Gpu memory limit

[songminkyu]

Hello. I want to limit C # tensorflow gpu memory, but I want to know if it is possible

tensorflow/tensorflow#25138

[Superpiffer]

Of course you can:

var config = new ConfigProto
{
    GpuOptions = new GPUOptions
    {
        AllowGrowth = true,
        PerProcessGpuMemoryFraction = 0.5
    }
};

var session = tf.Session(graph, config);

[abhikirk]

Hi @Superpiffer , I have the same issue as @songminkyu even with your proposed solution. Almost all of the GPU memory is being allocated no matter what PerProcessGpuMemoryFraction is chosen, and no matter the values of AllowGrowth boolean.

I am using:

• M1000M GPU (compute capability = 5),
• SciSharp.TensorFlow.Redist-Windows-GPU v1.15.1,
• SharpCV v0.4.5,
• TensorFlow.NET v0.15.1,
• Code: CnnInYourOwnData.cs

Modifications:
Added

public ConfigProto GPUConfig()
        {
            ConfigProto config = new ConfigProto();
            GPUOptions gpuOptions = new GPUOptions();
            gpuOptions.AllowGrowth = true;
            //gpuOptions.PerProcessGpuMemoryFraction = 0.5;
            config.GpuOptions = gpuOptions;
            return config;
        }

With

private NDArray ReadTensorFromImageFile(string file_name)
        {
            using (var graph = tf.Graph().as_default())
            {
                Tensor file_reader = tf.read_file(file_name, "file_reader");
                Tensor decodeJpeg = tf.image.decode_jpeg(file_reader, channels: n_channels, name: "DecodeJpeg");
                Tensor cast = tf.cast(decodeJpeg, tf.float32);
                Tensor dims_expander = tf.expand_dims(cast, 0);
                Tensor resize = tf.constant(new int[] { img_h, img_w });
                Tensor bilinear = tf.image.resize_bilinear(dims_expander, resize);
                Tensor sub = tf.subtract(bilinear, new float[] { img_mean });
                Tensor normalized = tf.divide(sub, new float[] { img_std });
                
                using (Session sess = tf.Session(graph, GPUConfig()))
                {
                    return sess.run(normalized);
                }
            }
        }

And:

public void Run()
        {
            PrepareData();
            BuildGraph();           

            using (Session sess = tf.Session(GPUConfig()))
            {
                Train(sess);
                //Test(sess);
                Console.ReadLine();
            }
        }

I have also checked my GPU memory allocation separately with TensorFlow/Keras in Python, and there the GPU allocation works correctly.

Is there a bug in TensorFlow.NET or am I doing something wrong?
Thanks.

[Oceania2018]

Been fixed in latest release.