hyperverse 0.0.1__py3-none-any.whl

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/avg_pool2d/operations_generic.h

@@ -0,0 +1,186 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_AVG_POOL2D_OPERATIONS_GENERIC_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_AVG_POOL2D_OPERATIONS_GENERIC_H
+#include "layer.h"
+#include "../../../rl_tools.h"
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools{
+    // ======================== malloc / free ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::avg_pool2d::LayerForward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::avg_pool2d::LayerForward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer) {
+        malloc(device, layer.output);
+    }
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer) {
+        free(device, layer.output);
+    }
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::avg_pool2d::State&) {}
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::avg_pool2d::State&) {}
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::avg_pool2d::Buffer&) {}
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::avg_pool2d::Buffer&) {}
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE&, TARGET_DEVICE&, nn::layers::avg_pool2d::State&, nn::layers::avg_pool2d::State&) {}
+    template<typename SPEC, typename DEVICE, typename RNG, typename MODE>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset(DEVICE&, const nn::layers::avg_pool2d::LayerForward<SPEC>&, nn::layers::avg_pool2d::State&, RNG&, Mode<MODE>) {}
+
+    // ======================== init_weights (no-op) ========================
+    template<typename DEVICE, typename SPEC, typename RNG>
+    RL_TOOLS_FUNCTION_PLACEMENT void init_weights(DEVICE& device, nn::layers::avg_pool2d::LayerForward<SPEC>& layer, RNG& rng) {}
+
+    // ======================== evaluate ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void evaluate(DEVICE& device, const nn::layers::avg_pool2d::LayerForward<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<OUTPUT_SPEC>& output, nn::layers::avg_pool2d::Buffer&, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        static_assert(nn::layers::avg_pool2d::check_input_output<LAYER_SPEC, INPUT_SPEC, OUTPUT_SPEC>);
+        using TI = typename DEVICE::index_t;
+        using T = typename OUTPUT_SPEC::T;
+        constexpr TI IH = LAYER_SPEC::INPUT_HEIGHT;
+        constexpr TI IW = LAYER_SPEC::INPUT_WIDTH;
+        constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+        constexpr TI BATCH_SIZE = product(typename INPUT_SPEC::SHAPE{}) / (IH * IW * C);
+        using INTERNAL_INPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, IH, IW, C>;
+        using INTERNAL_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, C>;
+        auto input_4d = view_memory<INTERNAL_INPUT_SHAPE>(device, input);
+        auto output_2d = view_memory<INTERNAL_OUTPUT_SHAPE>(device, output);
+        constexpr T scale = (T)1 / (T)(IH * IW);
+        for(TI bi = 0; bi < BATCH_SIZE; bi++){
+            for(TI c_i = 0; c_i < C; c_i++){
+                T acc = 0;
+                for(TI h = 0; h < IH; h++){
+                    for(TI w = 0; w < IW; w++){
+                        acc += get(device, input_4d, bi, h, w, c_i);
+                    }
+                }
+                set(device, output_2d, acc * scale, bi, c_i);
+            }
+        }
+    }
+
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void evaluate_step(DEVICE& device, const nn::layers::avg_pool2d::LayerForward<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, nn::layers::avg_pool2d::State& state, Tensor<OUTPUT_SPEC>& output, nn::layers::avg_pool2d::Buffer& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        evaluate(device, layer, input, output, buffer, rng, mode);
+    }
+
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<OUTPUT_SPEC>& output, nn::layers::avg_pool2d::Buffer& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        evaluate(device, static_cast<const nn::layers::avg_pool2d::LayerForward<LAYER_SPEC>&>(layer), input, output, buffer, rng, mode);
+    }
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, nn::layers::avg_pool2d::Buffer& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        forward(device, static_cast<nn::layers::avg_pool2d::LayerBackward<LAYER_SPEC>&>(layer), input, layer.output, buffer, rng, mode);
+    }
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<OUTPUT_SPEC>& output, nn::layers::avg_pool2d::Buffer& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        forward(device, layer, input, buffer, rng, mode);
