Memory leak

[davipeag]

System information

• Windows 10
• .net 5.0
• ONNX Runtime installed from nugget
• ONNX Runtime version: 1.5.2
• Microsoft.ML version: 1.5.2
• CPU only

Issue

Even after disposing the Microsoft.ML.PredictionEngine and Microsoft.ML.Transforms.Onnx.OnnxTransformer objects, memory is not freed

Source code

using System;
using System.Collections.Generic;
using Microsoft.ML.Data;
using Microsoft.ML;

namespace Test
{
    class LstmCellInput
    {
        [ColumnName("attributes")]
        [VectorType(1,1,8)]
        public float[] Attributes { get; set; }

        [ColumnName("h_in")]
        [VectorType(1, 1, 16)]
        public float[] HiddenState { get; set; }

        [ColumnName("c_in")]
        [VectorType(1, 1, 16)]
        public float[] CellState { get; set; }
    }

    class LSTMCellOutput
    {
        [ColumnName("prediction")]
        [VectorType(1, 1, 1)]
        public float[] Prediction { get; set; }

        [ColumnName("h_out")]
        [VectorType(1, 1, 16)]
        public float[] HiddenState { get; set; }

        [ColumnName("c_out")]
        [VectorType(1, 1, 16)]
        public float[] CellState { get; set; }
    }

    class Program
    {
        static void Main(string[] args)
        {
            Console.WriteLine("start");

            var x = new LstmCellInput {
                HiddenState = new float[16],
                CellState = new float[16],
                Attributes = new float[8]
            };

            for (int i = 0; i < 5000; i++)
            {
                var context = new MLContext();
                var transforms = context.Transforms.ApplyOnnxModel(
                    modelFile: "celllstm.onnx",
                    inputColumnNames: new String[] { "attributes", "h_in", "c_in" },
                    outputColumnNames: new String[] { "prediction", "h_out", "c_out" }
                );
                
                var model = transforms.Fit(context.Data.LoadFromEnumerable(new List<LstmCellInput>()));
                var engine = context.Model.CreatePredictionEngine<LstmCellInput, LSTMCellOutput>(model);
                var y = engine.Predict(x);
                model.Dispose();
                engine.Dispose();
            }

            Console.WriteLine("finish");
        }
    }
}

Additional Info

The onnx model was converted from pytorch
This sample program along with the onnx model is attached
I originally posted this question to ONNX RUNTIME (microsoft/onnxruntime#7303)

Test (1).zip

[davipeag]

It seems like the problem has been solved in the latest version of the libraries. Thank you!