Merge pull request #119 from cbovar/Simplify_shape

Simplify shape

src/ConvNetSharp.Core.Tests/FullyConnLayerTests.cs

@@ -49,10 +49,10 @@ public void Forward()
  layer.Init(inputWidth, inputHeight, inputDepth);
 
  // Make sure filter shape had flatten input shape
- Assert.AreEqual(1, layer.Filters.Shape.GetDimension(0));
- Assert.AreEqual(1, layer.Filters.Shape.GetDimension(1));
- Assert.AreEqual(8, layer.Filters.Shape.GetDimension(2));
- Assert.AreEqual(2, layer.Filters.Shape.GetDimension(3));
+ Assert.AreEqual(1, layer.Filters.Shape.Dimensions[0]);
+ Assert.AreEqual(1, layer.Filters.Shape.Dimensions[1]);
+ Assert.AreEqual(8, layer.Filters.Shape.Dimensions[2]);
+ Assert.AreEqual(2, layer.Filters.Shape.Dimensions[3]);
 
  for (var i = 0; i < 8; i++)
  {

src/ConvNetSharp.Core.Tests/PoolLayerTests.cs

@@ -29,10 +29,10 @@ public void Forward()
  var input = BuilderInstance.Volume.From(data, new Shape(inputWidth, inputHeight, inputDepth, inputBatchSize));
  layer.DoForward(input);
 
- Assert.AreEqual(2, layer.OutputActivation.Shape.GetDimension(0));
- Assert.AreEqual(2, layer.OutputActivation.Shape.GetDimension(1));
- Assert.AreEqual(4, layer.OutputActivation.Shape.GetDimension(2));
- Assert.AreEqual(4, layer.OutputActivation.Shape.GetDimension(3));
+ Assert.AreEqual(2, layer.OutputActivation.Shape.Dimensions[0]);
+ Assert.AreEqual(2, layer.OutputActivation.Shape.Dimensions[1]);
+ Assert.AreEqual(4, layer.OutputActivation.Shape.Dimensions[2]);
+ Assert.AreEqual(4, layer.OutputActivation.Shape.Dimensions[3]);
 
  Assert.AreEqual(5.0, layer.OutputActivation.Get(0,0,0,0));
  Assert.AreEqual(21.0, layer.OutputActivation.Get(0, 0, 1, 0));

src/ConvNetSharp.Core.Tests/SoftMaxLayerTests.cs

@@ -30,10 +30,10 @@ public SoftmaxLayerTests()
  public void OutputIsNormalized()
  {
  var output = this.layer.DoForward(input, true);
- Assert.AreEqual(1, output.Shape.GetDimension(0));
- Assert.AreEqual(1, output.Shape.GetDimension(1));
- Assert.AreEqual(4, output.Shape.GetDimension(2));
- Assert.AreEqual(3, output.Shape.GetDimension(3));
+ Assert.AreEqual(1, output.Shape.Dimensions[0]);
+ Assert.AreEqual(1, output.Shape.Dimensions[1]);
+ Assert.AreEqual(4, output.Shape.Dimensions[2]);
+ Assert.AreEqual(3, output.Shape.Dimensions[3]);
 
  var values = output.ToArray();
  Assert.AreEqual(0.25, values[0]);

src/ConvNetSharp.Core/Fluent/FluentNet.cs

@@ -92,8 +92,8 @@ public int[] GetPrediction()
  }
 
  var activation = softmaxLayer.OutputActivation;
- var N = activation.Shape.GetDimension(3);
- var C = activation.Shape.GetDimension(2);
+ var N = activation.Shape.Dimensions[3];
+ var C = activation.Shape.Dimensions[2];
  var result = new int[N];
 
  for (var n = 0; n < N; n++)

src/ConvNetSharp.Core/Layers/FullyConnLayer.cs

@@ -63,8 +63,8 @@ public override void Backward(Volume<T> outputGradient)
  this.OutputActivationGradients = outputGradient;
 
