Function sam3_extract_obj_ptr_cpu

sam3.cpp:9348–9420

Extract object pointer from SAM output token via 3-layer MLP (CPU-side).

// Extract object pointer from SAM output token via 3-layer MLP (CPU-side).
static void sam3_extract_obj_ptr_cpu(
    const sam3_model& model,
    const float* sam_token_data,  // [D]
    float obj_score,
    float* out_ptr)  // [D]
{
    const auto& hp = model.hparams;
    const int D = hp.neck_dim;

    // SAM2/EdgeTAM with fixed_no_obj_ptr: blend projected ptr with no_obj_ptr
    // based on presence score λ.
    // SAM3 / SAM2 without fixed_no_obj_ptr: binary threshold.
    if ((hp.is_sam2() || hp.is_edgetam()) && hp.fixed_no_obj_ptr) {
        // λ = (obj_score > 0) ? 1.0 : 0.0  (hard threshold, not sigmoid)
        float lambda = (obj_score > 0.0f) ? 1.0f : 0.0f;

        // Project token through MLP
        std::vector<float> h(D), tmp(D);
        std::copy(sam_token_data, sam_token_data + D, h.data());
        for (int j = 0; j < 3; ++j) {
            auto* w = model.obj_ptr_proj_w[j];
            auto* b = model.obj_ptr_proj_b[j];
            int nel_w = (int)(w->ne[0] * w->ne[1]);
            std::vector<float> w_data(nel_w), b_data(D);
            sam3_read_f32(w, w_data.data(), nel_w);
            sam3_read_f32(b, b_data.data(), D);
            for (int o = 0; o < D; ++o) {
                float sum = b_data[o];
                for (int i = 0; i < D; ++i) sum += w_data[o * D + i] * h[i];
                tmp[o] = (j < 2) ? std::max(0.0f, sum) : sum;
            }
            std::swap(h, tmp);
        }

        // Blend: obj_ptr = λ * projected + (1-λ) * no_obj_ptr
        std::vector<float> no_ptr(D);
        ggml_backend_tensor_get(model.no_obj_ptr, no_ptr.data(), 0, D * sizeof(float));
        for (int i = 0; i < D; ++i)
            out_ptr[i] = lambda * h[i] + (1.0f - lambda) * no_ptr[i];
        return;
    }

    // SAM3 / default path: binary threshold
    if (obj_score <= 0.0f) {
        ggml_backend_tensor_get(model.no_obj_ptr, out_ptr, 0, D * sizeof(float));
        return;
    }

    std::vector<float> h(D), tmp(D);
    std::copy(sam_token_data, sam_token_data + D, h.data());

    for (int j = 0; j < 3; ++j) {
        auto* w = model.obj_ptr_proj_w[j];
        auto* b = model.obj_ptr_proj_b[j];

        int nel_w = (int)(w->ne[0] * w->ne[1]);
        std::vector<float> w_data(nel_w);
        sam3_read_f32(w, w_data.data(), nel_w);