ANSCENTER/ANSCORE
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Use yuv and fix FR to use batch 4

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-22 17:11:17 +10:00
parent a0d5c81814
commit 6135649835
1 changed files with 17 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
modules/ANSFR/ANSFaceRecognizer.cpp
View File

@@ -255,8 +255,23 @@ namespace ANSCENTER {
if (engineType == EngineType::NVIDIA_GPU) { if (engineType == EngineType::NVIDIA_GPU) {
optimizedModelFolder = GetParentFolder(_modelFilePath); optimizedModelFolder = GetParentFolder(_modelFilePath);
m_options.optBatchSize = 8; // Must match the runtime Initialize() config EXACTLY — otherwise the
m_options.maxBatchSize = 32; // engine built/looked-up here has a different profile than the one
// Initialize() expects, causing either a stale `.b32` rebuild every
// launch (wasting 30-60s) or a load-time profile mismatch.
// Keep in sync with lines 100-114.
m_options.optBatchSize = 4;
m_options.maxBatchSize = 4;
m_options.maxInputHeight = GPU_FACE_HEIGHT;
m_options.minInputHeight = GPU_FACE_HEIGHT;
m_options.optInputHeight = GPU_FACE_HEIGHT;
m_options.maxInputWidth = GPU_FACE_WIDTH;
m_options.minInputWidth = GPU_FACE_WIDTH;
m_options.optInputWidth = GPU_FACE_WIDTH;
m_options.calibrationBatchSize = 8;
m_options.engineFileDir = optimizedModelFolder; m_options.engineFileDir = optimizedModelFolder;
m_options.precision = fp16 ? Precision::FP16 : Precision::FP32; m_options.precision = fp16 ? Precision::FP16 : Precision::FP32;