Why should you care?

AIPredict legacy TF sentiment can migrate to in-process scoring without rewriting training.

See it live — copy this example

Use a .NET console or Web API project with Microsoft.ML. Run dotnet run after pasting.

var tf = ml.Transforms.LoadImages("Image", null, "path")
    .Append(ml.Transforms.ResizeImages("Image", 224, 224))
    .Append(ml.Transforms.ExtractPixels("Pixels", "Image"))
    .Append(ml.Transforms.TensorFlowModelLoader("Scores", "models/saved_model.pb", "Pixels", new[] { "Softmax" }));
var pred = tf.Fit(dummies).Transform(sample);

What happened?

• TensorFlowModelLoader wires TF graphs to ML.NET columns.
• Prefer ONNX for new work unless TF asset already exists.

Practice next

1. Place saved_model.pb.
2. Map input/output tensor names.
3. Fit loader on sample schema.
4. Convert TF export to ONNX.
5. Batch Transform for throughput test.

Remember

TF loader in pipeline. Tensor name map. ONNX preferred for greenfield.

AIPredict TF legacy

Old TF sentiment graph scores in ML.NET.

Outcome: Decommission Python sidecar.