import sys

CPP = r'C:\Projects\ANLS\ANSLIB\ANSODEngine\ANSTENSORRTPOSE.cpp'

with open(CPP, 'rb') as f:
    data = f.read()

CRLF = b'\r\n' in data
print(f'Line ending: {"CRLF" if CRLF else "LF"}')

def L(s):
    if CRLF:
        return s.replace(b'\n', b'\r\n')
    return s

errors = []

def replace_once(data, old, new, label):
    count = data.count(old)
    if count != 1:
        errors.append(f'ERROR [{label}]: expected 1, found {count}')
        print(errors[-1])
        return data
    print(f'OK [{label}]')
    return data.replace(old, new, 1)

def replace_all(data, old, new, expected, label):
    count = data.count(old)
    if count != expected:
        errors.append(f'ERROR [{label}]: expected {expected}, found {count}')
        print(errors[-1])
        return data
    print(f'OK [{label}] ({count} replacements)')
    return data.replace(old, new)

def replace_slice(data, start_anchor, end_anchor, new_content, label):
    """Find start_anchor, then end_anchor after it, replace [start..end+len] with new_content."""
    s = data.find(start_anchor)
    if s == -1:
        errors.append(f'ERROR [{label}]: start_anchor not found: {start_anchor[:50]!r}')
        print(errors[-1])
        return data
    e = data.find(end_anchor, s + len(start_anchor))
    if e == -1:
        errors.append(f'ERROR [{label}]: end_anchor not found: {end_anchor[:50]!r}')
        print(errors[-1])
        return data
    e += len(end_anchor)
    old_chunk = data[s:e]
    if data.count(old_chunk) != 1:
        errors.append(f'ERROR [{label}]: extracted chunk not unique (count={data.count(old_chunk)})')
        print(errors[-1])
        return data
    print(f'OK [{label}]')
    return data[:s] + new_content + data[e:]

orig_size = len(data)

# 1. Add #include <future>
data = replace_once(data,
    L(b'#include <opencv2/cudaimgproc.hpp>\n'),
    L(b'#include <opencv2/cudaimgproc.hpp>\n#include <future>\n'),
    'add future')

# 2. Three buildLoadNetwork calls -> add m_maxSlotsPerGpu
data = replace_all(data,
    b'buildLoadNetwork(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE)',
    b'buildLoadNetwork(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE, m_maxSlotsPerGpu)',
    3, 'buildLoadNetwork 3x')

# 3. RunInference(camera_id) - release mutex before DetectObjects
# Old body: single lock scope containing all validation then try/catch
# New body: brief lock scope for validation, then try { return DetectObjects; } catch { ... }
old_ri = L(
b'    std::vector<Object> ANSTENSORRTPOSE::RunInference(const cv::Mat& inputImgBGR,const std::string& camera_id)\n'
b'    {\n'
b'        std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'\n'
b'        // Validation checks\n'
b'        if (!_modelLoadValid) {\n'
b'            _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                "Cannot load the TensorRT model. Please check if it exists",\n'
b'                __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        if (!_licenseValid) {\n'
b'            _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                "Runtime license is not valid or expired. Please contact ANSCENTER",\n'
b'                __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        if (!_isInitialized) {\n'
b'            _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                "Model is not initialized",\n'
b'                __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        if (inputImgBGR.empty() || inputImgBGR.cols < 10 || inputImgBGR.rows < 10) {\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        try {\n'
b'            return DetectObjects(inputImgBGR, camera_id);\n'
b'        }\n'
b'        catch (const std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::RunInference", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n')

