cudnn批处理中取出训练成绩

平时batch=1时，我们都会：（cpu训练一次更新一次或测试时情形）

// Perform classification
std::vector<float> class_vec(10);

// Copy back result
error_handling(cudaMemcpy(&class_vec[0], d_fc2smax, sizeof(float) * 10, cudaMemcpyDeviceToHost));

// Determine classification according to maximal response
int chosen = 0;
for (int id = 1; id < 10; ++id)
{
if (class_vec[chosen] < class_vec[id]) chosen = id;
}

// if (chosen != test_labels[i])// (float) Datasets.labels[i]
if (chosen != DatasetsTest.labels[i])
++num_errors;

batch很大时：特别是用cudnn训练，我们希望跑一轮，看一次成绩：

std::vector<float> class_vec(10 * m_batchSize);

error_handling(cudaMemcpy(&class_vec[0], d_fc2smax, sizeof(float) * 10 * m_batchSize, cudaMemcpyDeviceToHost));

for (int b = 0; b < m_batchSize; b++)
{
int chosen = 0;
for (int id = 1; id < 10; ++id)
{
if (class_vec[chosen + b * 10] < class_vec[id + b * 10]) chosen = id;
}

if (chosen != Datasets.labels[iter* m_batchSize + b])
++num_errorsTRain;
}

这都是小经验，做个记录