pyorch的坑
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [2, 3, 13, 13, 85]], which is output 0 of AsStridedBackward, is at version 6; expected version 3 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!
他跟我说good luck？？？？？？？？？？？？？？？？？？？？？？？？？？？？？？？？
反正我找了很多解释方法，比如inplace、+=都改了没B用，看了源码，就是一个简单的element 乘，我很无语。并且神奇的是，在A电脑可以跑，在B电脑不能跑，一个没GPU一个有GPU
tensorflow的坑
tensorflow底层代码不会错，但是写起来很麻烦，一个batchnormalization都要自己写，可读性又是极差。用keras的话，涉及到多gpu时候我感觉里写的很乱，比如并行多gpu
wtih tf.device(cpu0):
model
parallel_model=parallel_model(model)
有些电脑就会报错，但是你改成：
wtih tf.device(cpu0):
model.predict()
就不会报错，我贼蒙蔽。