Resize with mode linear is missing output elements

[diyessi]

Describe the bug
Resize with mode="linear" and a specified output shape has garbage in the last output tensor elements. I see this for multiple coordinate transformation modes and input and output shapes. For a trivial example, with an input size of [1,1,1,1] on a float input tensor whose single element is 1.0, resizing to [1, 2, 2, 1] I get output

[[[[[ 1.]
    [ 1.]]

   [[ 5.]
    [10.]]]]]

where the [5, 10] is wrong. I am running a number of these tests, using i^2+1 as inputs, so these particular outputs seem likely to be uninitialized memory from earlier inputs. I also see bunches of NaNs in the bad elements.

Urgency
None

System information

• Ubuntu 20.04.4 LTS (GNU/Linux 5.11.0-1028-azure x86_64)
• ONNX Runtime installed from binary (pip onnxruntime 1.11.1)
• ONNX Runtime version: 1.11.1
• Python version: 3.8.10
• CUDA/cuDNN version: None
• GPU model and memory: None

To Reproduce
I can give you a branch of a simple repo if you need one.

Expected behavior
In this case, all elements should be 1.

Additional context
nearest mode doesn't seem to have any problems. I was actually trying to use ONNX Runtime to generate gold values for testing my own implementation for an accelerator and for creating a doc PR for the ONNX specification.

[ytaous]

@hariharans29 @yihonglyu - any comment/feedback is more than welcome.

[yihonglyu]

@diyessi Could you provide simple reproducer? Thanks.

[diyessi]

import onnxruntime as ort
import numpy as np

model = b'\x08\x08\x12\x0cONNX Builder:\xd7\x01\nF\x12\nConstant_1\x1a\nConstant_0"\x08Constant*"\n\x05value*\x16\x08\x04\x10\x07:\x04\x01\x02\x02\x01B\nConstant_1\xa0\x01\x04\nD\n\x05input\n\x00\n\x00\n\nConstant_1\x12\x06output\x1a\x08Resize_2"\x06Resize*\x11\n\x04mode"\x06linear\xa0\x01\x03\x12\x04testZ\x1f\n\x05input\x12\x16\n\x14\x08\x01\x12\x10\n\x02\x08\x01\n\x02\x08\x01\n\x02\x08\x01\n\x02\x08\x01b \n\x06output\x12\x16\n\x14\x08\x01\x12\x10\n\x02\x08\x01\n\x02\x08\x02\n\x02\x08\x02\n\x02\x08\x01B\x02\x10\x10'
sess = ort.InferenceSession(model)
sess.run(['output'], {'input': ort.OrtValue.ortvalue_from_numpy(np.ones([1,1,1,1], dtype=np.float32))})

produces

[array([[[[1.],
          [1.]],
 
         [[0.],
          [0.]]]], dtype=float32)]

I've also seen NaN's in the output where those 0s are.
It also happens for more realistic shapes.

[satyajandhyala]

Could not reproduce using onnxruntime 1.14 or 1.15.