#include "bb_regressor.h"
#include <iostream>
#include <fstream>
#include <torch/script.h>
#include <torch/serialize.h>
#include <vector>
#include <stdexcept>
// Add CUDA includes for required CUDA implementation
#include <cuda_runtime.h>
#include <ATen/cuda/CUDAContext.h>

// Use the PrRoIPooling implementation
#include "prroi_pooling_gpu.h"
#include "prroi_pooling_gpu_impl.cuh"

// PrRoIPool2D implementation (requires CUDA)
PrRoIPool2D::PrRoIPool2D(int pooled_height, int pooled_width, float spatial_scale) 
    : pooled_height_(pooled_height), pooled_width_(pooled_width), spatial_scale_(spatial_scale) {}

torch::Tensor PrRoIPool2D::forward(torch::Tensor feat, torch::Tensor rois) {
    // Print shape info for debugging
    std::cout << "  PrRoIPool2D inputs: " << std::endl;
    std::cout << "    Features: [" << feat.size(0) << ", " << feat.size(1) << ", " 
              << feat.size(2) << ", " << feat.size(3) << "]" << std::endl;
    std::cout << "    ROIs: [" << rois.size(0) << ", " << rois.size(1) << "]" << std::endl;
    std::cout << "    Pooled size: [" << pooled_height_ << ", " << pooled_width_ << "]" << std::endl;
    std::cout << "    Spatial scale: " << spatial_scale_ << std::endl;
    
    // Calculate output shape
    int channels = feat.size(1);
    int num_rois = rois.size(0);
    
    // Ensure both tensors are on CUDA
    if (!feat.is_cuda() || !rois.is_cuda()) {
        throw std::runtime_error("PrRoIPool2D requires CUDA tensors - CPU mode is not supported");
    }
    
    // Print ROI values for debugging
    std::cout << "    ROI values: " << std::endl;
    for (int i = 0; i < std::min(num_rois, 3); i++) {
        std::cout << "      ROI " << i << ": [";
        for (int j = 0; j < rois.size(1); j++) {
            std::cout << rois[i][j].item<float>();
            if (j < rois.size(1) - 1) std::cout << ", ";
        }
        std::cout << "]" << std::endl;
    }
    
    // Create output tensor on the same device
    auto output = torch::zeros({num_rois, channels, pooled_height_, pooled_width_}, 
                              feat.options());
                              
    // Copy tensors to CPU for the C implementation
    auto feat_cpu = feat.to(torch::kCPU).contiguous();
    auto rois_cpu = rois.to(torch::kCPU).contiguous();
    auto output_cpu = output.to(torch::kCPU).contiguous();
    
    // Call the C wrapper function
    std::cout << "    Calling prroi_pooling_forward_cuda..." << std::endl;
    prroi_pooling_forward_cuda(
        feat_cpu.data_ptr<float>(),
        static_cast<float*>(rois_cpu.data_ptr()),
        static_cast<float*>(output_cpu.data_ptr()),
        channels,
        feat.size(2),
        feat.size(3),
        num_rois,
        pooled_height_,
        pooled_width_,
        spatial_scale_
    );
    std::cout << "    prroi_pooling_forward_cuda completed" << std::endl;
    
    // Copy result back to GPU
    output.copy_(output_cpu);
    
    return output;
}

// LinearBlock implementation
LinearBlock::LinearBlock(int in_planes, int out_planes, int input_sz, bool bias, bool batch_norm, bool relu) {
    // Create the linear layer with proper input dimensions
    auto linear_options = torch::nn::LinearOptions(in_planes * input_sz * input_sz, out_planes).bias(bias);
    linear = register_module("linear", torch::nn::Linear(linear_options));
    
    use_bn = batch_norm;
    if (use_bn) {
        // Important: use BatchNorm2d to match Python implementation
        bn = register_module("bn", torch::nn::BatchNorm2d(torch::nn::BatchNorm2dOptions(out_planes)));
    }
    
    use_relu = relu;
    if (use_relu) {
        relu_ = register_module("relu", torch::nn::ReLU(torch::nn::ReLUOptions().inplace(true)));
    }
}

torch::Tensor LinearBlock::forward(torch::Tensor x) {
    // Store original dtype for later
    auto original_dtype = x.dtype();
    
    // Use double precision for higher accuracy
    auto x_double = x.to(torch::kFloat64);
    
    // Reshape exactly as in Python: x.reshape(x.shape[0], -1)
    x_double = x_double.reshape({x_double.size(0), -1}).contiguous();
    
    // Convert back to original precision for the linear operation
    auto x_float = x_double.to(original_dtype);
    x_float = linear->forward(x_float);
    
    // Back to double precision for further operations
    x_double = x_float.to(torch::kFloat64);
    
    if (use_bn) {
        // This is crucial: reshape to 4D tensor for BatchNorm2d exactly as in Python
        // In Python: x = self.bn(x.reshape(x.shape[0], x.shape[1], 1, 1))
        x_double = x_double.reshape({x_double.size(0), x_double.size(1), 1, 1}).contiguous();
        
        // Apply batch norm (convert to float32 for the operation)
        x_float = x_double.to(original_dtype);
        x_float = bn->forward(x_float);
        x_double = x_float.to(torch::kFloat64);
    }
    
    // Apply ReLU if needed
    if (use_relu) {
        // Apply ReLU in float32 precision
        x_float = x_double.to(original_dtype);
        x_float = relu_->forward(x_float);
        x_double = x_float.to(torch::kFloat64);
    }
    
