System information

• OS version/distro: Windows 10 1809
• .NET Version (eg., dotnet --info): 2.2.202

Issue

I recently tried to use ML.NET in a Xamarin-based UWP app. I targeted the earliest version of UWP that support .NET Standard 2.0 and everything installed correctly. My intention was to make use a separately-trained TensorFlow model to predict some data from an image.

While my code worked without issues in a .NET Core 2.2 console app, in the UWP it was failing. I was constantly getting an exception when attempting to load the target image - Image [whatever] was not found. No additional details, no inner exception.

Only after a fair amount of headscratching and trying various things I managed to find the culprit (which was obvious in hindsight): UWP does not support bitmaps. After I tried to replicate a little bit of the ImageLoader code, I got an unsupported platform exception.

This is all fair but then looking into the code some more, I found this:

The code is catching all exceptions and throwing a custom exception in their place, without providing the base exception. I think this could be improved as part of making the API surface friendlier to use - spending an hour on this, I started to think I was doing something insanely wrong. The inner exception would have told me the root cause in 10 seconds flat.

So there are basically two issues:

1. Hiding the base exceptions
2. Image analytics won't work in UWP at all because of bitmaps

I think a fix for 1. could be relatively simple. 2. will be much more difficult and might not be desirable but I wanted to throw it out there. Maybe using a platform-agnostic implementation of image handling could be useful.

Opinions? Thoughts?