Error: The given key 'Rsqrt' was not present in the dictionary.

[mshanak]

I am trying to train a CNN, But when the code reach the model.fit() function it throws the following error

System.Collections.Generic.KeyNotFoundException: 'The given key 'Rsqrt' was not present in the dictionary.'

The training function is:

   public override void Train(List<(NDArray,NDArray,  int)>  examples)
        {

            var boards = examples.Select(x => x.Item1).ToArray();
            NDArray input_boards = np.concatenate(boards).reshape((-1,7,7));
        
            NDArray target_pis = np.concatenate(examples.Select(x => x.Item2).ToArray()).reshape((-1,196));
            NDArray target_vs = np.array(examples.Select(x =>(float) x.Item3).ToArray());
            log.Info("Training Net");
            nnet.model.fit( x: input_boards,
                            y: target_pis,
                            batch_size:64
                            );
        }

How could i solve this issue? do i missing any library for the Rsqrt?

[Oceania2018]

Is it possible to paste the reproduce sample code here?

[mshanak]

This is the CNN definition

using Core;
using Tensorflow;
using Tensorflow.Keras.Layers;
using static Tensorflow.KerasApi;
using Tensorflow.NumPy;
using Tensorflow.Keras.Engine;
using static Tensorflow.Binding;
using Tensorflow.Operations.Activation;

namespace Games.XO
{
    public class SGameNeuralNet 
    {
        public Model model;
        public SGameNeuralNet(GameBase _game,Dictionary<string, dynamic> args)
        {
            var (board_x, board_y) = _game.GetBoardSize();
            var action_size = _game.GetActionSize(); 
            var layers = new LayersApi();
           
            // input layer
            var input_boards = layers.Input(shape: (board_x, board_y));
            var inputs = layers.Reshape( ( board_x, board_y,1 )).Apply(input_boards);// (input_boards);
            //inputs = tf.cast(inputs, dtype: dtypes.float16);
            // convolutional layer
            var h_conv1 = layers.Conv2D(args["num_channels"], 3, padding: "same", activation: "relu",use_bias:false).Apply(inputs);
                h_conv1 = layers.BatchNormalization(name:"n1",axis:3).Apply(h_conv1);

            var h_conv2 = layers.Conv2D(args["num_channels"], 3, padding: "same", activation: "relu", use_bias: false).Apply(h_conv1);
                h_conv2= layers.BatchNormalization(name: "n1", axis: 3).Apply(h_conv2);

            var h_conv3 = layers.Conv2D(args["num_channels"], 3, padding: "valid", activation: "relu", use_bias: false).Apply(h_conv2);
            h_conv3 = layers.BatchNormalization(name: "n1", axis: 3).Apply(h_conv3);


            var h_conv4 = layers.Conv2D(args["num_channels"], 3, padding: "valid", activation: "relu", use_bias: false).Apply(h_conv3);
            h_conv4 = layers.BatchNormalization(name: "n1", axis: 3).Apply(h_conv4);

 
            var h_conv4_flat = layers.Flatten().Apply(h_conv4);

            var s_fc1 = layers.Dense(1024, activation: "relu").Apply(layers.BatchNormalization(axis: 1).Apply(h_conv4_flat));
            var s_fc2 = layers.Dense(512, activation: "relu").Apply(layers.BatchNormalization(axis: 1).Apply(s_fc1));

                   
            var pi =layers.Dense(action_size,activation:"softmax").Apply(s_fc2) ;//Dense(action_size, activation='softmax', name='pi')(s_fc2)   # batch_size x self.action_size
            var v = layers.Dense(1, activation:"tanh").Apply(s_fc2) ;

            // output layer
            var outputs =new Tensor[2]{pi,v} ;//new NDArray() { };//layers.Dense(10).Apply(x);

            // build keras model
            model = keras.Model(input_boards, outputs, name: "toy_resnet");
            model.summary();

            // compile keras model in tensorflow static graph
            model.compile(optimizer: keras.optimizers.RMSprop(1e-3f),
                loss: keras.losses.CategoricalCrossentropy(from_logits: true),
                metrics: new[] { "acc" });



                
        }


    }
}

In another class, i call the fit function :

   public override void Train(List<(NDArray,NDArray,  int)>  examples)
        {

            var boards = examples.Select(x => x.Item1).ToArray();
            NDArray input_boards = np.concatenate(boards).reshape((-1,7,7));
        
            NDArray target_pis = np.concatenate(examples.Select(x => x.Item2).ToArray()).reshape((-1,196));
            NDArray target_vs = np.array(examples.Select(x =>(float) x.Item3).ToArray());
            log.Info("Training Net");
            nnet.model.fit( x: input_boards,
                            y: target_pis,
                            batch_size:64
                            );
        }

[AsakusaRinne]

It seems to be fixed by #980. Please wait for the merge or have a try with the branch of #980. Thank you for your patience.