new_ri = L(
b'    std::vector<Object> ANSTENSORRTPOSE::RunInference(const cv::Mat& inputImgBGR,const std::string& camera_id)\n'
b'    {\n'
b'        {\n'
b'            std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'\n'
b'            // Validation checks\n'
b'            if (!_modelLoadValid) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                    "Cannot load the TensorRT model. Please check if it exists",\n'
b'                    __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'\n'
b'            if (!_licenseValid) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                    "Runtime license is not valid or expired. Please contact ANSCENTER",\n'
b'                    __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'\n'
b'            if (!_isInitialized) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInference",\n'
b'                    "Model is not initialized",\n'
b'                    __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'\n'
b'            if (inputImgBGR.empty() || inputImgBGR.cols < 10 || inputImgBGR.rows < 10) {\n'
b'                return {};\n'
b'            }\n'
b'        }\n'
b'        try {\n'
b'            return DetectObjects(inputImgBGR, camera_id);\n'
b'        }\n'
b'        catch (const std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::RunInference", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n')
data = replace_once(data, old_ri, new_ri, 'RunInference refactor')

# 4. DetectObjects - three-phase mutex
# Old body: single lock, try { Preprocess -> runInference -> if(succ) PostProcessPose }
# New body: Phase1 lock+Preprocess, Phase2 inference, Phase3 lock+PostProcessPose
old_det = L(
b'    // private\n'
b'    std::vector<Object> ANSTENSORRTPOSE::DetectObjects(const cv::Mat& inputImage, const std::string& camera_id) {\n'
b'        std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'        try {\n'
b'            // Preprocess the input image\n'
b'            const auto input = Preprocess(inputImage);\n'
b'            std::vector<std::vector<std::vector<float>>> featureVectors;\n'
b'            auto succ = m_trtEngine->runInference(input, featureVectors);\n'
b'            if (succ) {\n'
b'                // Check if our model does only object detection or also supports segmentation\n'
b'                std::vector<Object> ret;\n'
b'                const auto& numOutputs = m_trtEngine->getOutputDims().size();\n'
b'                std::vector<float> featureVector;\n'
b'                Engine<float>::transformOutput(featureVectors, featureVector);\n'
b'                const auto& outputDims = m_trtEngine->getOutputDims();\n'
b'                int numChannels = outputDims[outputDims.size() - 1].d[1];\n'
b'                ret = PostProcessPose(featureVector, camera_id);\n'
b'                return ret;\n'
b'            }\n'
b'            else {\n'
b'                this->_logger.LogError("ANSTENSORRTPOSE::DetectObjects", "Error running inference", __FILE__, __LINE__);\n'
b'                std::vector<Object> ret;\n'
b'                ret.clear();\n'
b'                return ret;\n'
b'            }\n'
b'        }\n'
b'        catch (std::exception& e) {\n'
b'            this->_logger.LogFatal("ANSTENSORRTPOSE::DetectObjects", e.what(), __FILE__, __LINE__);\n'
b'            std::vector<Object> ret;\n'
b'            ret.clear();\n'
b'            return ret;\n'
b'        }\n'
b'    }\n')

new_det = L(
b'    // private\n'
b'    std::vector<Object> ANSTENSORRTPOSE::DetectObjects(const cv::Mat& inputImage, const std::string& camera_id) {\n'
b'        // Phase 1: Preprocess under lock\n'
b'        ImageMetadata meta;\n'
b'        std::vector<std::vector<cv::cuda::GpuMat>> input;\n'
b'        {\n'
b'            std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'            input = Preprocess(inputImage, meta);\n'
b'        }\n'
b'        if (input.empty()) return {};\n'
b'\n'
b'        // Phase 2: Inference -- mutex released; pool dispatches to idle GPU slot\n'
b'        std::vector<std::vector<std::vector<float>>> featureVectors;\n'
b'        auto succ = m_trtEngine->runInference(input, featureVectors);\n'
b'        if (!succ) {\n'
b'            this->_logger.LogError("ANSTENSORRTPOSE::DetectObjects", "Error running inference", __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        // Phase 3: Postprocess under lock\n'
b'        std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'        std::vector<float> featureVector;\n'
b'        Engine<float>::transformOutput(featureVectors, featureVector);\n'
b'        return PostProcessPose(featureVector, camera_id, meta);\n'
b'    }\n')
data = replace_once(data, old_det, new_det, 'DetectObjects refactor')