+        copy(device, device, layer.output, output);
+    }
+
+    // ======================== backward ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename D_OUTPUT_SPEC, typename D_INPUT_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward_input(DEVICE& device, const nn::layers::avg_pool2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_INPUT_SPEC>& d_input, nn::layers::avg_pool2d::Buffer&, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        static_assert(nn::layers::avg_pool2d::check_input_output<LAYER_SPEC, D_INPUT_SPEC, D_OUTPUT_SPEC>);
+        using TI = typename DEVICE::index_t;
+        using T = typename D_OUTPUT_SPEC::T;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        constexpr TI IH = LAYER_SPEC::INPUT_HEIGHT;
+        constexpr TI IW = LAYER_SPEC::INPUT_WIDTH;
+        constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+        using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, C>;
+        using INTERNAL_D_INPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, IH, IW, C>;
+        auto d_output_2d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+        auto d_input_4d = view_memory<INTERNAL_D_INPUT_SHAPE>(device, d_input);
+        constexpr T scale = (T)1 / (T)(IH * IW);
+        for(TI bi = 0; bi < BATCH_SIZE; bi++){
+            for(TI c_i = 0; c_i < C; c_i++){
+                T grad = get(device, d_output_2d, bi, c_i) * scale;
+                for(TI h = 0; h < IH; h++){
+                    for(TI w = 0; w < IW; w++){
+                        set(device, d_input_4d, grad, bi, h, w, c_i);
+                    }
+                }
+            }
+        }
+    }
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename D_OUTPUT_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<D_OUTPUT_SPEC>& d_output, nn::layers::avg_pool2d::Buffer&, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        // No parameters to accumulate gradients for
+    }
+    template<typename DEVICE, typename LAYER_SPEC, typename INPUT_SPEC, typename D_OUTPUT_SPEC, typename D_INPUT_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward_full(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_INPUT_SPEC>& d_input, nn::layers::avg_pool2d::Buffer& buffer, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        backward_input(device, static_cast<const nn::layers::avg_pool2d::LayerBackward<LAYER_SPEC>&>(layer), d_output, d_input, buffer, mode);
+    }
+
+    // ======================== zero_gradient / update / _reset_optimizer_state (no-ops) ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void zero_gradient(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer) {}
+    template<typename DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void add_gradient(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::avg_pool2d::LayerGradient<TARGET_SPEC>& target) {}
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy_gradient(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::avg_pool2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::avg_pool2d::LayerGradient<TARGET_SPEC>& target) {}
+    template<typename DEVICE, typename SPEC, typename OPTIMIZER>
+    RL_TOOLS_FUNCTION_PLACEMENT void update(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer, OPTIMIZER& optimizer) {}
+    template<typename DEVICE, typename SPEC, typename OPTIMIZER>
+    RL_TOOLS_FUNCTION_PLACEMENT void _reset_optimizer_state(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer, OPTIMIZER& optimizer) {}
+
+    // ======================== copy ========================
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::avg_pool2d::LayerForward<SOURCE_SPEC>& source, nn::layers::avg_pool2d::LayerForward<TARGET_SPEC>& target) {
+        static_assert(nn::layers::avg_pool2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+    }
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::avg_pool2d::LayerBackward<SOURCE_SPEC>& source, nn::layers::avg_pool2d::LayerBackward<TARGET_SPEC>& target) {
+        static_assert(nn::layers::avg_pool2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+    }
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::avg_pool2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::avg_pool2d::LayerGradient<TARGET_SPEC>& target) {
+        static_assert(nn::layers::avg_pool2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+        copy(source_device, target_device, source.output, target.output);
