Strange classification accuracy drop

I’m training an incremental classification model (ResNet-32) on the CIFAR-100 dataset.

Using cross-entropy loss, I can get the final 43.2% accuracy. I fixed random seed for each random module (eg Python built-in random module, np.random, torch, etc.), so I can get the same accuracy each time if I trained the model many times.

However, if I put some other torch.tensor on GPU, it causes the accuracy drop (42.6%). And indeed this torch.tensor is not used in training.

(torch.tensor example, self.class_sizes = torch.tensor([20, 20, 500, 500,.....]).view(-1, 1).to(self._device), and self.class_sizes doesn’t be used in training process. I just put this variable there.)

Does anyone have any ideas or encountered similar situations? Thanks!

conda environment


0809 Update:
Another similar situation
If I have a Python class (BigModel below), which includes the training/test procedure.

class AuxModel(nn.Module):
    def __init__(self):
        super(AuxModel, self).__init__()

        self.align_FC = nn.Linear(64, 1)

        for m in self.modules():
            if isinstance(m, nn.Linear):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                if m.bias is not None:

    def forward(self, x):
        return self.align_FC(x)

class BigModel():
    def __init__(self, args):
           self._network = ResNet32(...)

           self.SP = AuxModel()

    def train(self, ...):

    def test(self, ...):

If I train the model (ResNet32), and totally not used self.SP, I will get lower accuracy (74.2%).

But if I directly comment out the #self.SP, I can get higher accuracy (76.6%). Another thing I can know is that I do not use the self.SP in my code, because I directly commented out that, I did not get any error message.

Since AuxModel initialized modules randomly, the pseudorandom number generator will be called and will thus change the random values of all following operations. The accuracy decrease seems to be a symptom of the changed seed and you should see the same effect, if you remove self.SP and rerun the code with different seeds.

1 Like

Many thanks for your reply!

I will try to use different seeds and self.SP architectures to do some experiments, and report the results later.

Not mentioned above, different self.SP architectures cause different final accuracy drops! I think this observation also results from the symptom that you had mentioned, which mainly related to pseudorandom number generator)