go-tensorflow

Unified interface for TensorFlow prediction for both embedded models and calls to Tensorflow Serving. Implementations automatically convert go types into matching TensorFlow Tensors.

Models should be exported in the SavedModel format.

Example

Example uses pre-trained model found under testdata/models wide_deep

import "github.com/Applifier/go-tensorflow/savedmodel"
// import "github.com/Applifier/go-tensorflow/serving"

// Uncomment line below to switch implementation
// predictor := serving.NewPredictor(servingModelClient)
predictor, _ := savedmodel.NewPredictor("testdata/models", "wide_deep", 1527087570, "serving_default")


m := map[string]interface{}{
    "age":            35.0,
    "capital_gain":   0.0,
    "capital_loss":   0.0,
    "education":      "Masters",
    "education_num":  14.0,
    "gender":         "Female",
    "hours_per_week": 29.0,
    "native_country": "United-States",
    "occupation":     "Prof-specialty",
    "relationship":   "Husband",
    "workclass":      "Private",
}

res, modelInfo, _ := predictor.Predict(
    context.Background(),
    map[string]interface{}{
        "inputs": m,
    },
    nil,
)

scores := res["scores"].Value().([][]float32)

fmt.Printf("scores %+v\n", scores[0])

// Output: scores [0.54612064 0.45387936]

fmt.Printf("model name %s, version %s", modelInfo.Name, modelInfo.Version)
// Output: model name wide_deep, version 1527087570

License

MIT