  // compute gradient wrt weights and data
- using (var reshapedInput = this.InputActivation.ReShape(1, 1, -1, this.InputActivation.Shape.GetDimension(3)))
- using (var reshapedInputGradients = this.InputActivationGradients.ReShape(1, 1, -1, this.InputActivationGradients.Shape.GetDimension(3)))
+ using (var reshapedInput = this.InputActivation.ReShape(1, 1, -1, this.InputActivation.Shape.Dimensions[3]))
+ using (var reshapedInputGradients = this.InputActivationGradients.ReShape(1, 1, -1, this.InputActivationGradients.Shape.Dimensions[3]))
  {
  reshapedInput.ConvolveGradient(
  this.Filters, this.OutputActivationGradients,
@@ -77,7 +77,7 @@ public override void Backward(Volume<T> outputGradient)
 
  protected override Volume<T> Forward(Volume<T> input, bool isTraining = false)
  {
- using (var reshapedInput = input.ReShape(1, 1, -1, input.Shape.GetDimension(3)))
+ using (var reshapedInput = input.ReShape(1, 1, -1, input.Shape.Dimensions[3]))
  {
  reshapedInput.DoConvolution(this.Filters, 0, 1, this.OutputActivation);
  this.OutputActivation.DoAdd(this.Bias, this.OutputActivation);

src/ConvNetSharp.Core/Layers/LayerBase.cs

@@ -67,7 +67,7 @@ public virtual Volume<T> DoForward(Volume<T> input, bool isTraining = false)
 
  this.InputActivation = input;
 
- var outputShape = new Shape(this.OutputWidth, this.OutputHeight, this.OutputDepth, input.Shape.DimensionCount == 4 ? input.Shape.GetDimension(3) : 1);
+ var outputShape = new Shape(this.OutputWidth, this.OutputHeight, this.OutputDepth, input.Shape.Dimensions[3]);
 
  if (this.OutputActivation == null ||
  !this.OutputActivation.Shape.Equals(outputShape))

src/ConvNetSharp.Core/Layers/RegressionLayer.cs

@@ -40,7 +40,7 @@ public override void Backward(Volume<T> outputGradient)
  public override void Backward(Volume<T> y, out T loss)
  {
  var reshape = y.ReShape(new Shape(1, 1, -1, Shape.Keep));
- var dy = this.InputActivationGradients.ReShape(this.OutputActivation.Shape.Dimensions.ToArray());
+ var dy = this.InputActivationGradients.ReShape(this.OutputActivation.Shape.Dimensions);
  reshape.DoSubtractFrom(this.OutputActivation, dy);
 
  if (this._result == null)
@@ -54,7 +54,7 @@ public override void Backward(Volume<T> y, out T loss)
  var half = (T)Convert.ChangeType(0.5, typeof(T));
  this._result.DoMultiply(this._result, half); // dy * dy * 0.5
  this._result.DoSum(this._sum); // sum over all batch
- var batchSize = y.Shape.GetDimension(3);
+ var batchSize = y.Shape.Dimensions[3];
  loss = Ops<T>.Divide(this._sum.Get(0), Ops<T>.Cast(batchSize)); // average
  }
 

src/ConvNetSharp.Core/Layers/SoftMaxLayer.cs

@@ -26,17 +26,17 @@ public SoftmaxLayer(int classCount)
  public override void Backward(Volume<T> y, out T loss)
  {
  // input gradient = pi - yi
- y.DoSubtractFrom(this.OutputActivation, this.InputActivationGradients.ReShape(this.OutputActivation.Shape.Dimensions.ToArray()));
+ y.DoSubtractFrom(this.OutputActivation, this.InputActivationGradients.ReShape(this.OutputActivation.Shape.Dimensions));
 
  //loss is the class negative log likelihood
  loss = Ops<T>.Zero;
- for (var n = 0; n < y.Shape.GetDimension(3); n++)
+ for (var n = 0; n < y.Shape.Dimensions[3]; n++)
  {
- for (var d = 0; d < y.Shape.GetDimension(2); d++)
+ for (var d = 0; d < y.Shape.Dimensions[2]; d++)
  {
- for (var h = 0; h < y.Shape.GetDimension(1); h++)
+ for (var h = 0; h < y.Shape.Dimensions[1]; h++)
  {
- for (var w = 0; w < y.Shape.GetDimension(0); w++)
+ for (var w = 0; w < y.Shape.Dimensions[0]; w++)
  {
  var expected = y.Get(w, h, d, n);
  var actual = this.OutputActivation.Get(w, h, d, n);

src/ConvNetSharp.Core/Net.cs

@@ -68,8 +68,8 @@ public int[] GetPrediction()
  }
 
  var activation = softmaxLayer.OutputActivation;
- var N = activation.Shape.GetDimension(3);
- var C = activation.Shape.GetDimension(2);
+ var N = activation.Shape.Dimensions[3];
+ var C = activation.Shape.Dimensions[2];
  var result = new int[N];
 
  for (var n = 0; n < N; n++)

src/ConvNetSharp.Core/Training/TrainerBase.cs

@@ -27,15 +27,15 @@ protected virtual void Backward(Volume<T> y)
  {
  var chrono = Stopwatch.StartNew();
 