    // Final reshape to 2D tensor, exactly matching Python's behavior
    x_double = x_double.reshape({x_double.size(0), -1}).contiguous();
    
    // Return tensor in original precision
    return x_double.to(original_dtype);
}

// Create convolutional block
torch::nn::Sequential BBRegressor::create_conv_block(int in_planes, int out_planes, 
                                                  int kernel_size, int stride, 
                                                  int padding, int dilation) {
    // Print dimensions for debugging
    std::cout << "Creating conv block: in_planes=" << in_planes << ", out_planes=" << out_planes << std::endl;

    torch::nn::Sequential seq;
    
    // Add convolutional layer
    seq->push_back(torch::nn::Conv2d(torch::nn::Conv2dOptions(in_planes, out_planes, kernel_size)
                          .stride(stride).padding(padding).dilation(dilation).bias(true)));
    
    // Add batch normalization layer
    seq->push_back(torch::nn::BatchNorm2d(torch::nn::BatchNorm2dOptions(out_planes)));
    
    // Add ReLU activation
    seq->push_back(torch::nn::ReLU(torch::nn::ReLUOptions().inplace(true)));
    
    return seq;
}

// Helper function to verify BatchNorm dimensions
void BBRegressor::verify_batchnorm_dimensions() {
    std::cout << "Verifying BatchNorm dimensions..." << std::endl;
    
    // Get children of conv3_1r
    std::cout << "conv3_1r has " << conv3_1r->size() << " modules" << std::endl;
    if (conv3_1r->size() > 1) {
        auto module = conv3_1r[1];
        std::cout << "conv3_1r module[1] type: " << module->name() << std::endl;
    }
    
    // Get children of conv3_1t
    std::cout << "conv3_1t has " << conv3_1t->size() << " modules" << std::endl;
    if (conv3_1t->size() > 1) {
        auto module = conv3_1t[1];
        std::cout << "conv3_1t module[1] type: " << module->name() << std::endl;
    }
    
    // Get children of conv3_2t
    std::cout << "conv3_2t has " << conv3_2t->size() << " modules" << std::endl;
    if (conv3_2t->size() > 1) {
        auto module = conv3_2t[1];
        std::cout << "conv3_2t module[1] type: " << module->name() << std::endl;
    }
}

// Helper function to read file to bytes
std::vector<char> BBRegressor::read_file_to_bytes(const std::string& file_path) {
    std::ifstream file(file_path, std::ios::binary | std::ios::ate);
    if (!file.is_open()) {
        throw std::runtime_error("Could not open file: " + file_path);
    }
    
    std::streamsize size = file.tellg();
    file.seekg(0, std::ios::beg);
    
    std::vector<char> buffer(size);
    if (!file.read(buffer.data(), size)) {
        throw std::runtime_error("Could not read file: " + file_path);
    }
    
    return buffer;
}

// Load tensor from file
torch::Tensor BBRegressor::load_tensor(const std::string& file_path) {
    try {
        // Read file into bytes first
        std::vector<char> data = read_file_to_bytes(file_path);
        // Use pickle_load with byte data
        torch::Tensor tensor = torch::pickle_load(data).toTensor();
        
        // Always move tensor to the specified device
        return tensor.to(device);
    } catch (const c10::Error& e) {
        std::cerr << "Error loading tensor from " << file_path << ": " << e.what() << std::endl;
        throw;
    }
}

// Constructor
BBRegressor::BBRegressor(const std::string& model_weights_dir, torch::Device dev) 
    : device(dev), model_dir(model_weights_dir), 
      fc3_rt(256, 256, 5, true, true, true),
      fc4_rt(256, 256, 3, true, true, true) {
    
    // Check if model directory exists
    if (!fs::exists(model_dir)) {
        throw std::runtime_error("Model directory does not exist: " + model_dir);
    }
    
    // Initialize convolution blocks - match Python's AtomIoUNet implementation exactly
    std::cout << "Initializing conv blocks..." << std::endl;
    
    // In Python: self.conv3_1r = conv(input_dim[0], 128, kernel_size=3, stride=1)
    conv3_1r = create_conv_block(512, 128, 3, 1, 1, 1);
    
    // In Python: self.conv3_1t = conv(input_dim[0], 256, kernel_size=3, stride=1)
    conv3_1t = create_conv_block(512, 256, 3, 1, 1, 1);
    
    // In Python: self.conv3_2t = conv(256, pred_input_dim[0], kernel_size=3, stride=1)
    conv3_2t = create_conv_block(256, 256, 3, 1, 1, 1);
    
    // Update pooling sizes to match the Python model exactly
    // In Python: self.prroi_pool3r = PrRoIPool2D(3, 3, 1/8)
    prroi_pool3r = std::make_shared<PrRoIPool2D>(3, 3, 0.125); // 1/8 scale for layer2
    
    // In Python: self.prroi_pool3t = PrRoIPool2D(5, 5, 1/8)
    prroi_pool3t = std::make_shared<PrRoIPool2D>(5, 5, 0.125); // 1/8 scale for layer2
    
    // Create sequential blocks
    // In Python: self.fc3_1r = conv(128, 256, kernel_size=3, stride=1, padding=0)
    fc3_1r = create_conv_block(128, 256, 3, 1, 0, 1);  // padding=0 for this layer
    
    // In Python: self.conv4_1r = conv(input_dim[1], 256, kernel_size=3, stride=1)
    conv4_1r = create_conv_block(1024, 256, 3, 1, 1, 1);
    
    // In Python: self.conv4_1t = conv(input_dim[1], 256, kernel_size=3, stride=1)
    conv4_1t = create_conv_block(1024, 256, 3, 1, 1, 1);
    
