Fixes for backpropogation

src/main.cpp

@@ -38,7 +38,7 @@ public:
       : Layer(layer), inputFeatures(inputFeatures),
         weights(layer.getOuputFeatures(), inputFeatures) {}
   ConnectedLayer(const Layer &a, const Layer &b)
-      : ConnectedLayer(b.getOuputFeatures(), a) {}
+      : ConnectedLayer(a.getOuputFeatures(), b) {}
 
   int getInputFeatures() const { return inputFeatures; }
   const Matrix &getWeights() const { return weights; }
@@ -47,7 +47,6 @@ public:
 
 class LearnLayer : public ConnectedLayer {
 protected:
-  // Matrix gradients;
   Matrix internal;
   Matrix outputs;
 
@@ -57,7 +56,7 @@ public:
         internal(layer.getOuputFeatures(), inputFeatures, false),
         outputs(layer.getOuputFeatures(), inputFeatures, false) {}
   LearnLayer(const Layer &a, const Layer &b)
-      : LearnLayer(b.getOuputFeatures(), a) {}
+      : LearnLayer(a.getOuputFeatures(), b) {}
 
   const Matrix &getInternal() const { return internal; }
   const Matrix &getOutputs() const { return outputs; }
@@ -101,16 +100,16 @@ public:
     // employ back
     layers.push_back(LearnLayer(inputFeatures, l[0]));
     for (size_t i = 1; i < l.size(); i++)
-      layers.push_back(LearnLayer(l[i - 1].getOuputFeatures(), l[i]));
+      layers.push_back(LearnLayer(l[i - 1], l[i]));
   }
 
   Matrix learn(Matrix inputs, Matrix target, float speed = 1.0f) {
     MatrixMath mm;
     VectorMath vm;
     for (size_t i = 0; i < layers.size(); i++) {
-      layers[i].setInternal(mm.dot(i == 0 ? inputs : layers[i - 1].getOutputs(),
+      layers[i].setInternal(mm.dot(layers[i].getWeights(),
-                                   layers[i].getWeights(), false, true,
+                                   i == 0 ? inputs : layers[i - 1].getOutputs(),
-                                   &layers[i].getBias()));
+                                   false, false, &layers[i].getBias()));
       layers[i].setOutputs(mm.activate(layers[i].getInternal(),
                                        layers[i].getActivation(),
                                        layers[i].getAlpha()));
@@ -120,17 +119,26 @@ public:
     std::vector<float> io = inputs.toVector();
     std::cout << "I: ";
     for (size_t i = 0; i < io.size(); ++i)
-      printf("%5.3f ", io[i]);
+      printf("%4.2f ", io[i]);
+
+    std::vector<float> ni = layers[layers.size() - 1].getInternal().toVector();
+    std::cout << "| NNI: ";
+    for (size_t i = 0; i < ni.size(); ++i)
+      printf("%4.2f ", ni[i]);
+
     std::vector<float> no = layers[layers.size() - 1].getOutputs().toVector();
-    std::cout << "| NN: ";
+    std::cout << "| NNO: ";
     for (size_t i = 0; i < no.size(); ++i)
-      printf("%5.3f ", no[i]);
+      printf("%4.2f ", no[i]);
+
     std::vector<float> to = target.toVector();
     std::cout << "| T: ";
     for (size_t i = 0; i < to.size(); ++i)
-      printf("%5.3f ", to[i]);
+      printf("%4.2f ", to[i]);
+
     Matrix mse =
         mm.loss(layers[layers.size() - 1].getOutputs(), target, Loss::MSE);
+
     std::vector<float> lo = mse.toVector();
     std::cout << "| L: ";
     for (size_t i = 0; i < lo.size(); ++i)
@@ -139,25 +147,48 @@ public:
 
     Matrix dAnl =
         mm.d_loss(layers[layers.size() - 1].getOutputs(), target, Loss::MSE);
+
     for (int i = layers.size() - 1; i >= 0; --i) {
+      printf("=== Layer %d ===\n", i + 1);
+      printf("dAnl: ");
+      dAnl.print();
+
       Matrix dZl = mm.mult(dAnl, mm.d_activate(layers[i].getInternal()));
-      Matrix dWl = mm.mult(
+      printf("dZl: ");
-          mm.dot(dZl, i == 0 ? inputs : layers[i - 1].getOutputs(), true),
+      dZl.print();
-          1.0f / (float)inputs.getRows());
+
+      Matrix dWl =
+          mm.mult(mm.dot(dZl, i == 0 ? inputs : layers[i - 1].getOutputs(),
+                         false, true),
+                  1.0f / (float)inputs.getRows());
+      printf("dWl: ");
+      dWl.print();
+
       Vector dbl = mm.axis_sum(mm.mult(dZl, 1.0f / (float)inputs.getRows()));
-      dAnl = mm.dot(dZl, layers[i].getWeights(), false, false); // false true?!
+      printf("dbl: ");
+      dbl.print();
+
+      dAnl = mm.dot(layers[i].getWeights(), dZl, true); // false true?!
 