+    }
+
+    // ======================== abs_diff ========================
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::avg_pool2d::LayerForward<S1>&, const nn::layers::avg_pool2d::LayerForward<S2>&) { return 0; }
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::avg_pool2d::LayerBackward<S1>&, const nn::layers::avg_pool2d::LayerBackward<S2>&) { return 0; }
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::avg_pool2d::LayerGradient<S1>& l1, const nn::layers::avg_pool2d::LayerGradient<S2>& l2) { return abs_diff(device, l1.output, l2.output); }
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT auto abs_diff(DEVICE& device, const nn::layers::avg_pool2d::State&, const nn::layers::avg_pool2d::State&) { return 0; }
+
+    // ======================== reset_forward_state ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset_forward_state(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<SPEC>&) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset_forward_state(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& l) { set_all(device, l.output, 0); }
+
+    // ======================== is_nan ========================
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE&, const nn::layers::avg_pool2d::LayerForward<SPEC>&, const Mode<MODE>& = Mode<mode::Default<>>{}) { return false; }
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE&, const nn::layers::avg_pool2d::LayerBackward<SPEC>&, const Mode<MODE>& = Mode<mode::Default<>>{}) { return false; }
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE& device, const nn::layers::avg_pool2d::LayerGradient<SPEC>& l, const Mode<MODE>& mode = Mode<mode::Default<>>{}) { return is_nan(device, l.output, mode); }
+    template<typename DEVICE, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE&, nn::layers::avg_pool2d::State&, const Mode<MODE>& = Mode<mode::Default<>>{}) { return false; }
+
+    // ======================== output ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT auto output(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& l){
+        return view_memory<typename SPEC::OUTPUT_SHAPE>(device, l.output);
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/avg_pool2d/persist.h

@@ -0,0 +1,33 @@
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_AVG_POOL2D_PERSIST_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_AVG_POOL2D_PERSIST_H
+#include "../../../version.h"
+#include "layer.h"
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools {
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::avg_pool2d::LayerForward<SPEC>& layer, GROUP& group) {
+        set_attribute(device, group, "type", "avg_pool2d");
+        write_attributes(device, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::avg_pool2d::LayerForward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::avg_pool2d::LayerBackward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::avg_pool2d::LayerForward<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::avg_pool2d::LayerBackward<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::avg_pool2d::LayerGradient<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename GROUP>
+    void save(DEVICE& device, nn::layers::avg_pool2d::State& state, GROUP& group) {}
+    template<typename DEVICE, typename GROUP>
+    bool load(DEVICE& device, nn::layers::avg_pool2d::State& state, GROUP& group) { return true; }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/conv2d/layer.h

@@ -0,0 +1,300 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_CONV2D_LAYER_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_CONV2D_LAYER_H
+#include "../../../nn/activation_functions.h"
+#include "../../../utils/generic/typing.h"
+#include "../../../containers/tensor/tensor.h"
+
+#include "../../../nn/capability/capability.h"
+#include "../../../nn/parameters/parameters.h"
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools::nn::layers::conv2d {
+
+    enum class Normalization { NONE, BATCH_NORM, LAYER_NORM };
+
+    template <typename T_TYPE_POLICY, typename T_TI>
+    struct KaimingUniformSpecification{
+        using T = typename T_TYPE_POLICY::DEFAULT;
+        using TI = T_TI;
+        static constexpr bool INIT_LEGACY = true;
+        static constexpr T SCALE = 1;
+    };
+    template<typename SPEC>
+    struct KaimingUniform {
+    };
+    template<typename T_TYPE_POLICY, typename T_TI>
+    using DefaultInitializer = KaimingUniform<KaimingUniformSpecification<T_TYPE_POLICY, T_TI>>;