    // In Python: self.conv4_2t = conv(256, pred_input_dim[1], kernel_size=3, stride=1)
    conv4_2t = create_conv_block(256, 256, 3, 1, 1, 1);
    
    // In Python: self.prroi_pool4r = PrRoIPool2D(1, 1, 1/16)
    prroi_pool4r = std::make_shared<PrRoIPool2D>(1, 1, 0.0625); // 1/16 scale for layer3
    
    // In Python: self.prroi_pool4t = PrRoIPool2D(3, 3, 1/16)
    prroi_pool4t = std::make_shared<PrRoIPool2D>(3, 3, 0.0625); // 1/16 scale for layer3
    
    // In Python: self.fc34_3r = conv(256 + 256, pred_input_dim[0], kernel_size=1, stride=1, padding=0)
    fc34_3r = create_conv_block(512, 256, 1, 1, 0, 1);  // kernel_size=1, padding=0
    
    // In Python: self.fc34_4r = conv(256 + 256, pred_input_dim[1], kernel_size=1, stride=1, padding=0)
    fc34_4r = create_conv_block(512, 256, 1, 1, 0, 1);  // kernel_size=1, padding=0
    
    // Linear blocks - exactly match Python's implementation dimensions and parameters
    // In Python: self.fc3_rt = LinearBlock(pred_input_dim[0], pred_inter_dim[0], 5)
    fc3_rt = LinearBlock(256, 256, 5, true, true, true);
    
    // In Python: self.fc4_rt = LinearBlock(pred_input_dim[1], pred_inter_dim[1], 3)
    fc4_rt = LinearBlock(256, 256, 3, true, true, true);
    
    // In Python: self.iou_predictor = nn.Linear(pred_inter_dim[0]+pred_inter_dim[1], 1, bias=True)
    iou_predictor = torch::nn::Linear(torch::nn::LinearOptions(256 + 256, 1).bias(true));
    
    // Load all weights
    load_weights();
    
    // Set the model to evaluation mode
    this->eval();
    
    // Debug information
    std::cout << "BB Regressor initialized in evaluation mode" << std::endl;
}

// Set the model to evaluation mode
void BBRegressor::eval() {
    // Set all sequential modules to eval mode
    conv3_1r->eval();
    conv3_1t->eval();
    conv3_2t->eval();
    fc3_1r->eval();
    conv4_1r->eval();
    conv4_1t->eval();
    conv4_2t->eval();
    fc34_3r->eval();
    fc34_4r->eval();
    
    // Linear blocks also need to be set to eval mode for BatchNorm layers
    fc3_rt.eval();
    fc4_rt.eval();
    
    // Set linear layers to eval mode (though this usually doesn't have any effect)
    iou_predictor->eval();
}

// Load weights
void BBRegressor::load_weights() {
    // Helper lambda to load weights for a sequential module
    auto load_sequential_weights = [this](torch::nn::Sequential& seq, const std::string& prefix) {
        try {
            // Load weights for conv layer (index 0)
            std::string weight_path = model_dir + "/" + prefix + "_0_weight.pt";
            std::string bias_path = model_dir + "/" + prefix + "_0_bias.pt";
            
            if (fs::exists(weight_path) && fs::exists(bias_path)) {
                auto conv_weight = load_tensor(weight_path);
                auto conv_bias = load_tensor(bias_path);
                
                // Get the conv2d module from sequential
                // Fix: Get the number of output channels from the weight tensor
                int out_channels = conv_weight.size(0);
                int in_channels = conv_weight.size(1);
                int kernel_size = conv_weight.size(2);
                
                std::cout << "Loading " << prefix << " conv weights: " 
                          << "[out_ch=" << out_channels 
                          << ", in_ch=" << in_channels 
                          << ", kernel=" << kernel_size << "]" << std::endl;
                
                // FIXED: Use the correct padding based on the layer name
                int padding = 1; // Default padding
                
                // Special cases for layers with different padding
                if (prefix == "fc3_1r" || prefix == "fc34_3r" || prefix == "fc34_4r") {
                    padding = 0; // These layers use padding=0 in the Python implementation
                }
                
                std::cout << "  Using padding=" << padding << " for " << prefix << std::endl;
                
                auto conv_options = torch::nn::Conv2dOptions(in_channels, out_channels, kernel_size)
                    .stride(1).padding(padding).bias(true);
                auto conv_module = torch::nn::Conv2d(conv_options);
                
                // Set weights and bias
                conv_module->weight = conv_weight;
                conv_module->bias = conv_bias;
                
                // Debug info - print some weight stats
                std::cout << "  Conv weight stats: mean=" << conv_weight.mean().item<float>()
                          << ", std=" << conv_weight.std().item<float>()
                          << ", min=" << conv_weight.min().item<float>()
                          << ", max=" << conv_weight.max().item<float>() << std::endl;
                
                // Create a new sequence with the proper conv module
                auto new_seq = torch::nn::Sequential();
                new_seq->push_back(conv_module);
                
                // Load batch norm parameters (index 1)
                std::string bn_weight_path = model_dir + "/" + prefix + "_1_weight.pt";
                std::string bn_bias_path = model_dir + "/" + prefix + "_1_bias.pt";
                std::string bn_mean_path = model_dir + "/" + prefix + "_1_running_mean.pt";
                std::string bn_var_path = model_dir + "/" + prefix + "_1_running_var.pt";
                
                if (fs::exists(bn_weight_path) && fs::exists(bn_bias_path) &&
                    fs::exists(bn_mean_path) && fs::exists(bn_var_path)) {
                    
                    auto bn_weight = load_tensor(bn_weight_path);
                    auto bn_bias = load_tensor(bn_bias_path);
                    auto bn_mean = load_tensor(bn_mean_path);
                    auto bn_var = load_tensor(bn_var_path);
                    
