bug of NDarray.reshape

[marongxing]

`
In [26]: import mxnet as mx

In [27]: mx.version
Out[27]: '1.6.0'

In [28]: nd = mx.nd

In [29]: a = nd.arange(27).reshape((3,3,3))

In [30]: a.reshape((3,3))
Out[30]:

[[0. 1. 2.]
[3. 4. 5.]
[6. 7. 8.]]
<NDArray 3x3 @cpu(0)>
`
the the old shape (3,3,3) and new shape (3,3) is not in the same size.

is there some special purpose for this?

[leezu]

I verified that this doesn't affect the np interface: mx.np.arange(27).reshape((3,3)) works as expected on master

[szha]

@cs8105 congrats that you have found mxnet's built-in way of creating memory leaks🤦‍♂️

Joke aside, I think what we need here is a check for consistent size. the logic for ndarray can be found here:
https://github.com/apache/incubator-mxnet/blob/cc287a0fc2de030c5cbb826092cd1863c72931aa/python/mxnet/ndarray/ndarray.py#L1501-L1520

Would you like to open a PR to add a check that makes sure the size is consistent before and after the reshape, and raise ValueError otherwise? The process is described at https://cwiki.apache.org/confluence/display/MXNET/Git+Setup+and+Workflow

[r3stl355]

I'm new so forgive my ignorance. The same bug seems to be present v2.0, shall I just give PR a go on the v2 branch?

[leezu]

@r3stl355 you can open the PR against the master branch and v1.x branch

[r3stl355]

This was fixed by #19095 few months ago and merged into v1.x branch but was not picked for 1.8 release. Should be closed.