[F#] Report merge of BinaryTestClassification

dedale · Oceania2018 · commit d75f35478308 · 2021-04-11T15:22:58.000-05:00
diff --git a/src/TensorFlowNET.Examples.FSharp/ImageProcessing/MnistCnnKerasSubclass.fs b/src/TensorFlowNET.Examples.FSharp/ImageProcessing/MnistCnnKerasSubclass.fs
@@ -79,7 +79,7 @@ module MnistCnnKerasSubclass =
 
             // Fully connected layer.
             let fc1 = layers.Dense(1024) :> Layer
-            // Apply Dropout (if is_training is False, dropout is not applied).
+            // Apply Dropout (if training is False, dropout is not applied).
             let dropout = layers.Dropout(rate = 0.5f) :> Layer
 
             // Output layer, class prediction.
@@ -99,19 +99,19 @@ module MnistCnnKerasSubclass =
             cnn
 
         //let state = defaultArg state null
-        //let is_training = defaultArg is_training false
 
-        override x.Call(inputs, state, is_training) =
+        override x.Call(inputs, state, training) =
+            let training = defaultArg (Option.ofNullable training) false
             let inputs = tf.reshape(inputs.asTensor, TensorShape (-1, 28, 28, 1)).asTensors
             let inputs = conv1.Apply(inputs)
             let inputs = maxpool1.Apply(inputs)
             let inputs = conv2.Apply(inputs)
             let inputs = maxpool2.Apply(inputs)
             let inputs = flatten.Apply(inputs)
             let inputs = fc1.Apply(inputs)
-            let inputs = dropout.Apply(inputs, is_training = is_training)
+            let inputs = dropout.Apply(inputs, training = training)
             let inputs = output.Apply(inputs)
-            if not is_training then tf.nn.softmax(inputs.asTensor).asTensors else inputs
+            if not training then tf.nn.softmax(inputs.asTensor).asTensors else inputs
 
     let cross_entropy_loss x y =
         // Convert labels to int 64 for tf cross-entropy function.
@@ -123,7 +123,7 @@ module MnistCnnKerasSubclass =
 
     let run_optimization (conv_net : ConvNet) (optimizer : OptimizerV2) (x : Tensor) y =
         use g = tf.GradientTape()
-        let pred = conv_net.Apply(x.asTensors, is_training = true)
+        let pred = conv_net.Apply(x.asTensors, training = true)
         let loss = cross_entropy_loss pred.asTensor y
 
         // Compute gradients.
diff --git a/src/TensorFlowNET.Examples.FSharp/NeuralNetworks/FullyConnectedKeras.fs b/src/TensorFlowNET.Examples.FSharp/NeuralNetworks/FullyConnectedKeras.fs
@@ -16,6 +16,8 @@ limitations under the License.
 
 namespace TensorFlowNET.Examples.FSharp
 
+open System
+
 open NumSharp
 open Tensorflow
 open Tensorflow.Keras
@@ -93,11 +95,12 @@ module FullyConnectedKeras =
             x.Call(inputs, null, false)
 
         // Set forward pass.
-        override _.Call(inputs : Tensors, state : Tensor, is_training : bool) =
+        override _.Call(inputs : Tensors, state : Tensor, training : Nullable<bool>) =
+            let training = defaultArg (Option.ofNullable training) false
             let inputs = fc1.Apply(inputs)
             let inputs = fc2.Apply(inputs)
             let inputs = output.Apply(inputs)
-            if not is_training then
+            if not training then
                 tf.nn.softmax(inputs.asTensor).asTensors
             else
                 inputs
@@ -140,7 +143,7 @@ module FullyConnectedKeras =
             // Wrap computation inside a GradientTape for automatic differentiation.
             use g = tf.GradientTape()
             // Forward pass.
-            let pred = neural_net.Apply(x.asTensors, is_training = true)
+            let pred = neural_net.Apply(x.asTensors, training = true)
             let loss = cross_entropy_loss pred.asTensor y
 
             // Compute gradients.
@@ -156,13 +159,13 @@ module FullyConnectedKeras =
             run_optimization batch_x batch_y
 
             if step % display_step = 0 then
-                let pred = neural_net.Apply(batch_x.asTensors, is_training = true)
+                let pred = neural_net.Apply(batch_x.asTensors, training = true)
                 let loss = cross_entropy_loss pred.asTensor batch_y
                 let acc = get_accuracy pred.asTensor batch_y
                 print($"step: {step}, loss: {float32 loss}, accuracy: {float32 acc}")
 
         // Test model on validation set.
-        let pred = neural_net.Apply(x_test.asTensor.asTensors, is_training = false)
+        let pred = neural_net.Apply(x_test.asTensor.asTensors, training = false)
         let accuracy = float32 (get_accuracy pred.asTensor y_test.asTensor)
         print($"Test Accuracy: {accuracy}")
 
diff --git a/src/TensorFlowNET.Examples.FSharp/TensorFlowNET.Examples.FSharp.fsproj b/src/TensorFlowNET.Examples.FSharp/TensorFlowNET.Examples.FSharp.fsproj
@@ -35,11 +35,11 @@
     <PackageReference Include="Argu" Version="6.1.1" />
     <PackageReference Include="Colorful.Console" Version="1.2.15" />
     <PackageReference Include="Newtonsoft.Json" Version="12.0.3" />
-    <PackageReference Include="SciSharp.TensorFlow.Redist" Version="2.3.1" />
-    <PackageReference Include="SharpCV" Version="0.6.0" />
+    <PackageReference Include="SciSharp.TensorFlow.Redist" Version="2.4.1" />
+    <PackageReference Include="SharpCV" Version="0.7.0" />
     <PackageReference Include="OpenCvSharp4.runtime.win" Version="4.4.0.20200915" />
     <PackageReference Include="System.Drawing.Common" Version="5.0.0" />
-    <PackageReference Include="TensorFlow.Keras" Version="0.4.0" />
+    <PackageReference Include="TensorFlow.Keras" Version="0.5.0" />
   </ItemGroup>
 
   <ItemGroup>