# 5a. Preprocess - remove inner lock, add outMeta param
old_pre_sig = L(
b'    std::vector<std::vector<cv::cuda::GpuMat>> ANSTENSORRTPOSE::Preprocess(const cv::Mat& inputImage) {\n'
b'        std::lock_guard<std::recursive_mutex> lock(_mutex);\n')
new_pre_sig = L(
b'    std::vector<std::vector<cv::cuda::GpuMat>> ANSTENSORRTPOSE::Preprocess(const cv::Mat& inputImage, ImageMetadata& outMeta) {\n')
data = replace_once(data, old_pre_sig, new_pre_sig, 'Preprocess sig + remove lock')

# 5b. Preprocess member writes -> outMeta
# Actual file has blank line + "// Set image size parameters" comment before the member writes
old_pw = L(
b'            stream.waitForCompletion();\n'
b'\n'
b'            // Set image size parameters\n'
b'            m_imgHeight = imgRGB.rows;\n'
b'            m_imgWidth = imgRGB.cols;\n'
b'            if (m_imgHeight > 0 && m_imgWidth > 0) {\n'
b'                m_ratio = 1.f / std::min(inputDims[0].d[2] / static_cast<float>(imgRGB.cols),\n'
b'                    inputDims[0].d[1] / static_cast<float>(imgRGB.rows));\n')
new_pw = L(
b'            stream.waitForCompletion();\n'
b'\n'
b'            // Set image size parameters\n'
b'            outMeta.imgHeight = imgRGB.rows;\n'
b'            outMeta.imgWidth = imgRGB.cols;\n'
b'            if (outMeta.imgHeight > 0 && outMeta.imgWidth > 0) {\n'
b'                outMeta.ratio = 1.f / std::min(inputDims[0].d[2] / static_cast<float>(imgRGB.cols),\n'
b'                    inputDims[0].d[1] / static_cast<float>(imgRGB.rows));\n')
data = replace_once(data, old_pw, new_pw, 'Preprocess member writes')

# 5c. Preprocess error log
old_pe = L(
b'                this->_logger.LogFatal("TENSORRTCL::Preprocess",\n'
b'                    "Image height or width is zero after processing (Width: " + std::to_string(m_imgWidth) +\n'
b'                    ", Height: " + std::to_string(m_imgHeight) + ")",\n')
new_pe = L(
b'                this->_logger.LogFatal("TENSORRTCL::Preprocess",\n'
b'                    "Image height or width is zero after processing (Width: " + std::to_string(outMeta.imgWidth) +\n'
b'                    ", Height: " + std::to_string(outMeta.imgHeight) + ")",\n')
data = replace_once(data, old_pe, new_pe, 'Preprocess error log')

# 6a. PostProcessPose - remove lock, add const ImageMetadata& meta param
old_ppp_sig = L(
b'    std::vector<Object> ANSTENSORRTPOSE::PostProcessPose(std::vector<float>& featureVector, const std::string& camera_id) {\n'
b'        std::lock_guard<std::recursive_mutex> lock(_mutex);\n')
new_ppp_sig = L(
b'    std::vector<Object> ANSTENSORRTPOSE::PostProcessPose(std::vector<float>& featureVector, const std::string& camera_id, const ImageMetadata& meta) {\n')
data = replace_once(data, old_ppp_sig, new_ppp_sig, 'PostProcessPose sig + remove lock')

# 6b/c/d. Replace m_ratio/m_imgWidth/m_imgHeight -> meta.* in PostProcessPose
# After steps 5b+5c fix Preprocess (removes 1/3/3), remaining counts: 6/4/4 (all in PostProcessPose)
data = replace_all(data, b'm_ratio', b'meta.ratio', 6, 'meta.ratio (6x in PostProcessPose)')
data = replace_all(data, b'm_imgWidth', b'meta.imgWidth', 4, 'meta.imgWidth (4x in PostProcessPose)')
data = replace_all(data, b'm_imgHeight', b'meta.imgHeight', 4, 'meta.imgHeight (4x in PostProcessPose)')

