Different random initializations with a manual seed

I am very new to Pytorch and this is my first post here, so I apologise if the question is very straightforward.
My problem is that I have defined class net1 and initialised it randomly with a manual seed.

if torch.cuda.is_available():

class net1(nn.Module):
    def __init__(self):
        super(net1, self).__init__()
        self.main_body = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=96, kernel_size=11, stride=4, padding=0),
            # rest of the network...#
    def forward(self, x):
        output = self.main_body(x)
        return output

# custom weights initialization called on net1
def weights_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        m.weight.data.normal_(0.0, 0.02)
    elif classname.find('BatchNorm') != -1:
        m.weight.data.normal_(1.0, 0.02)

net1_ = net1()

However, when I add another class net2 to the code:

class net2(nn.Module):
    def __init__(self):
        super(net2, self).__init__()
        self.main_body = nn.Sequential(
            nn.Conv2d(4096, 1, 1)
    def forward(self, x):
        output = self.main_body(x)
        return output.view(-1, 1).squeeze(1)

net2_ = net2()

and instantiate it, even though I do not use it anywhere else and it is not connected to my main graph (which is built on net1_), I get different outputs from my graph.
Is it a reasonable outcome?

1 Like

I dont think this is related to net2. Some operations in the convolution layer are non-deterministic, and BatchNorm behaves differently based on different mini-batches.

If you inserted the code of net2 above net1, then what you see is also expected, because the weights of net2 get initialized, changing the state of the random number generator

1 Like

Thanks for your answer.
Though I get very consistent outputs every time I run the code, when class _net2 is not defined.
I guess I see, that defining a new module anywhere probably changes the behaviour of random generator.

I have some problem in understanding the meaning of torch.manual_seed(opt.manualSeed). You didn’t call this variable later on, how does it help initialize the net? Which parameters does it initialize? Thanks.

To my understanding, setting a manual seed makes all the initializations in the code (e.g. weight initializations embedded in the network definitions, or wherever else in the code that you call a random generator) to be deterministic. In other words, given a fixed manual seed, they are supposed to generate identical values, any time you run the program.

Yes, you don’t see it is called later. But in pytorch, the in-build random generator use the seed. Trace the related codes, you can find https://github.com/pytorch/pytorch/blob/1c51c185a1152ae7efe7d8d490102a9cdb7782e7/torch/lib/THC/THCTensorRandom.cpp. It is where you seed finally goes to.

Hi, I have a very simple question to ask.

If I manually set the seed and initialize three weight vectors of the same shape, then I am wondering whether all the initialized parameters for these three weight vectors will be the same?