在ML.NET中加载导出的TensorFlow Keras冻结图模型时遍历预测结果报错的技术求助

Let's break down the potential issues with your code and fix them step by step:

1. Input Data Type Mismatch #

Your TensorFlow model was trained with integer inputs, but there's a high chance the input dtype in the frozen graph is int32 (not int64/long which you're using in ML.NET). TensorFlow is strict about dtype matching, so this mismatch can cause silent failures or errors during inference.

Fix:
Change the input type in TensorData from long[] to int[]:

public class TensorData {
    [ColumnName(@"x")] // Make sure this matches your model's actual input node name!
    [VectorType(2)]
    public int[] input { get; set; } // Use int instead of long
}

2. Output Type Mismatch #

Your Keras model outputs a single value (Dense(units=1)), so the Identity node in the frozen graph returns a scalar tensor per sample. However, you're mapping it to a float[] in OutputScores, which expects an array—this will cause a deserialization error when you try to iterate over results.

Fix:
Change the output type from float[] to float:

class OutputScores {
    [ColumnName(@"Identity")]
    public float output { get; set; } // Single float instead of array
}

3. Verify Input Node Name #

Your ML.NET code uses "x" as the input node name, but TensorFlow often assigns default names like "input_1" to Keras model inputs. If the names don't match, ML.NET can't feed data into the model.

How to check in Python:
Run this snippet to get your model's actual input node name:

import tensorflow as tf

frozen_graph = tf.GraphDef()
with open("test/test.pb", "rb") as f:
    frozen_graph.ParseFromString(f.read())

# List all placeholder (input) nodes
for node in frozen_graph.node:
    if node.op == "Placeholder":
        print(f"Input node name: {node.name}, dtype: {tf.as_dtype(node.attr['dtype'].type)}")

If the input name isn't "x", update the ColumnName attribute in TensorData to match (e.g., [ColumnName(@"input_1")]).

4. Batch Dimension Validation #

You set addBatchDimensionInput: true in ScoreTensorFlowModel, which is correct (you used np.expand_dims in Python to add a batch dimension). Just confirm your VectorType(2) aligns with the model's expected input shape of (None, 2).

Modified Working Code #

After applying the fixes above, your code should look like this (adjust input name if needed):

public class TensorData {
    [ColumnName(@"input_1")] // Replace with your actual input node name
    [VectorType(2)]
    public int[] input { get; set; }
}

public static TensorData[] GetTensorData() {
    return new TensorData[] {
        new() { input = new int[] { 0, 0 } },
        new() { input = new int[] { 12, 10 } },
        new() { input = new int[] { 12, 800 } }
    };
}

class OutputScores {
    [ColumnName(@"Identity")]
    public float output { get; set; }
}

public static void Test() {
    var mlContext = new MLContext();
    // Enable logging for debugging if needed
    mlContext.Log += (sender, e) => Console.WriteLine($"ML.NET Log: {e.Message}");
    
    var data = GetTensorData();
    var idv = mlContext.Data.LoadFromEnumerable(data);
    
    using var model = mlContext.Model.LoadTensorFlowModel(@"path to your test.pb");
    var pipeline = model.ScoreTensorFlowModel(
        outputColumnNames: new[] { "Identity"}, 
        inputColumnNames: new[] { "input_1" }, // Match input node name
        addBatchDimensionInput: true);
    
    var estimator = pipeline.Fit(idv);
    var transformedValues = estimator.Transform(idv);
    var outScores = mlContext.Data.CreateEnumerable<OutputScores>(transformedValues, reuseRowObject: false);
    
    foreach (var prediction in outScores) {
        Console.WriteLine($"Prediction: {prediction.output}");
    }
}

Final Checks #

• Use an absolute path to your test.pb file during debugging to avoid path resolution issues.
• If errors persist, the ML.NET logging (added in the code above) will give you detailed insights into what's going wrong during model loading or inference.

内容的提问来源于stack exchange，提问作者tixina3684