                    // Important: Create BatchNorm with the correct number of features from the weights
                    int num_features = bn_weight.size(0);
                    std::cout << "  Creating BatchNorm2d with num_features=" << num_features << std::endl;
                    
                    // Create a proper batch norm module with the right number of features
                    auto bn_options = torch::nn::BatchNorm2dOptions(num_features)
                                      .eps(1e-5)  // Match Python default
                                      .momentum(0.1)  // Match Python default
                                      .affine(true)
                                      .track_running_stats(true);
                    auto bn_module = torch::nn::BatchNorm2d(bn_options);
                    
                    // Set batch norm parameters
                    bn_module->weight = bn_weight;
                    bn_module->bias = bn_bias;
                    bn_module->running_mean = bn_mean;
                    bn_module->running_var = bn_var;
                    
                    // Debug info - print some batch norm stats
                    std::cout << "  BN weight stats: mean=" << bn_weight.mean().item<float>()
                              << ", std=" << bn_weight.std().item<float>() << std::endl;
                    std::cout << "  BN running_mean stats: mean=" << bn_mean.mean().item<float>()
                              << ", std=" << bn_mean.std().item<float>() << std::endl;
                    std::cout << "  BN running_var stats: mean=" << bn_var.mean().item<float>()
                              << ", std=" << bn_var.std().item<float>() << std::endl;
                    
                    // Add the batch norm module to the sequence
                    new_seq->push_back(bn_module);
                }
                
                // Add the ReLU module with inplace=true to match Python
                auto relu_options = torch::nn::ReLUOptions().inplace(true);
                new_seq->push_back(torch::nn::ReLU(relu_options));
                
                // Replace the old sequence with the new one
                seq = new_seq;
                
                std::cout << "Loaded weights for " << prefix << std::endl;
            } else {
                std::cerr << "Weight files not found for " << prefix << std::endl;
            }
        } catch (const std::exception& e) {
            std::cerr << "Error loading weights for " << prefix << ": " << e.what() << std::endl;
            throw; // Re-throw to stop execution
        }
    };
    
    // Load weights for linear blocks
    auto load_linear_block_weights = [this](LinearBlock& block, const std::string& prefix) {
        try {
            // Load weights for linear layer
            std::string weight_path = model_dir + "/" + prefix + "_linear_weight.pt";
            std::string bias_path = model_dir + "/" + prefix + "_linear_bias.pt";
            
            if (fs::exists(weight_path) && fs::exists(bias_path)) {
                auto linear_weight = load_tensor(weight_path);
                auto linear_bias = load_tensor(bias_path);
                
                // Set weights and bias
                block.linear->weight = linear_weight;
                block.linear->bias = linear_bias;
                
                // Load batch norm parameters
                std::string bn_weight_path = model_dir + "/" + prefix + "_bn_weight.pt";
                std::string bn_bias_path = model_dir + "/" + prefix + "_bn_bias.pt";
                std::string bn_mean_path = model_dir + "/" + prefix + "_bn_running_mean.pt";
                std::string bn_var_path = model_dir + "/" + prefix + "_bn_running_var.pt";
                
                if (fs::exists(bn_weight_path) && fs::exists(bn_bias_path) &&
                    fs::exists(bn_mean_path) && fs::exists(bn_var_path)) {
                    auto bn_weight = load_tensor(bn_weight_path);
                    auto bn_bias = load_tensor(bn_bias_path);
                    auto bn_mean = load_tensor(bn_mean_path);
                    auto bn_var = load_tensor(bn_var_path);
                    
                    // Set batch norm parameters
                    block.bn->weight = bn_weight;
                    block.bn->bias = bn_bias;
                    block.bn->running_mean = bn_mean;
                    block.bn->running_var = bn_var;
                }
                
                std::cout << "Loaded weights for " << prefix << std::endl;
            } else {
                std::cerr << "Weight files not found for " << prefix << std::endl;
            }
        } catch (const std::exception& e) {
            std::cerr << "Error loading weights for " << prefix << ": " << e.what() << std::endl;
            throw; // Re-throw to stop execution
        }
    };
    
    // Load weights for all layers
    load_sequential_weights(conv3_1r, "conv3_1r");
    load_sequential_weights(conv3_1t, "conv3_1t");
    load_sequential_weights(conv3_2t, "conv3_2t");
    load_sequential_weights(fc3_1r, "fc3_1r");
    load_sequential_weights(conv4_1r, "conv4_1r");
    load_sequential_weights(conv4_1t, "conv4_1t");
    load_sequential_weights(conv4_2t, "conv4_2t");
    load_sequential_weights(fc34_3r, "fc34_3r");
    load_sequential_weights(fc34_4r, "fc34_4r");
    
    load_linear_block_weights(fc3_rt, "fc3_rt");
    load_linear_block_weights(fc4_rt, "fc4_rt");
    
    // Load IoU predictor weights
    try {
        std::string weight_path = model_dir + "/iou_predictor_weight.pt";
        std::string bias_path = model_dir + "/iou_predictor_bias.pt";
        
        if (fs::exists(weight_path) && fs::exists(bias_path)) {
            auto weight = load_tensor(weight_path);
            auto bias = load_tensor(bias_path);
            
            iou_predictor->weight = weight;
            iou_predictor->bias = bias;
            
            std::cout << "Loaded weights for iou_predictor" << std::endl;
        } else {
            std::cerr << "Weight files not found for iou_predictor" << std::endl;
        }
    } catch (const std::exception& e) {
        std::cerr << "Error loading weights for iou_predictor: " << e.what() << std::endl;
        throw; // Re-throw to stop execution
    }
}