       mm.await();
 
       layers[i].setWeights(mm.add(layers[i].getWeights(), dWl, -speed));
+      printf("Weights %d: ", i + 1);
+      layers[i].getWeights().print();
+
       layers[i].setBias(
           vm.add(layers[i].getBias(), dbl, -speed / (float)inputs.getRows()));
+      printf("Bias %d: ", i + 1);
+      layers[i].getBias().print();
     }
 
     return mse;
   }
 
   const LearnLayer &getLayer(int i) const { return layers[i]; }
+
+  // delete
+  LearnLayer &getLayer(int i) { return layers[i]; }
 };
 
 #ifndef NOGPU
@@ -165,41 +196,80 @@ OpenCL openCL;
 #endif
 
 int main() {
+  // LearnNerualNetrowk nn(
+  //     3, {Layer(3, Activation::SIGMOID), Layer(3, Activation::SIGMOID)});
+  //
+  // Matrix weights1(3, 3,
+  //                 {0.88f, 0.39f, 0.9f, 0.37f, 0.14f, 0.41f, 0.96f, 0.5f,
+  //                 0.6f});
+  // Matrix weights2(
+  //     3, 3, {0.29f, 0.57f, 0.36f, 0.73f, 0.53f, 0.68f, 0.01f, 0.02f, 0.58f});
+  //
+  // Vector bias1(std::vector<float>{0.23f, 0.89f, 0.08f});
+  // Vector bias2(std::vector<float>{0.78f, 0.83f, 0.8f});
+  //
+  // nn.getLayer(0).setWeights(weights1);
+  // nn.getLayer(0).setBias(bias1);
+  //
+  // nn.getLayer(1).setWeights(weights2);
+  // nn.getLayer(1).setBias(bias2);
+  //
+  // std::cout << std::endl;
+  //
+  // Matrix input(3, 1, {0.03f, 0.72f, 0.49f});
+  // Matrix target(3, 1, {0.93f, 0.74f, 0.17f});
+  //
+  // // for (int i = 0; i < 1000; i++)
+  // nn.learn(input, target, 0.01f);
+
   LearnNerualNetrowk nn(
-      2, {Layer(2, Activation::TANH), Layer(1, Activation::SIGMOID)});
+      2, {Layer(3, Activation::SIGMOID), Layer(1, Activation::SIGMOID)});
-  std::cout << std::endl;
 
-  // Matrix input(4, 2);
+  Matrix input(2, 4);
-  // Matrix target(4, 1);
+  Matrix target(1, 4);
-  //
-  // for (int batch = 0; batch < 4; batch++) {
-  //   for (int i = 0; i < 4; i++) {
-  //     int v1 = (i / 2) % 2;
-  //     int v2 = i % 2;
-  //
-  //     input(i, 0) = static_cast<float>(v1);
-  //     input(i, 1) = static_cast<float>(v2);
-  //     target(i, 0) = static_cast<float>(v1 ^ v2);
-  //   }
-  // }
-  //
-  // for (int i = 0; i < 10; i++) {
-  //   printf("%4d | ", i + 1);
-  //   Matrix mse = nn.learn(input, target, 0.1f * std::pow(0.99, i));
-  // }
 
-  for (int i = 0; i < 4 * 1000; i++) {
+  float min = 100.0f;
-    int v1 = (i / 2) % 2;
+  for (int batch = 0; batch < 4; batch++) {
-    int v2 = i % 2;
+    for (int i = 0; i < 4; i++) {
+      int v1 = (i / 2) % 2;
+      int v2 = i % 2;
 
-    Matrix input(1, 2, {static_cast<float>(v1), static_cast<float>(v2)});
+      input(0, i) = static_cast<float>(v1);
-    Matrix target(1, 1, static_cast<float>(v1 ^ v2));
+      input(1, i) = static_cast<float>(v2);
-
+      target(0, i) = static_cast<float>(v1 ^ v2);
-    printf("%5d | ", i + 1);
+    }
-    Matrix mse = nn.learn(input, target, 0.00003f);
-    if (i % 4 == 3)
-      std::cout << std::endl;
   }
 
