The network will not converge when backend is 'cupy'

[fangwei123456]

spikingjelly/spikingjelly/clock_driven/examples/speechcommands.py

Line 303 in 8fded1a

LIFWrapper(neuron.MultiStepLIFNode(tau=10.0 / 7, surrogate_function=surrogate.Sigmoid(alpha=10.), backend='torch')),

[fangwei123456]

        for audios, labels in tqdm(train_dataloader, disable=True):
            audios = audios.cuda(non_blocking=True)
            labels = labels.cuda(non_blocking=True)
            optimizer.zero_grad()

            for m in net.modules():
                if isinstance(m, neuron.MultiStepLIFNode):
                    m.backend = 'cupy'


            out_spikes_counter_frequency = net(audios)

            loss = criterion(out_spikes_counter_frequency, labels)

            loss.backward()

            gd_cupy = []
            for m in net.parameters():
                gd_cupy.append(m.grad.clone())

            reset_net(net)
            optimizer.zero_grad()

            for m in net.modules():
                if isinstance(m, neuron.MultiStepLIFNode):
                    m.backend = 'torch'

            out_spikes_counter_frequency = net(audios)

            loss = criterion(out_spikes_counter_frequency, labels)


            loss.backward()

            gd_torch = []
            for m in net.parameters():
                gd_torch.append(m.grad.clone())

            reset_net(net)

            for i in range(gd_cupy.__len__()):
                print(i, gd_cupy[i].shape, (gd_cupy[i] - gd_torch[i]).abs().max().item())


            # Rate-based output decoding
            correct_rate = (out_spikes_counter_frequency.argmax(
                dim=1) == labels).float().mean().item()
            print(net.train_times, 'acc=', correct_rate)

            net.train_times += 1

Epoch 0
0 torch.Size([64, 1, 4, 3]) 2.525382125639908e-09
1 torch.Size([64, 64, 4, 3]) 1.087092904583642e-08
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
0 acc= 0.09375
0 torch.Size([64, 1, 4, 3]) 3.250982372193789e-09
1 torch.Size([64, 64, 4, 3]) 7.123997392000092e-09
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
1 acc= 0.078125
0 torch.Size([64, 1, 4, 3]) 2.8865985157722207e-09
1 torch.Size([64, 64, 4, 3]) 1.2915431391036236e-08
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
2 acc= 0.0625
0 torch.Size([64, 1, 4, 3]) 3.8988936523765005e-09
1 torch.Size([64, 64, 4, 3]) 1.1206062744406609e-08
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
3 acc= 0.046875
0 torch.Size([64, 1, 4, 3]) 2.381410402207962e-09
1 torch.Size([64, 64, 4, 3]) 8.244219529274233e-09
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
4 acc= 0.0625
0 torch.Size([64, 1, 4, 3]) 2.9487301489439233e-09
1 torch.Size([64, 64, 4, 3]) 1.0766485480928623e-08
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
5 acc= 0.078125
0 torch.Size([64, 1, 4, 3]) 2.3316282238283748e-09
1 torch.Size([64, 64, 4, 3]) 9.149445645562082e-09
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
6 acc= 0.078125
0 torch.Size([64, 1, 4, 3]) 3.791615021953021e-09
1 torch.Size([64, 64, 4, 3]) 1.2025250128999687e-08
2 torch.Size([64, 64, 4, 3]) 0.0
3 torch.Size([12, 2560]) 0.0
4 torch.Size([12]) 0.0
7 acc= 0.078125

[fangwei123456]

import torch

device = 'cuda:0'
x = torch.rand([2, 3, 4], device=device).transpose(0, 2)
y = torch.zeros_like(x)
args_list = [x, y]
ret_list = []
for item in args_list:
    if isinstance(item, torch.Tensor):
        item = item.contiguous()
        ret_list.append(item.data_ptr())
print(ret_list)

[47284487168, 47284487168]

[fangwei123456]

https://discuss.pytorch.org/t/bug-two-tensors-have-the-same-data-ptr/132104