// Move model to device
void BBRegressor::to(torch::Device device) {
    // Verify the device is a CUDA device
    if (!device.is_cuda()) {
        throw std::runtime_error("BBRegressor requires a CUDA device");
    }
    
    this->device = device;
    
    // Move all components to device
    conv3_1r->to(device);
    conv3_1t->to(device);
    conv3_2t->to(device);
    fc3_1r->to(device);
    conv4_1r->to(device);
    conv4_1t->to(device);
    conv4_2t->to(device);
    fc3_rt.to(device);
    fc4_rt.to(device);
    
    iou_predictor->to(device);
}

// Get IoU features from backbone features
std::vector<torch::Tensor> BBRegressor::get_iou_feat(std::vector<torch::Tensor> feat2) {
    // Convert to double precision for better numerical stability
    auto feat2_double0 = feat2[0].to(torch::kFloat64);
    auto feat2_double1 = feat2[1].to(torch::kFloat64);
    
    // Reshape exactly as in Python implementation
    // In Python: feat2 = [f.reshape(-1, *f.shape[-3:]) if f.dim()==5 else f for f in feat2]
    if (feat2_double0.dim() == 5) {
        auto shape = feat2_double0.sizes();
        feat2_double0 = feat2_double0.reshape({-1, shape[2], shape[3], shape[4]}).contiguous();
    }
    
    if (feat2_double1.dim() == 5) {
        auto shape = feat2_double1.sizes();
        feat2_double1 = feat2_double1.reshape({-1, shape[2], shape[3], shape[4]}).contiguous();
    }
    
    // Convert back to float32 for convolution operations
    feat2[0] = feat2_double0.to(torch::kFloat32).contiguous();
    feat2[1] = feat2_double1.to(torch::kFloat32).contiguous();
    
    // Apply convolutions exactly as in Python
    torch::Tensor feat3_t = feat2[0];
    torch::Tensor feat4_t = feat2[1];
    
    // Ensure we're in evaluation mode
    torch::NoGradGuard no_grad;
    
    // Apply convolutions just like Python version
    torch::Tensor c3_t_1 = conv3_1t->forward(feat3_t);
    c3_t_1 = c3_t_1.contiguous();
    
    torch::Tensor c3_t = conv3_2t->forward(c3_t_1);
    c3_t = c3_t.contiguous();
    
    torch::Tensor c4_t_1 = conv4_1t->forward(feat4_t);
    c4_t_1 = c4_t_1.contiguous();
    
    torch::Tensor c4_t = conv4_2t->forward(c4_t_1);
    c4_t = c4_t.contiguous();
    
    // Return results
    return {c3_t, c4_t};
}

// Get modulation vectors for the target
std::vector<torch::Tensor> BBRegressor::get_modulation(std::vector<torch::Tensor> feat, torch::Tensor bb) {
    // feat should contain two tensors: feat3_r and feat4_r (backbone features)
    // bb is the initial bounding box [batch_size, 1, 4] (x,y,w,h) or [batch_size, 4]
    // Ensure inputs are on the correct device
    torch::NoGradGuard no_grad; // Ensure no gradients are computed

    auto feat3_r = feat[0].to(device);
    auto feat4_r = feat[1].to(device);
    auto current_bb = bb.to(device);

    // Reshape bb if it's [batch, 1, 4] to [batch, 4]
    if (current_bb.dim() == 3 && current_bb.size(1) == 1) {
        current_bb = current_bb.squeeze(1);
    }
    if (current_bb.dim() != 2 || current_bb.size(1) != 4) {
        throw std::runtime_error("BBRegressor::get_modulation: bb must be [batch, 4] or [batch, 1, 4]");
    }

    // Pass through early conv layers (reference branch)
    // Python: c3_r = self.conv3_1r(feat3_r)
    auto c3_r = conv3_1r->forward(feat3_r);

    // Prepare ROIs: convert bb from [x,y,w,h] to [batch_idx, x1,y1,x2,y2]
    int batch_size = current_bb.size(0);
    auto batch_idx = torch::arange(0, batch_size, current_bb.options().dtype(torch::kFloat)).unsqueeze(1);
    
    auto rois = torch::zeros({batch_size, 5}, current_bb.options());
    rois.index_put_({torch::indexing::Slice(), 0}, batch_idx.squeeze(1)); // batch index
    rois.index_put_({torch::indexing::Slice(), 1}, current_bb.index({torch::indexing::Slice(), 0})); // x1
    rois.index_put_({torch::indexing::Slice(), 2}, current_bb.index({torch::indexing::Slice(), 1})); // y1
    rois.index_put_({torch::indexing::Slice(), 3}, current_bb.index({torch::indexing::Slice(), 0}) + current_bb.index({torch::indexing::Slice(), 2})); // x2 = x1 + w
    rois.index_put_({torch::indexing::Slice(), 4}, current_bb.index({torch::indexing::Slice(), 1}) + current_bb.index({torch::indexing::Slice(), 3})); // y2 = y1 + h
    
    rois = rois.to(device); // Ensure ROIs are on the correct device

    std::cout << "  BBRegressor::get_modulation: Converted ROIs (first item): [";
    if (batch_size > 0) {
        for (int j = 0; j < rois.size(1); j++) {
            std::cout << rois[0][j].item<float>();
            if (j < rois.size(1) - 1) std::cout << ", ";
        }
    }
    std::cout << "]" << std::endl;
    std::cout << "  BBRegressor::get_modulation: c3_r shape: " << c3_r.sizes() << ", device: " << c3_r.device() << std::endl;


