Add DataView Type Register (supporting loading in-memory image and other in-memory custom types)#3263
Conversation
Codecov Report
@@ Coverage Diff @@
## master #3263 +/- ##
=========================================
Coverage ? 72.64%
=========================================
Files ? 807
Lines ? 145126
Branches ? 16220
=========================================
Hits ? 105424
Misses ? 35281
Partials ? 4421
|
TomFinley
left a comment
There was a problem hiding this comment.
Hi @wschin, thanks for working on this!
I do not consider it acceptable that our answer to the problem of what prediction engine does with its reflection based mechanism is that this core assembly has to have a direct, actual dependence on the type -- quite apart from the already excellent objections raised by @Ivanidzo4ka and @eerhardt, this is incompatible with IDataView's own principle as having an extensible type system.
I will post more details on what I'd consider at least a more correct alternative in the existing thread.
Register image type
|
The PR looks great, but we will need a sample for this work. It's fine if it comes in another PR. Could you adapt a test and make it a sample? #Resolved |
|
Also would it be useful to surface the |
|
For this advanced functionality, I'd keep it as is for now. If we add an alias for it, we will have to support it for years. In reply to: 496640481 [](ancestors = 496640481) |
|
Added two samples. In reply to: 496639086 [](ancestors = 496639086) |
|
Hi! |
|
Hello! I've tried some of your examples, but none of them worked for me. Please let us know if you manage to fully support TensorFlow integration with ML.NET. Regards, |
|
Phew, it's not just me going mad. I've been trying to use in memory images all day and have so far failed to get any of this to work. I have SO question open as well in case anyone wants to follow. |
|
@jamsoft Did you remove the SO question? I'm also looking for an ML.Net sample for image classification that takes a bitmap directly rather than a folder of images. Seems sad that so many people are asking for this and MSFT's ML.Net team still haven't put together a useful sample. |
|
@Webreaper I thought we had a sample or documentation about it, but I just went and looked and couldn't find anything myself. I'll make sure we have something. The basic idea is that you have 2 pipelines. 1 pipeline loads the images and stops when it has the images loaded as bytes, and the second pipeline starts from where you would use these image bytes already. I can' upload a zip file here since its not a github issue, but I do have a sample interactive notebook that shows it. If you can create a new issue I can get attach it for you there. |
|
Thanks @michaelgsharp. Hopefully this should be sufficient? #5980 |
Fix #3162 by considering image type in our bridge code. Also fix #3723, fix #3369, fix #3274, fix #445, fix #3460, fix #2121, fix #2495, fix #3784.