- var batchSize = y.Shape.GetDimension(3);
+ var batchSize = y.Shape.Dimensions[3];
  this.Loss = Ops<T>.Divide(this.Net.Backward(y), Ops<T>.Cast(batchSize));
  this.BackwardTimeMs = chrono.Elapsed.TotalMilliseconds/batchSize;
  }
 
  private void Forward(Volume<T> x)
  {
  var chrono = Stopwatch.StartNew();
- var batchSize = x.Shape.GetDimension(3);
+ var batchSize = x.Shape.Dimensions[3];
  this.Net.Forward(x, true); // also set the flag that lets the net know we're just training
  this.ForwardTimeMs = chrono.Elapsed.TotalMilliseconds/batchSize;
  }
@@ -46,7 +46,7 @@ public virtual void Train(Volume<T> x, Volume<T> y)
 
  Backward(y);
 
- var batchSize = x.Shape.GetDimension(3);
+ var batchSize = x.Shape.Dimensions[3];
  var chrono = Stopwatch.StartNew();
  TrainImplem();
  this.UpdateWeightsTimeMs = chrono.Elapsed.TotalMilliseconds/batchSize;

src/ConvNetSharp.Flow.Tests/OpTests.cs

@@ -334,7 +334,7 @@ public void ReshapeDerivate()
  result = session.Run(diff,
  new Dictionary<string, Volume<T>>
  {
- {"x", NewVolume(new[] {1.0, 2.0, 3.0, 4.0}, Volume.Shape.From(4))},
+ {"x", NewVolume(new[] {1.0, 2.0, 3.0, 4.0}, Volume.Shape.From(4, 1, 1, 1))},
  {"grad", NewVolume(new[] {1.0, 1.0, 1.0, 1.0}, Volume.Shape.From(1, 1, 4, 1))}
  });
  Assert.AreEqual(new Shape(4, 1, 1, 1), result.Shape);

src/ConvNetSharp.Flow/Net.cs

@@ -62,8 +62,8 @@ public List<ParametersAndGradients<T>> GetParametersAndGradients()
  public int[] GetPrediction()
  {
  var activation = this.Op.Evaluate(this.Session);
- var N = activation.Shape.GetDimension(3);
- var C = activation.Shape.GetDimension(2);
+ var N = activation.Shape.Dimensions[3];
+ var C = activation.Shape.Dimensions[2];
  var result = new int[N];
 
  for (var n = 0; n < N; n++)

src/ConvNetSharp.Flow/Ops/Concat.cs

@@ -44,9 +44,9 @@ public override Volume<T> Evaluate(Session<T> session)
  var left = this.Parents[0].Evaluate(session);
  var right = this.Parents[1].Evaluate(session);
 
- var batchSize = Math.Max(left.Shape.GetDimension(3), right.Shape.GetDimension(3));
+ var batchSize = Math.Max(left.Shape.Dimensions[3], right.Shape.Dimensions[3]);
 
- int totalLength = (int)(left.Shape.TotalLength / left.Shape.GetDimension(3) + right.Shape.TotalLength / right.Shape.GetDimension(3));
+ int totalLength = (int)(left.Shape.TotalLength / left.Shape.Dimensions[3] + right.Shape.TotalLength / right.Shape.Dimensions[3]);
  if (this.Result == null || this.lastTotalLength != totalLength)
  {
  this.Result?.Dispose();

src/ConvNetSharp.Flow/Ops/Const.cs

@@ -88,10 +88,10 @@ public override Dictionary<string, object> GetData()
 
  if (this.OutputShape != null)
  {
- data["dim0"] = this.OutputShape.GetDimension(0);
- data["dim1"] = this.OutputShape.GetDimension(1);
- data["dim2"] = this.OutputShape.GetDimension(2);
- data["dim3"] = this.OutputShape.GetDimension(3);
+ data["dim0"] = this.OutputShape.Dimensions[0];
+ data["dim1"] = this.OutputShape.Dimensions[1];
+ data["dim2"] = this.OutputShape.Dimensions[2];
+ data["dim3"] = this.OutputShape.Dimensions[3];
  }
 
  return data;

src/ConvNetSharp.Flow/Ops/Convolution.cs

@@ -91,19 +91,19 @@ public override Volume<T> Evaluate(Session<T> session)
 
  if (this.Parents[1].Result == null)
  {
- var count = this.Width * this.Height * x.Shape.GetDimension(2);
+ var count = this.Width * this.Height * x.Shape.Dimensions[2];
  var scale = Math.Sqrt(2.0 / count);
 
