Machinelearning: Cannot load a trained model: "A local file header is corrupt"

[larsbeck]

System information

• OS version/distro: Windows 10
• .NET Version (eg., dotnet --info): Microsoft.NETCore.App 3.0.0-preview4-27615-11

Issue

Cannot load a trained model.

• What did you do?
  Trained a model with AutoML, saved it to "model.zip" (see attached) and then I tried to load it via:
  var mlContext = new MLContext(); var mlModel = mlContext.Model.Load("model.zip", out _);
• What happened?
  An exception is thrown: "A local file header is corrupt"
• What did you expect?
  The model to be loaded so I can make a prediction.
  

  Source code / logs

model.zip

Please paste or attach the code or logs or traces that would be helpful to diagnose the issue you are reporting.

[sfilipi]

@justinormont is this a supported scenario, atm, to your knowledge?

[justinormont]

The model.zip should be a completely standard ML.NET model file once it's trained.

@vinodshanbhag, @daholste: any thoughts?

[larsbeck]

In case it is of any help. I can reproduce the error with a model I trained with "SdcaLogisticRegression" too (not only AutoML). Training and saving the model works fine, but loading it results in the same error.

[glebuk]

Hmm.. what overload do you use the save the model? If you use stream, make sure the stream is flushed and disposed properly (within using(){} statement).
Can you check to see if the zip file can be opened by the explorer or other zip file reader? Is it corrupt?

[larsbeck]

@glebuk Yes, the stream is flushed. It is encapsulated in a using block. The zip can be opened fine (see attached).

[larsbeck]

I found the issue when I tried to make a minimal sample. The stream was flushed and I was able to open the zip file (the first time), but then after restarting that didn't work anymore, because I was accidentally appending to the zip instead of creating a new file everytime. Closing this issue.