diff --git a/unit_tests/metrics/Test_Accuracy.cpp b/unit_tests/metrics/Test_Accuracy.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..8d849feede215fbadfc684c8da68c1612879b995
--- /dev/null
+++ b/unit_tests/metrics/Test_Accuracy.cpp
@@ -0,0 +1,178 @@
+/********************************************************************************
+ * 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/backend/cpu/operator/ArgMaxImpl.hpp"
+#include "aidge/backend/cpu/operator/AndImpl.hpp"
+#include "aidge/backend/cpu/operator/ReduceSumImpl.hpp"
+#include "aidge/learning/metrics/Accuracy.hpp"
+#include "aidge/data/Tensor.hpp"
+#include "aidge/utils/TensorUtils.hpp"
+
+#if USE_AIDGE_BACKEND_CUDA
+#include "aidge/backend/cuda/operator/ArgMaxImpl.hpp"
+#include "aidge/backend/cuda/operator/AndImpl.hpp"
+#include "aidge/backend/cuda/operator/ReduceSumImpl.hpp"
+#endif
+
+namespace Aidge {
+TEST_CASE("[metrics] Accuracy", "[metrics][Accuracy]") {
+
+ 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);
+
+ SECTION("CPU") {
+ for (std::uint16_t trial = 0; trial < NBTRIALS; ++trial) {
+ const std::size_t nb_dims = 2; // For Accuracy test, nb_dims is fixed as 2: NbBatch, NbClass
+ std::vector<std::size_t> dims(2);
+ std::int32_t classAxis = 1;
+
+ for (std::size_t i = 0; i < nb_dims; ++i) { dims[i] = dimsDist(gen); }
+
+ std::size_t numClasses = dims[1];
+ std::size_t numBatchess = dims[0];
+ 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);
+ }
+
+ // Calculate accuracy
+ int correct_predictions = 0;
+ for (std::size_t batch = 0; batch < dims[0]; ++batch) {
+ // Find the index of the maximum value in the current batch (for both pred and targ)
+ auto pred_start = pred.get() + batch * numClasses;
+ auto targ_start = targ.get() + batch * numClasses;
+
+ std::size_t pred_max_idx = std::distance(pred_start, std::max_element(pred_start, pred_start + numClasses));
+ std::size_t targ_max_idx = std::distance(targ_start, std::max_element(targ_start, targ_start + numClasses));
+
+ // If the indices match, it's a correct prediction
+ if (pred_max_idx == targ_max_idx) {
+ correct_predictions++;
+ }
+ }
+
+ // Calculate accuracy as the proportion of correct predictions
+ float accuracy = static_cast<float>(correct_predictions);
+
+ // compute the Accuracy using Aidge::metric::Accuracy function
+ std::shared_ptr<Tensor> pred_tensor = std::make_shared<Tensor>(dims);
+ pred_tensor->setBackend("cpu");
+ pred_tensor->getImpl()->setRawPtr(pred.get(), nb_elements);
+
+ std::shared_ptr<Tensor> targ_tensor = std::make_shared<Tensor>(dims);
+ targ_tensor->setBackend("cpu");
+ targ_tensor->getImpl()->setRawPtr(targ.get(), nb_elements);
+
+ const Tensor res_function = metrics::Accuracy(pred_tensor, targ_tensor, classAxis);
+
+ // compare results
+ Tensor res_manual_tensor = Tensor(accuracy);
+ REQUIRE(approxEq<float>(res_manual_tensor, res_function));
+ }
+ }
+#if USE_AIDGE_BACKEND_CUDA
+ SECTION("CUDA") {
+ for (std::uint16_t trial = 0; trial < NBTRIALS; ++trial) {
+ const std::size_t nb_dims = 2; // For Accuracy test, nb_dims is fixed as 2: NbBatch, NbClass
+ std::vector<std::size_t> dims(2);
+ std::int32_t classAxis = 1;
+
+ for (std::size_t i = 0; i < nb_dims; ++i) { dims[i] = dimsDist(gen); }
+
+ std::size_t numClasses = dims[1];
+ std::size_t numBatchess = dims[0];
+ 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);
+ }
+ float * d_pred;
+ cudaMalloc(&d_pred, nb_elements * sizeof(float));
+ cudaMemcpy(d_pred, pred.get(), nb_elements * sizeof(float), cudaMemcpyHostToDevice);
+
+ // 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);
+ }
+ float * d_targ;
+ cudaMalloc(&d_targ, nb_elements * sizeof(float));
+ cudaMemcpy(d_targ, targ.get(), nb_elements * sizeof(float), cudaMemcpyHostToDevice);
+
+
+ // Calculate accuracy
+ int correct_predictions = 0;
+ for (std::size_t batch = 0; batch < dims[0]; ++batch) {
+ // Find the index of the maximum value in the current batch (for both pred and targ)
+ auto pred_start = pred.get() + batch * numClasses;
+ auto targ_start = targ.get() + batch * numClasses;
+
+ std::size_t pred_max_idx = std::distance(pred_start, std::max_element(pred_start, pred_start + numClasses));
+ std::size_t targ_max_idx = std::distance(targ_start, std::max_element(targ_start, targ_start + numClasses));
+
+ // If the indices match, it's a correct prediction
+ if (pred_max_idx == targ_max_idx) {
+ correct_predictions++;
+ }
+ }
+
+ // Calculate accuracy as the proportion of correct predictions
+ float accuracy = static_cast<float>(correct_predictions);
+
+ // compute the MSE using Aidge::loss::MSE function
+ std::shared_ptr<Tensor> pred_tensor = std::make_shared<Tensor>(dims);
+ pred_tensor->setBackend("cuda");
+ pred_tensor->getImpl()->setRawPtr(d_pred, nb_elements);
+
+
+ std::shared_ptr<Tensor> targ_tensor = std::make_shared<Tensor>(dims);
+ targ_tensor->setBackend("cuda");
+ targ_tensor->getImpl()->setRawPtr(d_targ, nb_elements);
+
+ const Tensor res_function = metrics::Accuracy(pred_tensor, targ_tensor, classAxis);
+
+ // compare results
+ Tensor res_manual_tensor = Tensor(accuracy);
+ REQUIRE(approxEq<float>(res_manual_tensor, res_function));
+
+ cudaFree(d_pred);
+ cudaFree(d_targ);
+ }
+ }
+#endif
+}
+} // namespace Aidge