p1.join() p2.join() # print(f'in main {os.getpid()}, weight is {n.fc1.weight.item()}') if n.fc1.weight.item() != 0: print(f'{idx} in main {os.getpid()}, weight is {n.fc1.weight.item()}') break print('finished') ``` My pytorch version is 1.4.0+cpu However, as I find it works without lock in the official “Hogwild” example and other examples, maybe the safe issue is not that important… ---
**Author:** ![iffiX](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/iffix/32/24443_2.png) [@iffiX](https://discuss.pytorch.org/u/iffiX)\ **Post date:** [June 4, 2020, 5:24pm UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857/2 "2020-06-04T17:24:46Z")
thread safety is not important, the hogwild essay basically spent pages trying to explain that the update is sparse and you can update in **whatever** way(sequence). It’s magic, so forget about the safety. 🤣 ---
**Author:** ![Brando\_Miranda](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/brando_miranda/32/14355_2.png) [@Brando\_Miranda](https://discuss.pytorch.org/u/Brando_Miranda)\ **Post date:** [February 15, 2021, 12:52am UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857/3 "2021-02-15T00:52:54Z")
> [@iffiX](#): > > you can update in **whatever** way(sequence) Can you clarify what you mean by “it’s magic”? ---
**Author:** ![S\_Q](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/s_q/32/15892_2.png) [@S\_Q](https://discuss.pytorch.org/u/S_Q)\ **Post date:** [February 17, 2021, 10:20am UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857/4 "2021-02-17T10:20:08Z")
Hi, does “whatever way” include update with data race? 👻 ---
**Author:** ![OrNot](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ornot/32/12839_2.png) [@OrNot](https://discuss.pytorch.org/u/OrNot)\ **Post date:** [May 12, 2021, 12:20am UTC](https://discuss.pytorch.org/t/model-update-with-share-memory-need-lock-protection/72857/5 "2021-05-12T00:20:26Z")
Any clue on this? I have the same question.