# 7. Rewrite DetectObjectsBatch + PreprocessBatch + PostProcessPoseBatch beginning
# Use anchor-based slice replacement to handle the exact bytes without hardcoding the full block
new_batch_all = L(
b'    std::vector<std::vector<Object>> ANSTENSORRTPOSE::DetectObjectsBatch(const std::vector<cv::Mat>& inputImages, const std::string& camera_id) {\n'
b'        {\n'
b'            std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'            if (inputImages.empty()) {\n'
b'                _logger.LogFatal("ANSTENSORRTPOSE::DetectObjectsBatch", "Empty input images vector", __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'        }\n'
b'\n'
b'        // Auto-split if batch exceeds engine capacity\n'
b'        const int maxBatch = m_options.maxBatchSize > 0 ? m_options.maxBatchSize : 1;\n'
b'        if (static_cast<int>(inputImages.size()) > maxBatch) {\n'
b'            const size_t numImages = inputImages.size();\n'
b'            std::vector<std::vector<cv::Mat>> chunks;\n'
b'            for (size_t start = 0; start < numImages; start += static_cast<size_t>(maxBatch)) {\n'
b'                const size_t end = std::min(start + static_cast<size_t>(maxBatch), numImages);\n'
b'                chunks.emplace_back(inputImages.begin() + start, inputImages.begin() + end);\n'
b'            }\n'
b'            std::vector<std::future<std::vector<std::vector<Object>>>> futures;\n'
b'            futures.reserve(chunks.size());\n'
b'            for (size_t i = 0; i < chunks.size(); ++i) {\n'
b'                futures.push_back(std::async(std::launch::async,\n'
b'                    [this, c = chunks[i], cid = camera_id]() {\n'
b'                        return DetectObjectsBatch(c, cid);\n'
b'                    }));\n'
b'            }\n'
b'            std::vector<std::vector<Object>> allResults;\n'
b'            allResults.reserve(numImages);\n'
b'            for (auto& fut : futures) {\n'
b'                auto chunkResults = fut.get();\n'
b'                for (auto& r : chunkResults) allResults.push_back(std::move(r));\n'
b'            }\n'
b'            return allResults;\n'
b'        }\n'
b'\n'
b'        _logger.LogInfo("ANSTENSORRTPOSE::DetectObjectsBatch",\n'
b'            "Processing batch of " + std::to_string(inputImages.size()) + " images", __FILE__, __LINE__);\n'
b'\n'
b'        // Phase 1: Preprocess under brief lock\n'
b'        BatchMetadata metadata;\n'
b'        std::vector<std::vector<cv::cuda::GpuMat>> inputs;\n'
b'        {\n'
b'            std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'            inputs = PreprocessBatch(inputImages, metadata);\n'
b'        }\n'
b'        if (inputs.empty() || inputs[0].empty()) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::DetectObjectsBatch", "Preprocessing failed", __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        // Phase 2: Inference -- mutex released; pool dispatches to idle GPU slot\n'
b'        std::vector<std::vector<std::vector<float>>> featureVectors;\n'
b'        auto succ = m_trtEngine->runInference(inputs, featureVectors);\n'
b'        if (!succ) {\n'
b'            _logger.LogError("ANSTENSORRTPOSE::DetectObjectsBatch", "Error running inference", __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'\n'
b'        // Phase 3: Parallel postprocessing -- each image is independent\n'
b'        const size_t numBatch = featureVectors.size();\n'
b'        std::vector<std::vector<Object>> batchDetections(numBatch);\n'
b'        std::vector<std::future<std::vector<Object>>> postFutures;\n'
b'        postFutures.reserve(numBatch);\n'
b'        for (size_t batchIdx = 0; batchIdx < numBatch; ++batchIdx) {\n'
b'            const auto& batchOutput = featureVectors[batchIdx];\n'
b'            std::vector<float> fv =\n'
b'                batchOutput.empty() ? std::vector<float>{} : batchOutput[0];\n'
b'            postFutures.push_back(std::async(std::launch::async,\n'
b'                [this, fv = std::move(fv), cid = camera_id,\n'
b'                 idx = batchIdx, &metadata]() mutable {\n'
b'                    return PostProcessPoseBatch(fv, cid, idx, metadata);\n'
b'                }));\n'
b'        }\n'
b'        for (size_t i = 0; i < numBatch; ++i)\n'
b'            batchDetections[i] = postFutures[i].get();\n'