+  for (int i = 0; i < 1000; i++) {
+    printf("%4d | ", i + 1);
+    Matrix mse = nn.learn(input, target, 0.0001f * std::pow(0.99f, i));
+    std::vector<float> lv = mse.toVector();
+    float loss = 0.0f;
+    for (size_t i = 0; i < lv.size(); ++i)
+      loss += lv[i];
+    if (loss < min)
+      min = loss;
+  }
+  std::cout << min << std::endl;
+
+  // LearnNerualNetrowk nn(
+  //     2, {Layer(3, Activation::SIGMOID), Layer(1, Activation::SIGMOID)});
+  // float min = 100.0f;
+  // for (int i = 0; i < 4 * 10000; i++) {
+  //   int v1 = (i / 2) % 2;
+  //   int v2 = i % 2;
+  //
+  //   Matrix input(2, 1, {static_cast<float>(v1), static_cast<float>(v2)});
+  //   Matrix target(1, 1, static_cast<float>(v1 ^ v2));
+  //
+  //   printf("%5d | ", i + 1);
+  //   Matrix mse = nn.learn(input, target, 0.0001f * std::pow(0.95f, i));
+  //   if (i % 4 == 3)
+  //     std::cout << std::endl;
+  //   if (mse[0] < min)
+  //     min = mse[0];
+  // }
+  // std::cout << min << std::endl;
+
   return 0;
 }

src/math/tensor/cpu/math.hpp

@@ -116,13 +116,13 @@ private:
   Tensor2 mse(const Tensor2 &a, const Tensor2 &b) {
     Tensor2 result(a.getShape(), false);
     for (size_t i = 0; i < result.getSize(); ++i)
-      result[i] += (a[i] - b[i]) * (a[i] - b[i]) / (float)a.getCols();
+      result[i] = (a[i] - b[i]) * (a[i] - b[i]) / (float)a.getCols();
     return result;
   }
-  Tensor2 dmse(const Tensor2 &a, const Tensor2 &b) {
+  Tensor2 d_mse(const Tensor2 &a, const Tensor2 &b) {
     Tensor2 result(a.getShape(), false);
     for (size_t i = 0; i < result.getSize(); ++i)
-      result[i] += 2 * (a[i] - b[i]) / (float)a.getCols();
+      result[i] = 2 * (a[i] - b[i]) / (float)a.getCols();
     return result;
   }
 
@@ -133,7 +133,7 @@ public:
               float alpha = 0.01f) override {
     validateMultDimensions(a, b, transpose_a, transpose_b);
     if (bias != nullptr)
-      validateBiasDimensions(b, *bias, transpose_b);
+      validateBiasDimensions(a, *bias, transpose_a);
     Tensor2 result(transpose_a ? a.getCols() : a.getRows(),
                    transpose_b ? b.getRows() : b.getCols(), 0.0f);
     for (int i = 0; i < result.getRows(); ++i) {
@@ -143,7 +143,7 @@ public:
           sum += (transpose_a ? a(k, i) : a(i, k)) *
                  (transpose_b ? b(j, k) : b(k, j));
         result(i, j) =
-            activateX(sum + (bias == nullptr ? 0.0f : (*bias)(j)), type, alpha);
+            activateX(sum + (bias == nullptr ? 0.0f : (*bias)(i)), type, alpha);
       }
     }
     return result;
@@ -162,18 +162,18 @@ public:
     this->validateSameDimensions(a, b);
     switch (type) {
     case Loss::MSE:
-      return dmse(a, b);
+      return d_mse(a, b);
     default:
       throw std::invalid_argument("Unknown loss type");
     }
   }
 
   Tensor1 axis_sum(const Tensor2 &m) override {
-    Tensor1 result(m.getCols(), 0.0f);
+    Tensor1 result(m.getRows(), 0.0f);
-    for (int i = 0; i < m.getCols(); ++i) {
+    for (int i = 0; i < m.getRows(); ++i) {
       float sum = 0.0f;
-      for (int j = 0; j < m.getRows(); ++j)
+      for (int j = 0; j < m.getCols(); ++j)
-        sum += m(j, i);
+        sum += m(i, j);
       result(i) = sum;
     }
     return result;

src/math/tensor/math.hpp

@@ -64,9 +64,9 @@ public:
       throw std::invalid_argument(
           "Invalid matrix dimensions for multiplication");
   };
-  void validateBiasDimensions(const M &a, const V &b, bool transpose) const {
+  void validateBiasDimensions(const M &m, const V &v, bool transpose) const {
-    if ((!transpose && a.getCols() != b.getSize()) ||
+    if ((transpose && (size_t)m.getCols() != v.getSize()) ||
-        (transpose && a.getRows() != b.getSize()))
+        (!transpose && (size_t)m.getRows() != v.getSize()))
       throw std::invalid_argument("Invalid matrix bias");
   };
 };