Merge branch 'BCE_loss' into 'dev'

Add BCE loss function

See merge request eclipse/aidge/aidge_learning!8

include/aidge/loss/LossList.hpp

@@ -31,6 +31,8 @@ namespace loss {
  */
 Tensor MSE(std::shared_ptr<Tensor>& prediction,
            const std::shared_ptr<Tensor>& target);
+Tensor BCE(std::shared_ptr<Tensor>& prediction,
+           const std::shared_ptr<Tensor>& target);
 
 }  // namespace loss
 }  // namespace Aidge

python_binding/learning/loss/pybind_Loss.cpp

@@ -23,5 +23,6 @@ void init_Loss(py::module &m) {
     auto m_loss =
         m.def_submodule("loss", "Submodule dedicated to loss functions");
     m_loss.def("MSE", &loss::MSE, py::arg("graph"), py::arg("target"));
+    m_loss.def("BCE", &loss::BCE, py::arg("graph"), py::arg("target"));
 }
 }  // namespace Aidge

src/loss/classification/BCE.cpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#include <memory>
+#include <numeric>  // std::iota
+
+#include "aidge/data/Tensor.hpp"
+#include "aidge/graph/GraphView.hpp"
+#include "aidge/graph/OpArgs.hpp"
+#include "aidge/loss/LossList.hpp"
+#include "aidge/recipes/GraphViewHelper.hpp"
+#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
+
+#include "aidge/operator/OperatorTensor.hpp"
+#include "aidge/operator/Add.hpp"
+#include "aidge/operator/Sub.hpp"
+#include "aidge/operator/Mul.hpp"
+#include "aidge/operator/Div.hpp"
+#include "aidge/operator/Ln.hpp"
+#include "aidge/operator/ReduceMean.hpp"
+
+#include "aidge/backend/cpu/operator/AddImpl.hpp"
+#include "aidge/backend/cpu/operator/SubImpl.hpp"
+#include "aidge/backend/cpu/operator/MulImpl.hpp"
+#include "aidge/backend/cpu/operator/DivImpl.hpp"
+#include "aidge/backend/cpu/operator/LnImpl.hpp"
+#include "aidge/backend/cpu/operator/ReduceMeanImpl.hpp"
+
+
+Aidge::Tensor Aidge::loss::BCE(std::shared_ptr<Tensor>& prediction,
+                               const std::shared_ptr<Tensor>& target) {
+    /*
+	Binay Cross Entropy (BCE) loss function
+	
+    Implementation note:
+    loss function is computed using a graph in order to not be backend dependant.
+    */
+
+    AIDGE_ASSERT(target->dims().size() == 2,
+                 "Label must have two dims: [BatchSize, NbChannel]");
+    AIDGE_ASSERT(prediction->backend() == target->backend(),
+                 "'prediction' and 'target' Tensors must be on the "
+                 "same backend. Found {} and {}.\n",
+                 prediction->backend(), target->backend());
+    AIDGE_ASSERT(prediction->dims() == target->dims(),
+                 "'prediction' (shape {}) and 'target' (shape {}) Tensors must "
+                 "have the same dimensions.\n",
+                 prediction->dims(), target->dims());
+    AIDGE_ASSERT(prediction->dataType() == target->dataType(),
+                 "'prediction' (data type {}) and 'target' (data type {}) "
+                 "Tensors must have the same data type.\n",
+                 prediction->dataType(), target->dataType());
+
+    const float eps1 = 1.e-10f;
+    const float eps2 = 1.e-10f;
+
+    // Define nodes: inputs
+    const std::shared_ptr<Node> prediction_node = Producer(prediction, "pred");
+    const std::shared_ptr<Node> target_node = Producer(target, "label");
+
+    // Define nodes: add1 = prediction + eps1, add2 = target + eps1
+    const std::shared_ptr<Node> add1_node = Add(2, "add1");
+    const std::shared_ptr<Node> add2_node = Add(2, "add2");
+    prediction_node->addChild(add1_node, 0, 0);
+    Producer(std::make_shared<Tensor>(Array1D<float, 1>{{eps1}}))
+        ->addChild(add1_node, 0, 1);
+    target_node->addChild(add2_node, 0, 0);
+    Producer(std::make_shared<Tensor>(Array1D<float, 1>{{eps1}}))
+        ->addChild(add2_node, 0, 1);
+	
+    // Define nodes: sub1 = 1 - prediction + eps2 and sub2 = - (1 - target + eps2)
+    const std::shared_ptr<Node> sub1_node = Sub("sub1");