- var filterShape = new Shape(this.Width, this.Height, x.Shape.GetDimension(2), this.FilterCount);
+ var filterShape = new Shape(this.Width, this.Height, x.Shape.Dimensions[2], this.FilterCount);
  this.Parents[1].Result = BuilderInstance<T>.Volume.Random(filterShape, 0.0, scale);
  }
 
  var outputDepth = this.FilterCount;
- var outputWidth = (int)Math.Floor((x.Shape.GetDimension(0) + this.Pad * 2 - this.Width) / (double)this.Stride + 1);
- var outputHeight = (int)Math.Floor((x.Shape.GetDimension(1) + this.Pad * 2 - this.Height) / (double)this.Stride + 1);
+ var outputWidth = (int)Math.Floor((x.Shape.Dimensions[0] + this.Pad * 2 - this.Width) / (double)this.Stride + 1);
+ var outputHeight = (int)Math.Floor((x.Shape.Dimensions[1] + this.Pad * 2 - this.Height) / (double)this.Stride + 1);
 
  this.Result?.Dispose();
- this.Result = BuilderInstance<T>.Volume.SameAs(new Shape(outputWidth, outputHeight, outputDepth, x.Shape.GetDimension(3)));
+ this.Result = BuilderInstance<T>.Volume.SameAs(new Shape(outputWidth, outputHeight, outputDepth, x.Shape.Dimensions[3]));
  }
 
  x.DoConvolution(this.Parents[1].Evaluate(session), this.Pad, this.Stride, this.Result);

src/ConvNetSharp.Flow/Ops/Extract.cs

@@ -38,7 +38,7 @@ public override Volume<T> Evaluate(Session<T> session)
  var length = (int)Convert.ChangeType(this.Parents[1].Evaluate(session).Get(0), typeof(int)); // TODO: Find a way to keep this on host
  var offset = (int)Convert.ChangeType(this.Parents[2].Evaluate(session).Get(0), typeof(int)); // TODO: Find a way to keep this on host
 
- var batchSize = x.Shape.GetDimension(3);
+ var batchSize = x.Shape.Dimensions[3];
 
  int totalLength = length * batchSize;
  if (this.Result == null || this.lastTotalLength != totalLength)

src/ConvNetSharp.Flow/Ops/Max.cs

@@ -41,7 +41,7 @@ public override Volume<T> Evaluate(Session<T> session)
  this.IsDirty = false;
 
  var x = this.Parents[0].Evaluate(session);
- var reshape = x.ReShape(-1, x.Shape.GetDimension(-1));
+ var reshape = x.ReShape(-1, x.Shape.Dimensions[3]);
  var targetShape = new Shape(reshape.Shape);
  targetShape.SetDimension(0, 1);
 

src/ConvNetSharp.Flow/Ops/Pool.cs

@@ -67,10 +67,10 @@ public override Volume<T> Evaluate(Session<T> session)
  this._lastInputShape = new Shape(x.Shape);
 
  var outputShape = new Shape(
- (int) Math.Floor((x.Shape.GetDimension(0) + this.HorizontalPad * 2 - this.Width) / (double) this.HorizontalStride + 1),
- (int) Math.Floor((x.Shape.GetDimension(1) + this.VerticalPad * 2 - this.Height) / (double) this.VerticalStride + 1),
- x.Shape.GetDimension(2),
- x.Shape.GetDimension(3)
+ (int) Math.Floor((x.Shape.Dimensions[0] + this.HorizontalPad * 2 - this.Width) / (double) this.HorizontalStride + 1),
+ (int) Math.Floor((x.Shape.Dimensions[1] + this.VerticalPad * 2 - this.Height) / (double) this.VerticalStride + 1),
+ x.Shape.Dimensions[2],
+ x.Shape.Dimensions[3]
  );
 
  this.Result?.Dispose();

src/ConvNetSharp.Flow/Ops/Reshape.cs

@@ -66,7 +66,7 @@ public override Volume<T> Evaluate(Session<T> session)
  var shape = this.Parents[1].Evaluate(session);
  var s = new[] { shape.Get(0), shape.Get(1), shape.Get(2), shape.Get(3) };
  var t = s.Select(o => Convert.ToInt32(o)).ToArray();
- this._tempShape = new Shape(t);
+ this._tempShape = new Shape(t[0], t[1], t[2], t[3]);
  this._lastBatchSize = session.BatchSize;
  }
 
