Search Unity

  1. The 2022.1 beta is now available for testing. To find out what's new, have a look at our 2022.1 beta blog post.
    Dismiss Notice

Resolved Confused about input shape

Discussion in 'Barracuda' started by ADNCG, Apr 2, 2021.

  1. ADNCG

    ADNCG

    Joined:
    Jun 9, 2014
    Posts:
    959
    Hi all,

    I have a model that analyzes strokes data from users drawings. The input shape is formatted like the following [batchSize, numberOfPoints(varies between batches), 3] where the channels are the x and the y and a binary value that says wether or not the point is the last one in the stroke.

    In unity, the model says the expected input shape is [n:*, h:1, w:3, c:*] while the original keras model is expecting [*, *, 3]

    I'm having a hard time understanding how I should format my tensor.

    Any ideas?

    EDIT : Tried defining the keras input shape as [batchSize, numberOfPoints, 1, 3] and then squeezing it in the model to see if that'd help unity conversion process but it still expects [*, 1, 3, *], instead of [*, *, 1, 3]

    EDIT2 : If I try to feed the model inputs in the way it expects [batchSize(1), 1, 3, numberOfPoints], I get the error
    along with the warning
    EDIT3 : I tried permuting my datasets and then permuting inside the model back to the original shape, but the shape unity expects changed along with the permutation

    It's becoming clear that I will not make any progress without understanding more about barracuda's inner workings. Investigating... Will report if I make progress.

    EDIT4 [SOLVED] : Realized I misread the barracuda docs. When converting your model from keras to onnx, the following flag is mentioned in the docs :
    I read it as "none of the inputs should be transposed". I was able to get the proper input shape in unity after adding my input to the flag.
     
    Last edited: Apr 4, 2021
    amirebrahimi_unity likes this.
unityunity