search

CESARDELATORRE / TensorFlowImageClassificationWebApp

Web App (Razor) using WebAPI methods that run a TensorFlow model (.pb - Inception v3) with ML.NET API
MIT License
14 stars 12 forks source link

Suggestions: Add some infos in Readme #1

Closed codeclash closed 5 years ago

codeclash commented 5 years ago

  1. When running/debugging the app from Visual Studio, the server port might be a random port: One must adapt that port (and maybe even switch to http instead of https) in upload.js

  2. I trained a multiclass multilabel model in https://www.customvision.ai and downloaded the tensorflow model. Upon testing with my images, I noticed (after correcting the width/height with https://lutzroeder.github.io/netron/ according to your article, I got predictions, but all with very low score for each class. I think it was due to the image offset: After setting mean = 0 here it worked like expected.

Thx for providing this nice and tidy sample btw!

CESARDELATORRE commented 5 years ago

Have you tried the final version which is public at the official samples repo? Here: https://github.com/dotnet/machinelearning-samples/tree/master/samples/csharp/end-to-end-apps/DeepLearning_ImageClassification_TensorFlow I think that one shouldn't have the port issue?

bamurtaugh commented 5 years ago

Hi @codeclash - we have updated the official sample's readme about the need to potentially change image settings when using a custom Tensor Flow model. Thanks for the feedback!

codeclash commented 5 years ago

Just looked at it now that I saw the link. Yep using a relative url in js for the fetch is actually simpler 👍

codeclash commented 5 years ago

@bamurtaugh that was fast 👌

bamurtaugh commented 5 years ago

@codeclash We are doing our best to monitor issues and update our samples repo to best help the community 😄 Feel free to provide us any other feedback in our repo!

CESARDELATORRE commented 5 years ago

Closing this issue. 👍