+
+
+    template<typename T_TYPE_POLICY, typename T_TI,
+             T_TI T_OUTPUT_CHANNELS,
+             T_TI T_KERNEL_HEIGHT, T_TI T_KERNEL_WIDTH = T_KERNEL_HEIGHT,
+             T_TI T_STRIDE_H = 1, T_TI T_STRIDE_W = T_STRIDE_H,
+             T_TI T_PADDING_H = 0, T_TI T_PADDING_W = T_PADDING_H,
+             nn::activation_functions::ActivationFunction T_ACTIVATION_FUNCTION = nn::activation_functions::ActivationFunction::IDENTITY,
+             Normalization T_NORMALIZATION = Normalization::NONE,
+             typename T_INITIALIZER = DefaultInitializer<T_TYPE_POLICY, T_TI>,
+             typename T_PARAMETER_GROUP = parameters::groups::Normal>
+    struct Configuration{
+        using TYPE_POLICY = T_TYPE_POLICY;
+        using TI = T_TI;
+        static constexpr TI OUTPUT_CHANNELS = T_OUTPUT_CHANNELS;
+        static constexpr TI KERNEL_HEIGHT = T_KERNEL_HEIGHT;
+        static constexpr TI KERNEL_WIDTH = T_KERNEL_WIDTH;
+        static constexpr TI STRIDE_H = T_STRIDE_H;
+        static constexpr TI STRIDE_W = T_STRIDE_W;
+        static constexpr TI PADDING_H = T_PADDING_H;
+        static constexpr TI PADDING_W = T_PADDING_W;
+        static constexpr nn::activation_functions::ActivationFunction ACTIVATION_FUNCTION = T_ACTIVATION_FUNCTION;
+        static constexpr Normalization NORMALIZATION = T_NORMALIZATION;
+        using INITIALIZER = T_INITIALIZER;
+        using PARAMETER_GROUP = T_PARAMETER_GROUP;
+        static constexpr float NORM_EPSILON = 1e-5f;
+        static constexpr float BN_MOMENTUM = 0.1f;
+    };
+
+    // ======================== Normalization forward storage (learnable parameters + running stats) ========================
+    template<Normalization NORM, typename SPEC>
+    struct NormForward {};
+
+    template<typename SPEC>
+    struct NormForward<Normalization::BATCH_NORM, SPEC> {
+        using TYPE_POLICY = typename SPEC::TYPE_POLICY;
+        using TI = typename SPEC::TI;
+        using PARAM_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS>;
+
+        using GAMMA_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, PARAM_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Weights, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST, numeric_types::categories::NormParameter>;
+        typename SPEC::PARAMETER_TYPE::template Instance<GAMMA_PARAMETER_SPEC> gamma;
+
+        using BETA_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, PARAM_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Biases, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST, numeric_types::categories::NormParameter>;
+        typename SPEC::PARAMETER_TYPE::template Instance<BETA_PARAMETER_SPEC> beta;
+
+        // Running statistics (non-learnable, plain tensors)
+        using T_NORM_STAT = typename TYPE_POLICY::template GET<numeric_types::categories::NormStatistics>;
+        using STAT_TENSOR_SPEC = tensor::Specification<T_NORM_STAT, TI, PARAM_SHAPE, SPEC::DYNAMIC_ALLOCATION>;
+        Tensor<STAT_TENSOR_SPEC> running_mean;
+        Tensor<STAT_TENSOR_SPEC> running_var;
+    };
+
+    template<typename SPEC>
+    struct NormForward<Normalization::LAYER_NORM, SPEC> {
+        using TYPE_POLICY = typename SPEC::TYPE_POLICY;
+        using TI = typename SPEC::TI;
+        using PARAM_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS>;
+
+        using GAMMA_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, PARAM_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Weights, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST, numeric_types::categories::NormParameter>;
+        typename SPEC::PARAMETER_TYPE::template Instance<GAMMA_PARAMETER_SPEC> gamma;
+
+        using BETA_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, PARAM_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Biases, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST, numeric_types::categories::NormParameter>;
+        typename SPEC::PARAMETER_TYPE::template Instance<BETA_PARAMETER_SPEC> beta;
+    };
+
+    // ======================== Normalization backward storage (cached statistics) ========================
+    template<Normalization NORM, typename SPEC>
+    struct NormBackward {};
+
+    template<typename SPEC>
+    struct NormBackward<Normalization::BATCH_NORM, SPEC> {
+        using T = typename SPEC::TYPE_POLICY::template GET<numeric_types::categories::Activation>;
+        using TI = typename SPEC::TI;
+        using STAT_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS>;