    // Python: roi3r = self.prroi_pool3r(c3_r, roi1)
    auto roi3r = prroi_pool3r->forward(c3_r, rois);
    std::cout << "  BBRegressor::get_modulation: roi3r shape: " << roi3r.sizes() << std::endl;

    // Python: c4_r = self.conv4_1r(feat4_r)
    auto c4_r = conv4_1r->forward(feat4_r);
    std::cout << "  BBRegressor::get_modulation: c4_r shape: " << c4_r.sizes() << ", device: " << c4_r.device() << std::endl;

    // Python: roi4r = self.prroi_pool4r(c4_r, roi1)
    auto roi4r = prroi_pool4r->forward(c4_r, rois);
    std::cout << "  BBRegressor::get_modulation: roi4r shape: " << roi4r.sizes() << std::endl;


    // Python: fc3_r = self.fc3_1r(roi3r)
    // fc3_1r is a conv block: conv(128, 256, kernel_size=3, stride=1, padding=0)
    // Input roi3r is (batch, 128, 3, 3) -> Output fc3_r is (batch, 256, 1, 1)
    auto fc3_r = fc3_1r->forward(roi3r);
    std::cout << "  BBRegressor::get_modulation: fc3_r shape: " << fc3_r.sizes() << std::endl;

    // Python: fc34_r = torch.cat((fc3_r, roi4r), dim=1)
    // fc3_r is (batch, 256, 1, 1), roi4r is (batch, 256, 1, 1)
    // Result fc34_r is (batch, 512, 1, 1)
    auto fc34_r = torch::cat({fc3_r, roi4r}, 1);
    std::cout << "  BBRegressor::get_modulation: fc34_r shape: " << fc34_r.sizes() << std::endl;

    // Python: fc34_3_r = self.fc34_3r(fc34_r)
    // fc34_3r is conv(512, 256, kernel_size=1, stride=1, padding=0)
    // Output fc34_3_r is (batch, 256, 1, 1)
    auto mod_vec1 = fc34_3r->forward(fc34_r);
    std::cout << "  BBRegressor::get_modulation: mod_vec1 (fc34_3_r) shape: " << mod_vec1.sizes() << std::endl;

    // Python: fc34_4_r = self.fc34_4r(fc34_r)
    // fc34_4r is conv(512, 256, kernel_size=1, stride=1, padding=0)
    // Output fc34_4_r is (batch, 256, 1, 1)
    auto mod_vec2 = fc34_4r->forward(fc34_r);
    std::cout << "  BBRegressor::get_modulation: mod_vec2 (fc34_4_r) shape: " << mod_vec2.sizes() << std::endl;
    
    return {mod_vec1, mod_vec2};
}

// Predict IoU for proposals
torch::Tensor BBRegressor::predict_iou(std::vector<torch::Tensor> modulation, 
                                     std::vector<torch::Tensor> feat, 
                                     torch::Tensor proposals) {
    // Ensure all inputs are on the correct device
    auto target_device = device; // Assuming 'device' is a member of BBRegressor
    for (auto& t : feat) { t = t.to(target_device); }
    for (auto& m : modulation) { m = m.to(target_device); }
    proposals = proposals.to(target_device);

    // Get batch size and number of proposals
    int batch_size = proposals.size(0);
    int num_proposals = proposals.size(1);

    // Reshape proposals to [batch_size * num_proposals, 4]
    // and add batch index for PrRoIPooling
    auto proposals_view = proposals.reshape({batch_size * num_proposals, 4});
    auto roi_batch_index = torch::arange(0, batch_size, proposals.options().dtype(torch::kInt)).unsqueeze(1);
    roi_batch_index = roi_batch_index.repeat_interleave(num_proposals, 0);
    auto roi = torch::cat(std::vector<torch::Tensor>{roi_batch_index.to(proposals_view.options()), proposals_view}, 1);
    
    // Ensure ROI is on the correct device, matching features
    auto feat_device = feat[0].device();
    roi = roi.to(feat_device);
    
    // Apply ROI pooling to get features for each proposal
    auto pooled_feat1 = prroi_pool3t->forward(feat[0], roi); // Output: [batch_size * num_proposals, C, 5, 5]
    auto pooled_feat2 = prroi_pool4t->forward(feat[1], roi); // Output: [batch_size * num_proposals, C, 3, 3]
    
    std::cout << "  Pooled shapes:" << std::endl;
    std::cout << "    pooled_feat1 (from prroi_pool3t on feat[0]): [" << pooled_feat1.sizes() << "] dev: " << pooled_feat1.device() << std::endl;
    std::cout << "    pooled_feat2 (from prroi_pool4t on feat[1]): [" << pooled_feat2.sizes() << "] dev: " << pooled_feat2.device() << std::endl;
    
    std::cout << "  IoU predictor dimensions:" << std::endl;
    std::cout << "    weight: [" << iou_predictor->weight.sizes() << "]" << std::endl;
    std::cout << "    bias: [" << iou_predictor->bias.sizes() << "]" << std::endl;
    
    try {
        auto mod0_4d = modulation[0].to(target_device); 
        auto mod1_4d = modulation[1].to(target_device); 