@@ -82,10 +82,10 @@ public override Dictionary<string, object> GetData()
 
  if (this.OutputShape != null)
  {
- data["dim0"] = this.OutputShape.GetDimension(0);
- data["dim1"] = this.OutputShape.GetDimension(1);
- data["dim2"] = this.OutputShape.GetDimension(2);
- data["dim3"] = this.OutputShape.GetDimension(3);
+ data["dim0"] = this.OutputShape.Dimensions[0];
+ data["dim1"] = this.OutputShape.Dimensions[1];
+ data["dim2"] = this.OutputShape.Dimensions[2];
+ data["dim3"] = this.OutputShape.Dimensions[3];
  }
 
  return data;

src/ConvNetSharp.Flow/Ops/Shape.cs

@@ -5,6 +5,16 @@
 
 namespace ConvNetSharp.Flow.Ops
 {
+ /// <summary>
+ /// Describes the shape of the data.
+ /// Shape always has 4 dimensons: [width, height, class, batch size]
+ /// 
+ /// e.g. A 1D array fits in a volume that has a shape of [1,1,n,1]
+ /// A 2D array fits in a volume that has a shape of [w,h,1,1]
+ /// A 2D array with 3 channels (a color image for example) fits in a volume that has a shape of [w,h,3,1]
+ /// 10 2D arrays (e.g. 10 grayscale images) fits in a volume that a shape of [w,h,1,10]
+ /// </summary>
+ /// <typeparam name="T">type of data (double or float)</typeparam>
  public class Shape<T> : Op<T> where T : struct, IEquatable<T>, IFormattable
  {
  private readonly VolumeBuilder<T> _builder;
@@ -53,14 +63,14 @@ public override Volume<T> Evaluate(Session<T> session)
 
  if (this.Index == -1)
  {
- this.Result.Set(0, Ops<T>.Cast(y.Shape.GetDimension(0)));
- this.Result.Set(1, Ops<T>.Cast(y.Shape.GetDimension(1)));
- this.Result.Set(2, Ops<T>.Cast(y.Shape.GetDimension(2)));
- this.Result.Set(3, Ops<T>.Cast(y.Shape.GetDimension(3)));
+ this.Result.Set(0, Ops<T>.Cast(y.Shape.Dimensions[0]));
+ this.Result.Set(1, Ops<T>.Cast(y.Shape.Dimensions[1]));
+ this.Result.Set(2, Ops<T>.Cast(y.Shape.Dimensions[2]));
+ this.Result.Set(3, Ops<T>.Cast(y.Shape.Dimensions[3]));
  }
  else
  {
- this.Result.Set(0, Ops<T>.Cast(y.Shape.GetDimension(this.Index)));
+ this.Result.Set(0, Ops<T>.Cast(y.Shape.Dimensions[this.Index]));
  }
 
  return base.Evaluate(session);

src/ConvNetSharp.Flow/Ops/SoftMaxCrossEntropy.cs

@@ -44,13 +44,13 @@ public override Volume<T> Evaluate(Session<T> session)
  var outputActivation = this.Parents[0].Evaluate(session);
 
  var loss = Ops<T>.Zero;
- for (var n = 0; n < y.Shape.GetDimension(3); n++)
+ for (var n = 0; n < y.Shape.Dimensions[3]; n++)
  {
- for (var d = 0; d < y.Shape.GetDimension(2); d++)
+ for (var d = 0; d < y.Shape.Dimensions[2]; d++)
  {
- for (var h = 0; h < y.Shape.GetDimension(1); h++)
+ for (var h = 0; h < y.Shape.Dimensions[1]; h++)
  {
- for (var w = 0; w < y.Shape.GetDimension(0); w++)
+ for (var w = 0; w < y.Shape.Dimensions[0]; w++)
  {
  var expected = y.Get(w, h, d, n);
  var actual = outputActivation.Get(w, h, d, n);
@@ -66,7 +66,7 @@ public override Volume<T> Evaluate(Session<T> session)
  }
  }
 
- var batchSize = outputActivation.Shape.GetDimension(3);
+ var batchSize = outputActivation.Shape.Dimensions[3];
  loss = Ops<T>.Divide(Ops<T>.Negate(loss), Ops<T>.Cast(batchSize));
  this.Result.Set(0, loss);