mirror of
https://github.com/lordmathis/CUDANet.git
synced 2025-11-05 17:34:21 +00:00
Add more softmax tests
This commit is contained in:
@@ -5,8 +5,7 @@
|
||||
|
||||
#include "activation_functions.cuh"
|
||||
|
||||
TEST(ActivationsTest, SigmoidSanityCheck) {
|
||||
|
||||
TEST(ActivationFunctionsTest, SigmoidSanityCheck) {
|
||||
cudaError_t cudaStatus;
|
||||
|
||||
float input[3] = {-100.0f, 0.0f, 100.0f};
|
||||
@@ -22,7 +21,8 @@ TEST(ActivationsTest, SigmoidSanityCheck) {
|
||||
cudaStatus = cudaMalloc((void**)&d_output, sizeof(float) * 3);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
cudaStatus = cudaMemcpy(d_input, input, sizeof(float) * 3, cudaMemcpyHostToDevice);
|
||||
cudaStatus =
|
||||
cudaMemcpy(d_input, input, sizeof(float) * 3, cudaMemcpyHostToDevice);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
CUDANet::Kernels::sigmoid<<<1, 3>>>(d_input, d_output, 3);
|
||||
@@ -31,7 +31,9 @@ TEST(ActivationsTest, SigmoidSanityCheck) {
|
||||
|
||||
std::vector<float> output(3);
|
||||
|
||||
cudaStatus = cudaMemcpy(output.data(), d_output, sizeof(float) * 3, cudaMemcpyDeviceToHost);
|
||||
cudaStatus = cudaMemcpy(
|
||||
output.data(), d_output, sizeof(float) * 3, cudaMemcpyDeviceToHost
|
||||
);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
for (int i = 0; i < 3; i++) {
|
||||
@@ -40,4 +42,59 @@ TEST(ActivationsTest, SigmoidSanityCheck) {
|
||||
|
||||
cudaFree(d_input);
|
||||
cudaFree(d_output);
|
||||
}
|
||||
|
||||
TEST(ActivationFunctionsTest, SoftmaxExpTest) {
|
||||
cudaError_t cudaStatus;
|
||||
|
||||
float input[6] = {22.496f, 36.9006f, 30.9904f,
|
||||
28.4213f, 26.4541f, 31.7887f};
|
||||
|
||||
std::vector<float> expected = {5886928896.0f, 1.06102872080384e+16f,
|
||||
28771323215872.0f, 2204012904448.0f,
|
||||
308226162688.0f, 63922983927808.0f};
|
||||
|
||||
float* d_input;
|
||||
float* d_output;
|
||||
|
||||
cudaStatus = cudaMalloc((void**)&d_input, sizeof(float) * 6);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
cudaStatus = cudaMalloc((void**)&d_output, sizeof(float) * 6);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
cudaStatus =
|
||||
cudaMemcpy(d_input, input, sizeof(float) * 6, cudaMemcpyHostToDevice);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
CUDANet::Kernels::softmax_exp<<<1, 6>>>(d_input, d_output, 6);
|
||||
cudaStatus = cudaDeviceSynchronize();
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
std::vector<float> output(6);
|
||||
|
||||
cudaStatus = cudaMemcpy(
|
||||
output.data(), d_output, sizeof(float) * 6, cudaMemcpyDeviceToHost
|
||||
);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
|
||||
for (int i = 0; i < 6; i++) {
|
||||
EXPECT_NEAR(expected[i], output[i], 1e7);
|
||||
}
|
||||
|
||||
cudaFree(d_input);
|
||||
cudaFree(d_output);
|
||||
}
|
||||
|
||||
TEST(ActivationFunctionsTest, SoftmaxSumTest) {
|
||||
cudaError_t cudaStatus;
|
||||
|
||||
std::vector<float> input = {5886928896.0f, 1.06102872080384e+16f,
|
||||
28771323215872.0f, 2204012904448.0f,
|
||||
308226162688.0f, 63922983927808.0f};
|
||||
|
||||
float* d_input;
|
||||
|
||||
cudaStatus = cudaMalloc((void**)&d_input, sizeof(float) * 6);
|
||||
EXPECT_EQ(cudaStatus, cudaSuccess);
|
||||
}
|
||||
@@ -3,7 +3,7 @@
|
||||
#include <cuda_runtime.h>
|
||||
#include <vector>
|
||||
|
||||
TEST(ActivationTest, SoftmaxTest) {
|
||||
TEST(ActivationTest, SoftmaxTest1) {
|
||||
CUDANet::Layers::Activation activation(
|
||||
CUDANet::Layers::ActivationType::SOFTMAX, 5
|
||||
);
|
||||
@@ -30,5 +30,35 @@ TEST(ActivationTest, SoftmaxTest) {
|
||||
|
||||
EXPECT_NEAR(sum, 1.0f, 1e-5f);
|
||||
|
||||
cudaFree(d_input);
|
||||
}
|
||||
|
||||
TEST(ActivationTest, SoftmaxTest2) {
|
||||
CUDANet::Layers::Activation activation(
|
||||
CUDANet::Layers::ActivationType::SOFTMAX, 6
|
||||
);
|
||||
|
||||
std::vector<float> input = {22.496f, 36.9006f, 30.9904f, 28.4213f, 26.4541f, 31.7887f};
|
||||
|
||||
float* d_input;
|
||||
cudaMalloc((void**)&d_input, sizeof(float) * 6);
|
||||
cudaMemcpy(d_input, input.data(), sizeof(float) * 6, cudaMemcpyHostToDevice);
|
||||
|
||||
activation.activate(d_input);
|
||||
std::vector<float> output(6);
|
||||
cudaMemcpy(
|
||||
output.data(), d_input, sizeof(float) * 6, cudaMemcpyDeviceToHost
|
||||
);
|
||||
|
||||
float sum = 0.0f;
|
||||
|
||||
std::vector<float> expected = {0.0f, 0.99111f, 0.00269f, 0.00021f, 3e-05f, 0.00597f};
|
||||
for (int i = 0; i < 5; ++i) {
|
||||
sum += output[i];
|
||||
EXPECT_NEAR(output[i], expected[i], 1e-5f);
|
||||
}
|
||||
|
||||
EXPECT_NEAR(sum, 1.0f, 1e-5f);
|
||||
|
||||
cudaFree(d_input);
|
||||
}
|
||||
Reference in New Issue
Block a user