Merge pull request #501 from NiklasGustafsson/bugs

Fixed bugs #499 and #500

RELEASENOTES.md

@@ -2,6 +2,22 @@
 
 Releases, starting with 9/2/2021, are listed with the most recent release at the top.
 
+## NuGet Version 0.96.0
+
+__API Changes:__
+
+__NOTE__: This release contains breaking changes.
+
+'Module.named_parameters()', 'parameters()', 'named_modules()', 'named_children()' all return IEnumerable instances instead of arrays.
+Adding weight and bias properties to the RNN modules.
+Lower-cased names: Module.Train --> Module.train and Module.Eval --> Module.eval
+
+__Fixed Bugs:__
+
+#500 BatchNorm1d throws exception during eval with batch size of 1
+#499 Setting Linear.weight is not reflected in 'parameters()'
+#496 Wrong output shape of torch.nn.Conv2d with 2d stride overload
+
 ## NuGet Version 0.95.4
 
 __API Changes:__
@@ -151,15 +167,15 @@ __Fixed Bugs:__
 
 __Added Features:__
 
-```
+'''
 torch.nn.MultiHeadAttention
 torch.linalg.cond
 torch.linalg.cholesky_ex
 torch.linalg.inv_ex
 torch.amax/amin
 torch.matrix_exp
 torch.distributions.*   (about half the namespace)
-```
+'''
 
 __API Changes:__
 

build/BranchInfo.props

@@ -1,8 +1,8 @@
 <Project>
   <PropertyGroup>
     <MajorVersion>0</MajorVersion>
-    <MinorVersion>95</MinorVersion>
-    <PatchVersion>4</PatchVersion>
+    <MinorVersion>96</MinorVersion>
+    <PatchVersion>0</PatchVersion>
   </PropertyGroup>
 
 </Project>

src/Examples.Utils/Examples.Utils.csproj

@@ -9,7 +9,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="SharpZipLib" Version="1.3.1" />
+    <PackageReference Include="SharpZipLib" Version="1.3.3" />
   </ItemGroup>
 
   <ItemGroup>

src/Examples/AdversarialExampleGeneration.cs

@@ -100,7 +100,7 @@ internal static void Main(string[] args)
                 }
 
                 model.to((Device)device);
-                model.Eval();
+                model.eval();
 
                 var epsilons = new double[] { 0, 0.05, 0.1, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50 };
 

src/Examples/AlexNet.cs

@@ -50,9 +50,7 @@ public AlexNet(string name, int numClasses, torch.Device device = null) : base(n
                 ("l3", Linear(4096, numClasses))
             );
 
-            register_module("features", features);
-            register_module("avg", avgPool);
-            register_module("classify", classifier);
+            RegisterComponents();
 
             if (device != null && device.type == DeviceType.CUDA)
                 this.to(device);

src/Examples/CIFAR10.cs

@@ -155,7 +155,7 @@ internal static void Main(string[] args)
             long batchSize,
             long size)
         {
-            model.Train();
+            model.train();
 
             int batchId = 1;
             long total = 0;
@@ -200,7 +200,7 @@ internal static void Main(string[] args)
             IEnumerable<(Tensor, Tensor)> dataLoader,
             long size)
         {
-            model.Eval();
+            model.eval();
 
             double testLoss = 0;
             long correct = 0;

src/Examples/Examples.csproj

@@ -15,7 +15,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="SharpZipLib" Version="1.3.1" />
+    <PackageReference Include="SharpZipLib" Version="1.3.3" />
     <PackageReference Include="SkiaSharp" Version="2.80.2" />
     <PackageReference Include="System.Memory" Version="4.5.3" />
   </ItemGroup>

src/Examples/MNIST.cs

@@ -170,7 +170,7 @@ public override Tensor forward(Tensor input)
             long batchSize,
             long size)
         {
-            model.Train();
+            model.train();
 
             int batchId = 1;
 
@@ -206,7 +206,7 @@ public override Tensor forward(Tensor input)
             IEnumerable<(Tensor, Tensor)> dataLoader,
             long size)
         {
-            model.Eval();
+            model.eval();
 
             double testLoss = 0;
             int correct = 0;

src/Examples/Properties/launchSettings.json

@@ -2,7 +2,6 @@
   "profiles": {
     "Examples": {
       "commandName": "Project",
-      "commandLineArgs": "VGG11 16 1800",
       "nativeDebugging": true
     }
   }

src/Examples/SequenceToSequence.cs

@@ -100,7 +100,7 @@ internal static void Main(string[] args)
 
         private static void train(int epoch, Tensor train_data, TransformerModel model, Loss criterion, int bptt, int ntokens, torch.optim.Optimizer optimizer)
         {
-            model.Train();
+            model.train();
 
             using (var d = torch.NewDisposeScope()) {
 
@@ -126,7 +126,7 @@ private static void train(int epoch, Tensor train_data, TransformerModel model,
                     using (var output = model.forward(data, src_mask)) {
                         var loss = criterion(output.view(-1, ntokens), targets);
                         loss.backward();
-                        torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5);
+                        torch.nn.utils.clip_grad_norm_(model.parameters().ToArray(), 0.5);
                         optimizer.step();
 
