ZZK

不连续的情况下，用pack需要保证连续。但是这里判断是否连续比较麻烦，后续用unroll展开处理

> add a oneflow version badge in `README.md`? like PyTorch Sure，it look like this: 

pytorch现在都是推荐用 DDP了，而不是DP

Try to use this: ```python input1 = oneflow.rand(9, 40, 20, 10,dtype=oneflow.float64) input2 = oneflow.randint(0,5,(1,),dtype=oneflow.int64) mod = oneflow.nn.MovingAverageMinMaxObserver(training= True, stop_update_after_iters=1, quantization_formula= 'google', quantization_bit=8, quantization_scheme='symmetric', momentum= 0.95).double() # Add .double() output =...

  I just follow your script, and get different result of `count`. It seems like a Floating point error by accident, I think it is not a Bug? Or...

Currently OneFlow only support ROIAlign Operator in CUDA, you can set the tensor `device` property as "cuda"

You can change to use this: ```python import oneflow input = oneflow.rand(4,10,32,32,dtype=oneflow.float64) m = oneflow.nn.Conv2d(10,10,3).double() output = m(input) ``` `.double()` will change the parameter to float64 dtype(Default is float32 dtype....

已经由 https://github.com/Oneflow-Inc/oneflow/pull/8936 该PR完成

我倾向不合并到master分支，以这个分支给用户编译使用

> 输入会恰好等于 0 或者 6 吗 是的，如果恰好等于这两个边界值，梯度的表现是不一样的