b'\n'
b'        _logger.LogInfo("ANSTENSORRTPOSE::DetectObjectsBatch",\n'
b'            "Batch processing complete. Images: " + std::to_string(numBatch), __FILE__, __LINE__);\n'
b'        return batchDetections;\n'
b'    }\n'
b'    std::vector<std::vector<cv::cuda::GpuMat>> ANSTENSORRTPOSE::PreprocessBatch(const std::vector<cv::Mat>& inputImages, BatchMetadata& outMetadata) {\n'
b'        try {\n'
b'            if (!_licenseValid) {\n'
b'                _logger.LogFatal("ANSTENSORRTPOSE::PreprocessBatch", "Invalid license", __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'            const auto& inputDims = m_trtEngine->getInputDims();\n'
b'            const int inputH = inputDims[0].d[1];\n'
b'            const int inputW = inputDims[0].d[2];\n'
b'            outMetadata.imgHeights.resize(inputImages.size());\n'
b'            outMetadata.imgWidths.resize(inputImages.size());\n'
b'            outMetadata.ratios.resize(inputImages.size());\n'
b'            std::vector<cv::cuda::GpuMat> batchProcessed;\n'
b'            batchProcessed.reserve(inputImages.size());\n'
b'            cv::cuda::Stream stream;\n'
b'            for (size_t i = 0; i < inputImages.size(); ++i) {\n'
b'                const auto& inputImage = inputImages[i];\n'
b'                if (inputImage.empty()) {\n'
b'                    _logger.LogFatal("ANSTENSORRTPOSE::PreprocessBatch",\n'
b'                        "Empty input image at index " + std::to_string(i), __FILE__, __LINE__);\n'
b'                    return {};\n'
b'                }\n'
b'                cv::cuda::GpuMat img;\n'
b'                if (inputImage.channels() == 1) {\n'
b'                    cv::Mat img3Channel;\n'
b'                    cv::cvtColor(inputImage, img3Channel, cv::COLOR_GRAY2BGR);\n'
b'                    img.upload(img3Channel, stream);\n'
b'                }\n'
b'                else {\n'
b'                    img.upload(inputImage, stream);\n'
b'                }\n'
b'                cv::cuda::GpuMat imgRGB;\n'
b'                cv::cuda::cvtColor(img, imgRGB, cv::COLOR_BGR2RGB, 0, stream);\n'
b'                outMetadata.imgHeights[i] = imgRGB.rows;\n'
b'                outMetadata.imgWidths[i] = imgRGB.cols;\n'
b'                if (outMetadata.imgHeights[i] <= 0 || outMetadata.imgWidths[i] <= 0) {\n'
b'                    _logger.LogFatal("ANSTENSORRTPOSE::PreprocessBatch",\n'
b'                        "Image " + std::to_string(i) + " has invalid dimensions (Width: " +\n'
b'                        std::to_string(outMetadata.imgWidths[i]) + ", Height: " +\n'
b'                        std::to_string(outMetadata.imgHeights[i]) + ")", __FILE__, __LINE__);\n'
b'                    return {};\n'
b'                }\n'
b'                outMetadata.ratios[i] = 1.f / std::min(inputW / static_cast<float>(imgRGB.cols),\n'
b'                    inputH / static_cast<float>(imgRGB.rows));\n'
b'                cv::cuda::GpuMat resized = imgRGB;\n'
b'                if (resized.rows != inputH || resized.cols != inputW) {\n'
b'                    resized = Engine<float>::resizeKeepAspectRatioPadRightBottom(imgRGB, inputH, inputW);\n'
b'                }\n'
b'                batchProcessed.push_back(std::move(resized));\n'
b'            }\n'
b'            stream.waitForCompletion();\n'
b'            std::vector<std::vector<cv::cuda::GpuMat>> inputs;\n'
b'            inputs.push_back(std::move(batchProcessed));\n'
b'            return inputs;\n'
b'        }\n'
b'        catch (const std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::PreprocessBatch", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n'
b'    std::vector<Object> ANSTENSORRTPOSE::PostProcessPoseBatch(std::vector<float>& featureVector, const std::string& camera_id, size_t batchIdx, const BatchMetadata& metadata) {\n'
b'        try {\n'
b'            const auto& outputDims = m_trtEngine->getOutputDims();\n'
b'            auto numChannels = outputDims[0].d[1];\n'
b'            auto numAnchors = outputDims[0].d[2];\n'
b'            float ratio = metadata.ratios[batchIdx];\n'
b'            float imgWidth = static_cast<float>(metadata.imgWidths[batchIdx]);\n'
b'            float imgHeight = static_cast<float>(metadata.imgHeights[batchIdx]);\n')