+    const std::shared_ptr<Node> sub2_node = Sub("sub2");
+    Producer(std::make_shared<Tensor>(Array1D<float, 1>{{1.0f + eps2}}))
+        ->addChild(sub1_node, 0, 0);
+    prediction_node->addChild(sub1_node, 0, 1);
+    target_node->addChild(sub2_node, 0, 0);
+    Producer(std::make_shared<Tensor>(Array1D<float, 1>{{1.0f + eps2}}))
+        ->addChild(sub2_node, 0, 1);
+	
+    // Define nodes: ln1 = ln(prediction + eps1) and ln2 = ln(1 - prediction + eps2)
+    const std::shared_ptr<Node> ln1_node = Ln("ln1");
+    const std::shared_ptr<Node> ln2_node = Ln("ln2");
+    add1_node-> addChild(ln1_node, 0, 0);
+    sub1_node-> addChild(ln2_node, 0, 0);
+
+    // Define nodes: mul1 = (target + eps1) * ln(prediction + eps1) and mul2 = - (1 - target + eps2) * ln(1 - prediction + eps2)
+    const std::shared_ptr<Node> mul1_node = Mul("mul1");
+    const std::shared_ptr<Node> mul2_node = Mul("mul2");
+    add2_node->addChild(mul1_node, 0, 0);
+    ln1_node->addChild(mul1_node, 0, 1);
+    sub2_node->addChild(mul2_node, 0, 0);
+    ln2_node->addChild(mul2_node, 0, 1);
+	
+    // Define node: sub3 = - [(target + eps1) * ln(prediction + eps1) + (1 - target + eps2) * ln(1 - prediction + eps2)]
+    const std::shared_ptr<Node> sub3_node = Sub("sub3");
+    mul2_node->addChild(sub3_node, 0, 0);
+    mul1_node->addChild(sub3_node, 0, 1);
+	
+    // Define nodes: div1 = (target + eps1) / (prediction + eps1) and div2 = - (1 - target + eps2)/(1 - prediction + eps2)
+    const std::shared_ptr<Node> div1_node = Div("div1");
+    const std::shared_ptr<Node> div2_node = Div("div2");
+    add2_node->addChild(div1_node, 0, 0);
+    add1_node->addChild(div1_node, 0, 1);
+    sub2_node->addChild(div2_node, 0, 0);
+    sub1_node->addChild(div2_node, 0, 1);
+	
+    // Define node: add3 = (target + eps1) / (prediction + eps1) - (1 - target + eps2)/(1 - prediction + eps2)
+    const std::shared_ptr<Node> add3_node = Add(2, "add3");
+    div1_node->addChild(add3_node, 0, 0);
+    div2_node->addChild(add3_node, 0, 1);
+
+    // Define node: loss    
+    std::vector<int> axes_dims(prediction->nbDims());
+    std::iota(std::begin(axes_dims), std::end(axes_dims), 0);
+    auto loss_node = ReduceMean(axes_dims, 1, "loss");
+    sub3_node->addChild(loss_node, 0, 0);
+		
+    // Define node: gradient
+    const std::shared_ptr<Node> gradient_node = Mul("gradient");
+    add3_node->addChild(gradient_node, 0, 0);
+    Producer(std::make_shared<Tensor>(Array1D<float, 1>{{-1.0f/float(target->dims()[0])}}))
+        ->addChild(gradient_node, 0, 1);
+    	
+    // Create GraphView
+    std::shared_ptr<GraphView> gv_loss = std::make_shared<GraphView>("BCE");
+    gv_loss->add({prediction_node, target_node, 
+                  add1_node->getParent(1), add1_node,
+                  add2_node->getParent(1), add2_node,
+                  sub1_node->getParent(0), sub1_node,
+                  sub2_node->getParent(1), sub2_node,
+                  ln1_node, ln2_node, mul1_node, mul2_node, div1_node, div2_node,
+                  sub3_node, loss_node,
+                  add3_node, gradient_node->getParent(1), gradient_node});
+    gv_loss->compile(prediction->getImpl()->backend(), prediction->dataType());
+	
+    // Compute loss and gradient
+    SequentialScheduler ss_loss{gv_loss};
+    ss_loss.forward(false);
+	
+    prediction->initGrad(); // Enable gradient for output
+    std::shared_ptr<Tensor> outputGrad = prediction->grad(); 
+    const std::shared_ptr<OperatorTensor> gradient_op = std::dynamic_pointer_cast<OperatorTensor>(gradient_node->getOperator());
+    outputGrad->copyFrom(gradient_op->getOutput(0)->clone()); // Update gradient
+	
+    const std::shared_ptr<OperatorTensor> loss_op = std::dynamic_pointer_cast<OperatorTensor>(loss_node->getOperator());
+    return loss_op->getOutput(0)->clone(); // Return loss
+}

