cimp-impl/cimp/bb_regressor/bb_regressor.h

#pragma once

#include <torch/torch.h>
#include <string>
#include <vector>
#include <filesystem>

namespace fs = std::filesystem;

// Forward declaration of PrRoIPool2D
class PrRoIPool2D;

// Linear block for IoU prediction
class LinearBlock : public torch::nn::Module {
public:
    LinearBlock(int in_planes = 256, int out_planes = 256, int input_sz = 5, bool bias = true, 
                bool batch_norm = true, bool relu = true);
    
    torch::Tensor forward(torch::Tensor x);
    
    // Set to evaluation mode
    void eval() {
        linear->eval();
        if (use_bn) {
            bn->eval();
        }
        if (use_relu) {
            relu_->eval();
        }
    }
    
    // Move to device
    void to(torch::Device device) {
        linear->to(device);
        if (use_bn) bn->to(device);
        if (use_relu) relu_->to(device);
    }
    
    // Public members for direct access to weights
    torch::nn::Linear linear{nullptr};
    torch::nn::BatchNorm2d bn{nullptr};
    torch::nn::ReLU relu_{nullptr};
    bool use_bn;
    bool use_relu;
};

// PrRoIPool2D implementation
class PrRoIPool2D {
public:
    PrRoIPool2D(int pooled_height, int pooled_width, float spatial_scale);
    torch::Tensor forward(torch::Tensor feat, torch::Tensor rois);

    // CPU-based fallback implementation
    torch::Tensor forward_cpu(torch::Tensor feat, torch::Tensor rois) {
        // Simple implementation that returns zeros (for fallback only)
        int channels = feat.size(1);
        int num_rois = rois.size(0);
        return torch::zeros({num_rois, channels, pooled_height_, pooled_width_}, feat.options());
    }
    
private:
    int pooled_height_;
    int pooled_width_;
    float spatial_scale_;
};

// BBRegressor class
class BBRegressor {
public:
    // Statistics structure for tensors
    struct TensorStats {
        std::vector<int64_t> shape;
        float mean;
        float std_dev;
        float min_val;
        float max_val;
        float sum;
        std::vector<float> samples;
    };
    
    // Constructor with base directory and device specification
    BBRegressor(const std::string& base_dir, torch::Device device = torch::kCUDA);
    
    // Set model to evaluation mode
    void eval();
    
    // Get IoU features from backbone features
    std::vector<torch::Tensor> get_iou_feat(std::vector<torch::Tensor> feat);
    
    // Get modulation vectors for target
    std::vector<torch::Tensor> get_modulation(std::vector<torch::Tensor> feat, torch::Tensor bb);
    
    // Predict IoU for proposals
    torch::Tensor predict_iou(std::vector<torch::Tensor> modulation, 
                           std::vector<torch::Tensor> feat, 
                           torch::Tensor proposals);
    
    // Move model to device
    void to(torch::Device device);
    
    // Print model information
    void print_model_info();
    
    // Compute statistics for a tensor
    TensorStats compute_stats(const torch::Tensor& tensor);
    
    // Save tensor statistics to a file
    void save_stats(const std::vector<TensorStats>& all_stats, const std::string& filepath);

private:
    // Helper functions
    torch::nn::Sequential create_conv_block(int in_planes, int out_planes, int kernel_size,
                                      int stride, int padding, int dilation);
    void verify_batchnorm_dimensions();
    std::vector<char> read_file_to_bytes(const std::string& file_path);
    torch::Tensor load_tensor(const std::string& file_path);
    void load_weights();

    // Model state
    torch::Device device;
    std::string model_dir;
    
    // Convolution blocks
    torch::nn::Sequential conv3_1r{nullptr};
    torch::nn::Sequential conv3_1t{nullptr};
    torch::nn::Sequential conv3_2t{nullptr};
    torch::nn::Sequential fc3_1r{nullptr};
    torch::nn::Sequential conv4_1r{nullptr};
    torch::nn::Sequential conv4_1t{nullptr};
    torch::nn::Sequential conv4_2t{nullptr};
    torch::nn::Sequential fc34_3r{nullptr};
    torch::nn::Sequential fc34_4r{nullptr};
    
    // Pooling layers
    std::shared_ptr<PrRoIPool2D> prroi_pool3r;
    std::shared_ptr<PrRoIPool2D> prroi_pool3t;
    std::shared_ptr<PrRoIPool2D> prroi_pool4r;
    std::shared_ptr<PrRoIPool2D> prroi_pool4t;
    
    // Linear blocks
    LinearBlock fc3_rt;
    LinearBlock fc4_rt;
    
    // IoU predictor
    torch::nn::Linear iou_predictor{nullptr};
};