data = replace_slice(data,
    start_anchor=L(b'    std::vector<Object> ANSTENSORRTPOSE::DetectObjectsBatch('),
    end_anchor=L(b'            float imgHeight = static_cast<float>(m_batchImgHeights[batchIdx]);\n'),
    new_content=new_batch_all,
    label='DetectObjectsBatch+PreprocessBatch+PostProcessPoseBatch rewrite')

# 8. Add RunInferencesBatch before namespace close
old_ns_close = L(
b'        catch (std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::PostProcessPoseBatch", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n'
b'}\n')

new_ns_close = L(
b'        catch (std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::PostProcessPoseBatch", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n'
b'    std::vector<std::vector<Object>> ANSTENSORRTPOSE::RunInferencesBatch(\n'
b'        const std::vector<cv::Mat>& inputs, const std::string& camera_id)\n'
b'    {\n'
b'        {\n'
b'            std::lock_guard<std::recursive_mutex> lock(_mutex);\n'
b'            if (!_modelLoadValid) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInferencesBatch",\n'
b'                    "Model not loaded", __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'            if (!_licenseValid) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInferencesBatch",\n'
b'                    "Invalid license", __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'            if (!_isInitialized) {\n'
b'                _logger.LogError("ANSTENSORRTPOSE::RunInferencesBatch",\n'
b'                    "Engine not initialized", __FILE__, __LINE__);\n'
b'                return {};\n'
b'            }\n'
b'            if (inputs.empty()) return {};\n'
b'        }\n'
b'        try {\n'
b'            return DetectObjectsBatch(inputs, camera_id);\n'
b'        }\n'
b'        catch (const std::exception& e) {\n'
b'            _logger.LogFatal("ANSTENSORRTPOSE::RunInferencesBatch", e.what(), __FILE__, __LINE__);\n'
b'            return {};\n'
b'        }\n'
b'    }\n'
b'}\n')
data = replace_once(data, old_ns_close, new_ns_close, 'Add RunInferencesBatch + namespace close')

# Summary
print(f'\nOriginal size: {orig_size} bytes')
print(f'New size:      {len(data)} bytes')
if errors:
    print(f'\n{len(errors)} ERROR(S) -- file NOT written')
    sys.exit(1)
else:
    with open(CPP, 'wb') as f:
        f.write(data)
    print('\nOK -- file written')