vision-mlp-oneflow 0.0.2

vision mlp model based on oneflow

vision-mlp-oneflow

Vision MLP Models Based on OneFlow

TODO LIST

• Mlp-Mixer
• ResMLP
• gMLP
• S2-MLP
• Vision-Permutator

复现过程中遇到的问题

• tensor.chunk()方法没有和torch.tensor.chunk()对齐, 在输入的dim参数为负数的时候未作判断, 并且整体功能还变成了repeat操作
• 缺少tensor.floor_()方法, 只能用tensor = tensor.floor()替代
• 缺少tensor.erfinv()方法及其inplace版本, 无法复现trunc_norm_
• 缺少tensor.new_empty()方法
• flow.shape返回的是oneflow._oneflow_internal.Size 无法和 tuple相加, 按以下方法使用会报错:

# oneflow里报错
x = flow.tensor(np.random.randn(1, 16, 512), dtype=flow.float32)
x.shape + (4, )
x.size() + (4, )
>>>TypeError: unsupported operand type(s) for +: 'oneflow._oneflow_internal.Size' and 'tuple'

# torch里不报错
x = torch.randn(1, 16, 512)
x.shape + (4, )
x.size() + (4, )
>>> torch.Size([1, 16, 512, 4])

• tensor.gather()未具体判定输入的dim为负数的情况