why do you usr cross entropy loss?

[EthanGuan]

I noticed in the paper and original repo, L2 loss function is used.
In your implementation:

# define loss function (criterion) and pptimizer
criterion = nn.CrossEntropyLoss().cuda()

cross entropy loss is used.

does it really work?

[tensorboy]

I didn't successfully implement the training part, but the model is converted right. ;-D.

Best,
Wangpeng

[tensorboy]

Try train.py