                         total_loss += loss.to(torch.CPU).item<float>();
@@ -145,7 +145,7 @@ private static void train(int epoch, Tensor train_data, TransformerModel model,
 
         private static double evaluate(Tensor eval_data, TransformerModel model, Loss criterion, int bptt, int ntokens, torch.optim.Optimizer optimizer)
         {
-            model.Eval();
+            model.eval();
 
             using (var d = torch.NewDisposeScope()) {
 

src/Examples/TextClassification.cs

@@ -105,7 +105,7 @@ internal static void Main(string[] args)
 
         static void train(int epoch, IEnumerable<(Tensor, Tensor, Tensor)> train_data, TextClassificationModel model, Loss criterion, torch.optim.Optimizer optimizer)
         {
-            model.Train();
+            model.train();
 
             double total_acc = 0.0;
             long total_count = 0;
@@ -123,7 +123,7 @@ static void train(int epoch, IEnumerable<(Tensor, Tensor, Tensor)> train_data, T
 
                     var loss = criterion(predicted_labels, labels);
                     loss.backward();
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5);
+                    torch.nn.utils.clip_grad_norm_(model.parameters().ToArray(), 0.5);
                     optimizer.step();
 
                     total_acc += (predicted_labels.argmax(1) == labels).sum().to(torch.CPU).item<long>();
@@ -141,7 +141,7 @@ static void train(int epoch, IEnumerable<(Tensor, Tensor, Tensor)> train_data, T
 
         static double evaluate(IEnumerable<(Tensor, Tensor, Tensor)> test_data, TextClassificationModel model, Loss criterion)
         {
-            model.Eval();
+            model.eval();
 
             double total_acc = 0.0;
             long total_count = 0;

src/FSharp.Examples/AdversarialExampleGeneration.fs

@@ -121,7 +121,7 @@ let run epochs =
 
     model.``to``(device) |> ignore
 
-    model.Eval()
+    model.eval()
 
     let epsilons = [| 0.0; 0.05; 0.1; 0.15; 0.20; 0.25; 0.30; 0.35; 0.40; 0.45; 0.50|]
 

src/FSharp.Examples/AlexNet.fs

@@ -87,7 +87,7 @@ let loss x y = functional.nll_loss().Invoke(x,y)
 
 let train (model:Model) (optimizer:Optimizer) (dataLoader: CIFARReader) epoch =
 
-    model.Train()
+    model.train()
 
     let size = dataLoader.Size
 
@@ -125,7 +125,7 @@ let train (model:Model) (optimizer:Optimizer) (dataLoader: CIFARReader) epoch =
         end
 
 let test (model:Model) (dataLoader:CIFARReader) =
-    model.Eval()
+    model.eval()
 
     let sz = float32 dataLoader.Size
 

src/FSharp.Examples/FSharp.Examples.fsproj

@@ -20,7 +20,7 @@
   </ItemGroup>
 
   <ItemGroup>
-    <PackageReference Include="SharpZipLib" Version="1.3.1" />
+    <PackageReference Include="SharpZipLib" Version="1.3.3" />
     <PackageReference Include="System.Memory" Version="4.5.3" />
   </ItemGroup>
 

src/FSharp.Examples/MNIST.fs

@@ -87,7 +87,7 @@ type Model(name,device:torch.Device) as this =
 let loss x y = functional.nll_loss(reduction=Reduction.Mean).Invoke(x,y)
 
 let train (model:Model) (optimizer:Optimizer) (dataLoader: MNISTReader) epoch =
-    model.Train()
+    model.train()
 
     let size = dataLoader.Size
     let batchSize = dataLoader.BatchSize
@@ -114,7 +114,7 @@ let train (model:Model) (optimizer:Optimizer) (dataLoader: MNISTReader) epoch =
         batchID <- batchID + 1
 
 let test (model:Model) (dataLoader:MNISTReader) =
-    model.Eval()
+    model.eval()
 
     let sz = float32 dataLoader.Size
 

src/FSharp.Examples/SequenceToSequence.fs

@@ -138,7 +138,7 @@ let get_batch (source:torch.Tensor) (index:int64) =
 
 let train epoch (model:TransformerModel) (optimizer:Optimizer) (trainData:torch.Tensor) ntokens =
 
-    model.Train()
+    model.train()
 
     use d = torch.NewDisposeScope()
 
@@ -183,7 +183,7 @@ let train epoch (model:TransformerModel) (optimizer:Optimizer) (trainData:torch.
 
 let evaluate (model:TransformerModel) (evalData:torch.Tensor) ntokens =
 
-    model.Eval()
+    model.eval()
 
     use d = torch.NewDisposeScope()
 

src/FSharp.Examples/TextClassification.fs

@@ -71,7 +71,7 @@ type TextClassificationModel(vocabSize, embedDim, nClasses, device:torch.Device)
 
 let train epoch (trainData:IEnumerable<torch.Tensor*torch.Tensor*torch.Tensor>) (model:TextClassificationModel) (optimizer:torch.optim.Optimizer) =
 
-    model.Train()
+    model.train()
 
     let mutable total_acc = 0.0
     let mutable total_count = 0L
@@ -103,7 +103,7 @@ let train epoch (trainData:IEnumerable<torch.Tensor*torch.Tensor*torch.Tensor>)
 
 let evaluate (testData:IEnumerable<torch.Tensor*torch.Tensor*torch.Tensor>) (model:TextClassificationModel) =
 
-    model.Eval()
+    model.eval()
 
     let mutable total_acc = 0.0
     let mutable total_count = 0L

src/Native/LibTorchSharp/THSModule.cpp

@@ -70,7 +70,7 @@ void THSNN_Module_register_module(const NNModule module, const char* name, const
 void THSNN_Module_register_parameter(const NNModule module, const char* name, const Tensor tensor, bool requires_grad)
 {
     CATCH(
-        (*module)->register_parameter(name, *tensor, requires_grad);
+        (*module)->register_parameter(name, (tensor == nullptr) ? at::Tensor() : *tensor, requires_grad);
     );
 }