Bug with arange inside a foreach

[matteosal]

symbol.zip

Download+extract the attached json and run this script:

import mxnet as mx

sym = mx.symbol.load('/path/to/symbol.json')
args = {
	".Inputs.Input": mx.nd.zeros([1, 1, 2]),
	".States.Element": mx.nd.zeros([1]),
	"zeroarray0": mx.nd.zeros([1]),
	".Inputs.Position": mx.nd.zeros([1, 1]),
	"fixedarray:.Parameters.Net.Nodes.2:Limit": mx.nd.zeros([1]),
}
ex = sym.bind(mx.cpu(), args=args)
outputs = ex.forward()
print(outputs)

It fails with the following error:

mxnet.base.MXNetError: [15:37:57] src/imperative/./imperative_utils.h:145: Operator _arange inferring shapes failed.
input shapes:
output shapes:
None
operator attributes:
start : 0
dtype : float32
infer_range : true

This example stopped working at this commit: 83d2c2d

Before that, it was producing the expected result:

[
[[0.]]
<NDArray 1x1 @cpu(0)>, 
[0.]
<NDArray 1 @cpu(0)>]

@sbodenstein

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Bug

[junrushao]

Thanks for raising this! I am not super familiar with the arange operator, but through the error message, I saw its attributes like "start = 0, infer_range = true". I got a little bit confused so would love to confirm with you...Is it indicating that it is trying to generate [0, 0)?

Thank you so much!

[matteosal]

Thanks for raising this! I am not super familiar with the arange operator, but through the error message, I saw its attributes like "start = 0, infer_range = true". I got a little bit confused so would love to confirm with you...Is it indicating that it is trying to generate [0, 0)?

Thank you so much!

Thanks for answering.
No, it means it is inferring the end of the interval from its output size, that is the input size of the node it's attached to: operator elemwise_add named .Parameters.Net.Nodes.2$8. So arange should generate [0, n), where n is the input length of .Parameters.Net.Nodes.2$8.
That should, in turn, infer its input shape from the output of the broadcast_like op named .Parameters.Net.Nodes.2$7, but apparently something is going wrong in the process.

[junrushao]

I see. A known limitation for dynamic shape operators is that it does not support reverse shape inference, and control flow operators are now assumed to be dynamic shape operators.

[matteosal]

Ok, so is this to be considered a bug or a feature no more supported? In case it's something intended to work, can you give a time estimate for the fix to happen? Thanks

[leezu]

You may be able to use the following workaround to work around the lack of reverse shape inference

            zeros = F.zeros_like(inputs)
            arange = F.arange(start=0, repeat=1, step=1, infer_range=True, dtype='int32')
            arange = F.elemwise_add(arange, zeros)

https://github.com/dmlc/gluon-nlp/blob/a9eb2fd38180b5affbd30ec78f33f6ef16ef9c20/src/gluonnlp/model/bert.py#L497-L499