amirhfarzaneh / lsoftmax-pytorch Goto Github PK

hi, very thanks for your contributions.
when i read calculate_cos_m_theta function in LSoftmaxLinear implementions,
i notice that torch.cos(2*torch.acos(cos_theta)) can simply compute cos_m_theta , why not use it ?
maybe i have some misunderstands

Can lsoftmax be transformed into ONNX?

Hi,
I am unable to understand why target would become None in line 55 of lsoftmax.py.

When I run your model, there will be this error, is there any solution? Thank you very much!

one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [256, 10]], which is output 0 of MmBackward, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

Hi, @amirhfarzaneh !

In __init__, you forgot to call method reset_parameters, because of this weight of the linear layer -- it's just zeros and the loss goes to nan

I tried to train with cifar10 and set margin to 2, but got nan.

learning rate is set to 0.1.

Any idea?

thanks for your work，but maybe there is some problems in models.py.
In paper，conv_2 is 3 conv2d

while in your implementation there are 4 conv2d
self.conv_2 = nn.Sequential( nn.Conv2d(64, 64, 3, padding=1), nn.PReLU(), nn.BatchNorm2d(64), nn.Conv2d(64, 64, 3, padding=1), nn.PReLU(), nn.BatchNorm2d(64), nn.Conv2d(64, 64, 3, padding=1), nn.PReLU(), nn.BatchNorm2d(64), nn.Conv2d(64, 64, 3, padding=1), nn.PReLU(), nn.BatchNorm2d(64), nn.MaxPool2d(2, 2) )