        if (mod0_4d.dim() == 2) {
            mod0_4d = mod0_4d.reshape({mod0_4d.size(0), mod0_4d.size(1), 1, 1});
        }
        if (mod1_4d.dim() == 2) {
            mod1_4d = mod1_4d.reshape({mod1_4d.size(0), mod1_4d.size(1), 1, 1});
        }
        
        if (mod0_4d.size(0) == 1 && pooled_feat1.size(0) > 1) {
            mod0_4d = mod0_4d.repeat({pooled_feat1.size(0), 1, 1, 1});
        }
        if (mod1_4d.size(0) == 1 && pooled_feat2.size(0) > 1) {
            mod1_4d = mod1_4d.repeat({pooled_feat2.size(0), 1, 1, 1});
        }

        std::cout << "  Modulation vector shapes (reshaped 4D):" << std::endl;
        std::cout << "    mod0_4d: [" << mod0_4d.sizes() << "] dev: " << mod0_4d.device() << std::endl;
        std::cout << "    mod1_4d: [" << mod1_4d.sizes() << "] dev: " << mod1_4d.device() << std::endl;
        
        auto feat_prod_0 = pooled_feat1 * mod0_4d; 
        auto feat_prod_1 = pooled_feat2 * mod1_4d; 
        std::cout << "  After element-wise product with modulation:\n    feat_prod_0 (pooled_feat1 * mod0_4d): [" << feat_prod_0.sizes() << "] dev: " << feat_prod_0.device() << "\n    feat_prod_1 (pooled_feat2 * mod1_4d): [" << feat_prod_1.sizes() << "] dev: " << feat_prod_1.device() << std::endl;
        
        std::cout << "  Applying fc3_rt to feat_prod_0..." << std::endl;
        auto x0 = fc3_rt.forward(feat_prod_0); // Corrected: . instead of ->
        std::cout << "  Applying fc4_rt to feat_prod_1..." << std::endl;
        auto x1 = fc4_rt.forward(feat_prod_1); // Corrected: . instead of ->
        std::cout << "  After fc_rt blocks:\n    x0 (fc3_rt output): [" << x0.sizes() << "] dev: " << x0.device() << "\n    x1 (fc4_rt output): [" << x1.sizes() << "] dev: " << x1.device() << std::endl;

        auto ioufeat_final = torch::cat(std::vector<torch::Tensor>{x0, x1}, 1).contiguous(); // Corrected: std::vector wrapper
        std::cout << "  Concatenated ioufeat_final: [" << ioufeat_final.sizes() << "] dev: " << ioufeat_final.device() << std::endl;

        torch::Tensor iou_scores;
        try {
            std::cout << "  Applying final iou_predictor on GPU" << std::endl;
            iou_predictor->to(target_device); 
            iou_scores = iou_predictor->forward(ioufeat_final.to(target_device));
            std::cout << "  Final iou_predictor on GPU successful. Output scores shape: [" << iou_scores.sizes() << "]" << std::endl;

        } catch (const std::exception& cuda_error) {
            std::cout << "  GPU iou_predictor->forward() failed: " << cuda_error.what() << std::endl;
            std::cout << "  Falling back to CPU for final iou_predictor" << std::endl;
            
            auto ioufeat_final_cpu = ioufeat_final.to(torch::kCPU).contiguous();
            auto weight_cpu = iou_predictor->weight.to(torch::kCPU).contiguous();
            auto bias_cpu = torch::Tensor(); 
            if (iou_predictor->bias.defined()) {
                bias_cpu = iou_predictor->bias.to(torch::kCPU).contiguous();
            }

            std::cout << "    DEBUG CPU Fallback: ioufeat_final_cpu device: " << ioufeat_final_cpu.device() << std::endl;
            std::cout << "    DEBUG CPU Fallback: weight_cpu device: " << weight_cpu.device() << std::endl;
            if (bias_cpu.defined()) {
                 std::cout << "    DEBUG CPU Fallback: bias_cpu device: " << bias_cpu.device() << std::endl;
            } else {
                 std::cout << "    DEBUG CPU Fallback: bias_cpu is undefined." << std::endl;
            }
            
            try {
                iou_scores = torch::nn::functional::linear(ioufeat_final_cpu, weight_cpu, bias_cpu);
                std::cout << "  CPU fallback torch::nn::functional::linear() successful. Output device: " << iou_scores.device() << std::endl;
            } catch (const std::exception& cpu_fwd_error) {
                std::cerr << "ERROR during CPU torch::nn::functional::linear(): " << cpu_fwd_error.what() << std::endl;
                iou_predictor->to(target_device); 
                throw; 
            }
            
            iou_predictor->to(target_device); 
            iou_scores = iou_scores.to(target_device);
        }
        
        std::cout << "  iou_scores raw shape: [" << iou_scores.size(0) << ", " << iou_scores.size(1) << "]" << std::endl;
        
        iou_scores = iou_scores.reshape({batch_size, num_proposals});
        std::cout << "  Final iou_scores shape: [" << iou_scores.size(0) << ", " << iou_scores.size(1) << "]" << std::endl;
        
        return iou_scores;
        
    } catch (const std::exception& e) {
        std::cerr << "CRITICAL: Unexpected error in predict_iou: " << e.what() << std::endl;
        std::cout << "  Propagating critical error. No fallback available for this stage." << std::endl;
        throw; 
    }
}

