[GitHub] [incubator-tvm] masahi commented on issue #5243: [Frontend][TensorFlow]Improve TensorFlow Static Shape Tensor Array

[GitBox <gi...@apache.org>]

masahi commented on issue #5243: [Frontend][TensorFlow]Improve TensorFlow Static Shape Tensor Array
URL: https://github.com/apache/incubator-tvm/pull/5243#issuecomment-609552445
 
 
   Hi @kevinthesun, I started experimenting with how to integrate static tensor array in Torch frontend. My use case is to support Python tensor list append and stack. I got two problems below:
   
   1. When I cons the tensor to tensor array, I can do infer shape on input tensor to get the fixed shape static tensor array expects. But after I've done some cons and try to stack the static tensor array, I don't have a way to tell what fixed shape the input tensor array to stack expects. See
   https://github.com/masahi/tvm/blob/support-more-rnn/python/tvm/relay/frontend/pytorch.py#L990
   Since the shape is fixed, I think there should be an easy way to query the shape associated with a static array. I see you have such function `check_tensor_array_shape` in this PR (by parsing op name). Is this the recommended way?
   
   2. The output type of stack is currently `static_tensor_float32_?_2_4_t[]` in my test. Is there a way to easily unwrap static tensor type wrapper and get  relay `Tensor`? @wweic had such unwrapper in https://github.com/apache/incubator-tvm/pull/4325 for generic arrays. We should have something equivalent for static arrays.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services