Unknown shape in flatten op.

[roywei]

Right now there is no way to represent unknown shape in MXNet

According to #14253, now it use -1 instead of 0 as unknown dim. This should only take effect if mx.npx.set_np() is used.

I have the following use case when I want to bind a shape with unknown batch size, and feed different batch size data during forward.

import mxnet as mx
a = mx.sym.Variable("a")
b = mx.sym.Flatten(a)
exec = b.simple_bind(ctx=mx.cpu(), a=(0,10,2))

if give the following error:

Traceback (most recent call last):
  File "//anaconda3/lib/python3.7/site-packages/mxnet/symbol/symbol.py", line 1780, in simple_bind
    ctypes.byref(exe_handle)))
  File "//anaconda3/lib/python3.7/site-packages/mxnet/base.py", line 246, in check_call
    raise get_last_ffi_error()
mxnet.base.MXNetError: Traceback (most recent call last):
  File "src/executor/../common/exec_utils.h", line 391
MXNetError: InferShape pass cannot decide shapes for the following arguments (-1 means unknown dimensions). Please consider providing them as inputs:
a: [-1,10,2],

But if I use -1, it also fails no matter if set_np() is used.

>>> exec = b.simple_bind(ctx=mx.cpu(), a=(-1, 10,2))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "//anaconda3/lib/python3.7/site-packages/mxnet/symbol/symbol.py", line 1758, in simple_bind
    array('I', provided_arg_shape_data)),
OverflowError: can't convert negative value to unsigned int

[apeforest]

It works with the np_shape API. @reminisce @haojin2 Is this by design that requires the np_shape API for backward compatibility? Thanks.

import mxnet as mx
with mx.np_shape(active=True):
    a = mx.sym.Variable("a")
    b = mx.sym.Flatten(a)
    exec = b.simple_bind(ctx=mx.cpu(), a=(0,10,2))

[roywei]

I think it still fails for unknow shape case. The above code works fine for 0 dim, but use -1 to represent unknown shape still fails:

>>> with mx.np_shape(active=True):
...     a = mx.sym.Variable("a")
...     b = mx.sym.Flatten(a)
...     exec = b.simple_bind(ctx=mx.cpu(), a=(-1,10,2))
...     exec.forward(a=mx.nd.ones((32,10,2)))
...
Traceback (most recent call last):
  File "<stdin>", line 4, in <module>
  File "//anaconda3/lib/python3.7/site-packages/mxnet/symbol/symbol.py", line 1754, in simple_bind
    array('I', provided_arg_shape_data)),
OverflowError: can't convert negative value to unsigned int