// Print model information
void BBRegressor::print_model_info() {
    std::cout << "BBRegressor Model Information:" << std::endl;
    std::cout << "  - Model directory: " << model_dir << std::endl;
    std::cout << "  - Device: CUDA:" << device.index() << std::endl;
    std::cout << "  - CUDA Device Count: " << torch::cuda::device_count() << std::endl;
    std::cout << "  - Using PreciseRoIPooling: " << 
    #ifdef WITH_PRROI_POOLING
                   "Yes" 
    #else
                   "No (will fail)"
    #endif
                   << std::endl;
}

// Compute statistics for a tensor
BBRegressor::TensorStats BBRegressor::compute_stats(const torch::Tensor& tensor) {
    TensorStats stats;
    
    // Get shape
    for (int i = 0; i < tensor.dim(); i++) {
        stats.shape.push_back(tensor.size(i));
    }
    
    // Compute basic stats - make sure we reduce to scalar values
    stats.mean = tensor.mean().item<float>();  // Mean of all elements
    stats.std_dev = tensor.std().item<float>(); // Std dev of all elements
    stats.min_val = tensor.min().item<float>(); // Min of all elements 
    stats.max_val = tensor.max().item<float>(); // Max of all elements
    stats.sum = tensor.sum().item<float>();    // Sum of all elements
    
    // Sample values at specific positions
    if (tensor.dim() >= 4) {
        // For 4D tensors (batch, channel, height, width)
        stats.samples.push_back(tensor.index({0, 0, 0, 0}).item<float>());
        
        if (tensor.size(1) > 1 && tensor.size(2) > 1 && tensor.size(3) > 1) {
            int mid_c = static_cast<int>(tensor.size(1) / 2);
            int mid_h = static_cast<int>(tensor.size(2) / 2);
            int mid_w = static_cast<int>(tensor.size(3) / 2);
            stats.samples.push_back(tensor.index({0, mid_c, mid_h, mid_w}).item<float>());
            
            // Use static_cast to convert int64_t to int to avoid type mismatch
            int64_t last_c_idx = tensor.size(1) - 1;
            int64_t last_h_idx = tensor.size(2) - 1;
            int64_t last_w_idx = tensor.size(3) - 1;
            
            // Limit indices to avoid accessing out of bounds
            if (last_c_idx > 10) last_c_idx = 10;
            if (last_h_idx > 10) last_h_idx = 10;
            if (last_w_idx > 10) last_w_idx = 10;
            
            stats.samples.push_back(tensor.index({0, static_cast<int>(last_c_idx), 
                                                static_cast<int>(last_h_idx), 
                                                static_cast<int>(last_w_idx)}).item<float>());
        }
    } else if (tensor.dim() == 3) {
        // For 3D tensors
        stats.samples.push_back(tensor.index({0, 0, 0}).item<float>());
        
        if (tensor.size(1) > 1 && tensor.size(2) > 1) {
            int mid_h = static_cast<int>(tensor.size(1) / 2);
            int mid_w = static_cast<int>(tensor.size(2) / 2);
            stats.samples.push_back(tensor.index({0, mid_h, mid_w}).item<float>());
            
            int last_h = static_cast<int>(tensor.size(1) - 1);
            int last_w = static_cast<int>(tensor.size(2) - 1);
            stats.samples.push_back(tensor.index({0, last_h, last_w}).item<float>());
        }
    } else if (tensor.dim() == 2) {
        // For 2D tensors
        stats.samples.push_back(tensor.index({0, 0}).item<float>());
        
        if (tensor.size(0) > 1 && tensor.size(1) > 1) {
            int mid_h = static_cast<int>(tensor.size(0) / 2);
            int mid_w = static_cast<int>(tensor.size(1) / 2);
            stats.samples.push_back(tensor.index({mid_h, mid_w}).item<float>());
            
            int last_h = static_cast<int>(tensor.size(0) - 1);
            int last_w = static_cast<int>(tensor.size(1) - 1);
            stats.samples.push_back(tensor.index({last_h, last_w}).item<float>());
        }
    } else {
        // For 1D tensors or scalars
        if (tensor.numel() > 0) {
            stats.samples.push_back(tensor.index({0}).item<float>());
        
            if (tensor.size(0) > 1) {
                int mid = static_cast<int>(tensor.size(0) / 2);
                stats.samples.push_back(tensor.index({mid}).item<float>());
                
                int last = static_cast<int>(tensor.size(0) - 1);
                stats.samples.push_back(tensor.index({last}).item<float>());
            }
        }
    }
    
    return stats;
}

// Save tensor statistics to a file
void BBRegressor::save_stats(const std::vector<TensorStats>& all_stats, const std::string& filepath) {
    std::ofstream file(filepath);
    if (!file.is_open()) {
        std::cerr << "Error opening file for writing: " << filepath << std::endl;
        return;
    }
    
    for (size_t i = 0; i < all_stats.size(); i++) {
        const auto& stats = all_stats[i];
        file << "Output " << i << ":" << std::endl;
        
        file << "  Shape: [";
        for (size_t j = 0; j < stats.shape.size(); j++) {
            file << stats.shape[j];
            if (j < stats.shape.size() - 1) file << ", ";
        }
        file << "]" << std::endl;
        
        file << "  Mean: " << stats.mean << std::endl;
        file << "  Std: " << stats.std_dev << std::endl;
        file << "  Min: " << stats.min_val << std::endl;
        file << "  Max: " << stats.max_val << std::endl;
        file << "  Sum: " << stats.sum << std::endl;
        
        file << "  Sample values: [";
        for (size_t j = 0; j < stats.samples.size(); j++) {
            file << stats.samples[j];
            if (j < stats.samples.size() - 1) file << ", ";
        }
        file << "]" << std::endl << std::endl;
    }
    
    file.close();
}