src/loss/regression/MSE.cpp

@@ -15,6 +15,7 @@
 #include "aidge/backend/cpu/operator/PowImpl.hpp"
 #include "aidge/backend/cpu/operator/ReduceMeanImpl.hpp"
 #include "aidge/backend/cpu/operator/SubImpl.hpp"
+#include "aidge/backend/cpu/operator/MulImpl.hpp"
 #include "aidge/data/Tensor.hpp"
 #include "aidge/graph/GraphView.hpp"
 #include "aidge/graph/OpArgs.hpp"
@@ -23,6 +24,7 @@
 #include "aidge/operator/Pow.hpp"
 #include "aidge/operator/ReduceMean.hpp"
 #include "aidge/operator/Sub.hpp"
+#include "aidge/operator/Mul.hpp"
 #include "aidge/recipes/GraphViewHelper.hpp"
 #include "aidge/scheduler/Scheduler.hpp"
 #include "aidge/scheduler/SequentialScheduler.hpp"

unit_tests/loss/classification/Test_BCE.cpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#include <catch2/catch_test_macros.hpp>
+#include <cstddef>     // std::size_t
+#include <cmath>       //
+#include <functional>  // std::multiplies, std::plus
+#include <memory>      // std::make_unique
+#include <numeric>     // std::accumulate
+#include <random>      // std::random_device, std::mt19937,
+                       // std::uniform_int_distribution
+#include <vector>
+
+#include "aidge/loss/LossList.hpp"
+#include "aidge/data/Tensor.hpp"
+#include "aidge/utils/TensorUtils.hpp"
+
+namespace Aidge {
+TEST_CASE("[loss/classification] BCE", "[loss][classification][BCE]") {
+    constexpr std::uint16_t NBTRIALS = 10;
+
+    // set random variables
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_int_distribution<std::size_t> dimsDist(1, 5);
+    std::uniform_int_distribution<std::size_t> nbDimsDist(1, 2);
+    std::uniform_real_distribution<float> valueDist(0.0f, 1.0f);
+
+    for (std::uint16_t trial = 0; trial < NBTRIALS; ++trial) {
+        const std::size_t nb_dims = 2; // For BCE test, nb_dims is fixed as 2: NbBatch, NbChan
+        std::vector<std::size_t> dims(2);
+
+        for (std::size_t i = 0; i < nb_dims; ++i) { dims[i] = dimsDist(gen); }
+        const std::size_t nb_elements = std::accumulate(dims.cbegin(), dims.cend(), std::size_t(1), std::multiplies<std::size_t>());
+
+        // create random predictions
+        std::unique_ptr<float[]> pred = std::make_unique<float[]>(nb_elements);
+        for (std::size_t i = 0; i < nb_elements; ++i) {
+            pred[i] = valueDist(gen);
+        }
+
+        // create random targets
+        std::unique_ptr<float[]> targ = std::make_unique<float[]>(nb_elements);
+        for (std::size_t i = 0; i < nb_elements; ++i) {
+            targ[i] = valueDist(gen);
+        }
+
+        // compute the BCE manually
+        const float eps1 = 1.0e-10f;
+        const float eps2 = 1.0e-10f;
+        std::unique_ptr<float[]> tmp_res_manual = std::make_unique<float[]>(nb_elements);
+        for (std::size_t i = 0; i < nb_elements; ++i) {
+            tmp_res_manual[i] = - ((targ[i] + eps1) * std::log(pred[i] + eps1) + (1.0f - targ[i] + eps2) * std::log(1.0f - pred[i] + eps2));
+        }
+        std::cout << "Output manual:" << std::endl;
+        std::shared_ptr<Tensor> tmp_tensor = std::make_shared<Tensor>(dims);
+        tmp_tensor->setBackend("cpu");
+        tmp_tensor->getImpl()->setRawPtr(tmp_res_manual.get(), nb_elements);
+        tmp_tensor->print();
+        const float res_manual = std::accumulate(&tmp_res_manual[0], &tmp_res_manual[nb_elements], 0.0f, std::plus<float>()) / static_cast<float>(nb_elements);
+
+        // compute the BCE using Aidge::loss::BCE function
+        std::cout << "Input 0 manual:" << std::endl;
+        std::shared_ptr<Tensor> pred_tensor = std::make_shared<Tensor>(dims);
+        pred_tensor->setBackend("cpu");
+        pred_tensor->getImpl()->setRawPtr(pred.get(), nb_elements);
+        pred_tensor->print();
+
+        std::cout << "Input 1 manual:" << std::endl;
+        std::shared_ptr<Tensor> targ_tensor = std::make_shared<Tensor>(dims);
+        targ_tensor->setBackend("cpu");
+        targ_tensor->getImpl()->setRawPtr(targ.get(), nb_elements);
+        targ_tensor->print();
+        
+        const Tensor res_function = loss::BCE(pred_tensor, targ_tensor);
+
+        // compare results
+        Tensor res_manual_tensor = Tensor(res_manual);
+        REQUIRE(approxEq<float>(res_manual, res_function));
+    }
+}
+}  // namespace Aidge