search

onnx / models

A collection of pre-trained, state-of-the-art models in the ONNX format
http://onnx.ai/models/
Apache License 2.0
7.77k stars 1.39k forks source link

input pb data for may not be what people expect #144

Open YanYas opened 5 years ago

YanYas commented 5 years ago

Hi, having had difficulty getting satisfying results from tiny yolo I tried loading and previewing the example data I was a bit surprised when I saw this was the result of the 416x416 image:

protobufattempt-DirectX Renderer_2019 03 18-22 24 29

So after some playing around with the pixel width, I retrieved this (500x332):

protobufattempt-DirectX Renderer_2019 03 18-22 23 41

So I guess the image was supposed to be resized or cropped but that hasn't happened here. Should this be corrected? Or would you have to throw out your test results? I mean, the Output.pb results may match what should get but that only counts as a test rather than an example because random colours are not what people will be using these models on.

YanYas commented 5 years ago

I'm not sure what squeezenet should be interpreting from here either:

protobufattempt-DirectX Renderer_2019 03 18-22 40 51

What should the correct pixel dimension be in this case, if at all?

snnn commented 5 years ago

No. It's not. It's random data, meaningless.