From eff8202099609940a906fe752dc50596bebd5c12 Mon Sep 17 00:00:00 2001
From: Jerome Hue <jerome.hue@cea.fr>
Date: Mon, 3 Mar 2025 16:44:35 +0100
Subject: [PATCH] Add randomness in heaviside backward test
---
src/operator/HeavisideImpl.cpp | 1 -
unit_tests/operator/Test_HeavisideImpl.cpp | 57 +++++++++++++++++-----
2 files changed, 45 insertions(+), 13 deletions(-)
diff --git a/src/operator/HeavisideImpl.cpp b/src/operator/HeavisideImpl.cpp
index 5bf77f87..8349a0ad 100644
--- a/src/operator/HeavisideImpl.cpp
+++ b/src/operator/HeavisideImpl.cpp
@@ -48,7 +48,6 @@ void Aidge::HeavisideImplCpu::backward() {
std::shared_ptr<Tensor> in0 = op_.getInput(0);
std::shared_ptr<Tensor> out0 = op_.getOutput(0);
- AIDGE_ASSERT(out0, "missing output #0 for current {} operator", op_.type());
impl.backward(gra_int0->size(), getCPUPtr(in0), getCPUPtr(gra_out0), getCPUPtr(gra_int0));
}
diff --git a/unit_tests/operator/Test_HeavisideImpl.cpp b/unit_tests/operator/Test_HeavisideImpl.cpp
index e6aa38b8..2743027c 100644
--- a/unit_tests/operator/Test_HeavisideImpl.cpp
+++ b/unit_tests/operator/Test_HeavisideImpl.cpp
@@ -100,27 +100,60 @@ TEST_CASE("[cpu/operator] Heaviside(forward)", "[Heaviside][CPU]") {
}
}
+// TODO: Make this work for random intput.
TEST_CASE("[cpu/operator] Heaviside(backward)", "[Heaviside][CPU]") {
+
+ std::random_device rd;
+ std::mt19937 gen(rd());
+ std::uniform_real_distribution<float> valueDist(-2.0f, 2.0f);
+ std::uniform_int_distribution<std::size_t> sizeDist(5, 100);
+
+ const std::size_t tensorSize = sizeDist(gen);
+
auto hs = Heaviside(1.0f);
auto op = std::static_pointer_cast<OperatorTensor>(hs->getOperator());
op->setDataType(DataType::Float32);
op->setBackend("cpu");
- auto input = Tensor(Array1D<float, 3>({1.0, -1.0, 1.0}));
- input.setDataType(DataType::Float32);
- input.setBackend("cpu");
- auto grad = Tensor(Array1D<float, 3>({1.0, 1.0, 1.0}));
- grad.setDataType(DataType::Float32);
- grad.setBackend("cpu");
+ auto inputTensor = std::make_shared<Tensor>(std::vector<std::size_t>{tensorSize});
+ inputTensor->setDataType(DataType::Float32);
+ inputTensor->setBackend("cpu");
+ auto* inputData = static_cast<float*>(inputTensor->getImpl()->rawPtr());
+
+ for(std::size_t i = 0; i < tensorSize; ++i) {
+ inputData[i] = valueDist(gen);
+ }
+
+ auto gradTensor = std::make_shared<Tensor>(std::vector<std::size_t>{tensorSize});
+ gradTensor->setDataType(DataType::Float32);
+ gradTensor->setBackend("cpu");
+ auto* gradData = static_cast<float*>(gradTensor->getImpl()->rawPtr());
+
+ for (std::size_t i = 0; i < tensorSize; ++i) {
+ gradData[i] = valueDist(gen);
+ }
- op->setInput(IOIndex_t(0), std::make_shared<Tensor>(input));
+ op->setInput(IOIndex_t(0), inputTensor);
op->forward();
-
- op->getOutput(0)->setGrad(std::make_shared<Tensor>(grad));
+
+ auto output = op->getOutput(0);
+ output->setGrad(gradTensor);
+
+ // Backward pass
op->backward();
-
- auto expectedResult = Tensor(Array1D<float,3>({0.0920, 0.0920, 0.0920}));
- REQUIRE(approxEq<float>(*(op->getInput(0)->grad()), expectedResult));
+
+ // Compute expected gradient manually
+ auto expectedGrad = std::make_shared<Tensor>(std::vector<std::size_t>{tensorSize});
+ expectedGrad->setDataType(DataType::Float32);
+ expectedGrad->setBackend("cpu");
+ auto* expectedGradData = static_cast<float*>(expectedGrad->getImpl()->rawPtr());
+
+ for (std::size_t i = 0; i < tensorSize; ++i) {
+ expectedGradData[i] = gradData[i] * (1.0f / (1.0f + (inputData[i] * M_PI) * (inputData[i] * M_PI)));
+ }
+
+ // Compare actual gradient with expected gradient
+ REQUIRE(approxEq<float>(*(op->getInput(0)->grad()), *expectedGrad));
}
}
--
GitLab