bug found

[tszdanger]

in "手写数字识别器_minst_convnet.ipynb"
in def forward()
x = F.log_softmax(x, dim = 0) #输出层为log_softmax，即概率对数值log(p(x))。采用log_softmax可以使得后面的交叉熵计算更快
here the dim should be 1 because we see dim=0 as per data line

[tszdanger]

And I also feel that in the same project
We use CrossEntropyLoss but the last layer of our network is log_softmax().
This may be not right