Tiny YOLOv2 produces incorrect output with v1.8.0

Bug Report

Which model does this pertain to?

Tiny YOLOv2

Describe the bug

The output when running tiny yolov2 through an ONNX runtime inference session appears to be incorrect, and outputs a large number of warnings. The code being used is given in this gist: https://gist.github.com/avarsh/cdbab0cc635fad6a403a5a616aff639e

When running this code with the threshold value set to "0.002" in the postprocessing, the following output is given: https://gist.github.com/avarsh/d5f4efd28e4e927ebcaad06163dff801

I was able to manually generate an ONNX file for tiny yolov2 from the relevant Darknet files, which gives the correct output with the same code (post processing threshold set to 0.3):

(1, 3, 416, 416)
(125, 13, 13)
Bounding Box 1 : [[75, 97, 163, 371], 0.7742048036087853, 'person']
Bounding Box 2 : [[2, 262, 86, 354], 0.7187669842556927, 'sheep']
Bounding Box 3 : [[283, 151, 409, 335], 0.6768828930128936, 'sheep']
Bounding Box 4 : [[305, 155, 410, 323], 0.5562674827033209, 'sheep']

Note that the input image for both models was: https://github.com/pjreddie/darknet/blob/master/data/person.jpg

Reproduction instructions

System Information

OS Platform and Distribution (e.g. Linux Ubuntu 16.04): Arch Linux
ONNX version (e.g. 1.6): 1.8
Backend/Runtime version (e.g. ONNX Runtime 1.1, PyTorch 1.2): Onnx Runtime 1.5.2

Start by running the inference code from the linked gist with the Tiny YOLOv2 ONNX model, using the supplied person.jpg input and comparing it with the manually generated ONNX file. Done means identifying why the v1.8.0 model produces warnings and incorrect detections, then confirming output comparable to the reported correct bounding boxes.