+        using STAT_SPEC = tensor::Specification<T, TI, STAT_SHAPE, SPEC::DYNAMIC_ALLOCATION>;
+        Tensor<STAT_SPEC> mean;
+        Tensor<STAT_SPEC> inv_std;
+    };
+
+    template<typename SPEC>
+    struct NormBackward<Normalization::LAYER_NORM, SPEC> {
+        using T = typename SPEC::TYPE_POLICY::template GET<numeric_types::categories::Activation>;
+        using TI = typename SPEC::TI;
+        using STAT_SHAPE = tensor::Shape<TI, SPEC::INTERNAL_BATCH_SIZE>;
+        using STAT_SPEC = tensor::Specification<T, TI, STAT_SHAPE, SPEC::DYNAMIC_ALLOCATION>;
+        Tensor<STAT_SPEC> mean;
+        Tensor<STAT_SPEC> inv_std;
+    };
+
+    // ======================== Specification ========================
+    template <typename T_CONFIG, typename T_CAPABILITY, typename T_INPUT_SHAPE>
+    struct Specification: T_CAPABILITY, T_CONFIG{
+        using CONFIG = T_CONFIG;
+        using TYPE_POLICY = typename CONFIG::TYPE_POLICY;
+        using TI = typename CONFIG::TI;
+        using CAPABILITY = T_CAPABILITY;
+        using INPUT_SHAPE = T_INPUT_SHAPE;
+        static_assert(length(INPUT_SHAPE{}) >= 4, "Conv2d input shape must have at least 4 dimensions (...BATCH x H x W x C)");
+        static constexpr TI INPUT_HEIGHT = get<length(INPUT_SHAPE{})-3>(INPUT_SHAPE{});
+        static constexpr TI INPUT_WIDTH = get<length(INPUT_SHAPE{})-2>(INPUT_SHAPE{});
+        static constexpr TI INPUT_CHANNELS = get_last(INPUT_SHAPE{});
+        static constexpr TI OUTPUT_HEIGHT = (INPUT_HEIGHT + 2 * CONFIG::PADDING_H - CONFIG::KERNEL_HEIGHT) / CONFIG::STRIDE_H + 1;
+        static constexpr TI OUTPUT_WIDTH = (INPUT_WIDTH + 2 * CONFIG::PADDING_W - CONFIG::KERNEL_WIDTH) / CONFIG::STRIDE_W + 1;
+
+        template <typename NEW_INPUT_SHAPE>
+        struct OUTPUT_SHAPE_FACTORY{
+            static_assert(length(NEW_INPUT_SHAPE{}) >= 4);
+            static constexpr TI NEW_H = get<length(NEW_INPUT_SHAPE{})-3>(NEW_INPUT_SHAPE{});
+            static constexpr TI NEW_W = get<length(NEW_INPUT_SHAPE{})-2>(NEW_INPUT_SHAPE{});
+            static constexpr TI NEW_C = get_last(NEW_INPUT_SHAPE{});
+            static_assert(NEW_H == INPUT_HEIGHT);
+            static_assert(NEW_W == INPUT_WIDTH);
+            static_assert(NEW_C == INPUT_CHANNELS);
+            static constexpr TI NEW_OH = (NEW_H + 2 * CONFIG::PADDING_H - CONFIG::KERNEL_HEIGHT) / CONFIG::STRIDE_H + 1;
+            static constexpr TI NEW_OW = (NEW_W + 2 * CONFIG::PADDING_W - CONFIG::KERNEL_WIDTH) / CONFIG::STRIDE_W + 1;
+            using SHAPE_HEIGHT = tensor::Replace<NEW_INPUT_SHAPE, NEW_OH, length(NEW_INPUT_SHAPE{}) - 3>;
+            using SHAPE_WIDTH = tensor::Replace<SHAPE_HEIGHT, NEW_OW, length(NEW_INPUT_SHAPE{}) - 2>;
+            using SHAPE = tensor::Replace<SHAPE_WIDTH, CONFIG::OUTPUT_CHANNELS, length(NEW_INPUT_SHAPE{}) - 1>;
+        };
+        using OUTPUT_SHAPE = typename OUTPUT_SHAPE_FACTORY<INPUT_SHAPE>::SHAPE;
+        static constexpr TI INTERNAL_BATCH_SIZE = tensor::shape_math::leading_product(tensor::shape_math::element_to_array<INPUT_SHAPE>(), 3);
+        static constexpr TI NUM_WEIGHTS = CONFIG::OUTPUT_CHANNELS * INPUT_CHANNELS * CONFIG::KERNEL_HEIGHT * CONFIG::KERNEL_WIDTH + CONFIG::OUTPUT_CHANNELS
+            + (CONFIG::NORMALIZATION != Normalization::NONE ? 2 * CONFIG::OUTPUT_CHANNELS : 0);
+    };
+
+    template<typename SPEC_1, typename SPEC_2>
+    constexpr bool check_spec_memory =
+            SPEC_1::INPUT_CHANNELS == SPEC_2::INPUT_CHANNELS
+            && SPEC_1::OUTPUT_CHANNELS == SPEC_2::OUTPUT_CHANNELS
+            && SPEC_1::KERNEL_HEIGHT == SPEC_2::KERNEL_HEIGHT
+            && SPEC_1::KERNEL_WIDTH == SPEC_2::KERNEL_WIDTH
+            && SPEC_1::INPUT_HEIGHT == SPEC_2::INPUT_HEIGHT
+            && SPEC_1::INPUT_WIDTH == SPEC_2::INPUT_WIDTH
+            && SPEC_1::NORMALIZATION == SPEC_2::NORMALIZATION;
+
+    template<typename SPEC_1, typename SPEC_2>
+    constexpr bool check_spec =
+        check_spec_memory<SPEC_1, SPEC_2>
+        && SPEC_1::ACTIVATION_FUNCTION == SPEC_2::ACTIVATION_FUNCTION
+        && SPEC_1::STRIDE_H == SPEC_2::STRIDE_H
+        && SPEC_1::STRIDE_W == SPEC_2::STRIDE_W
+        && SPEC_1::PADDING_H == SPEC_2::PADDING_H
+        && SPEC_1::PADDING_W == SPEC_2::PADDING_W;
+
+    template <typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC>
+    constexpr bool check_input_output =
+        length(typename INPUT_SPEC::SHAPE{}) >= 4 &&
+        length(typename OUTPUT_SPEC::SHAPE{}) >= 4 &&
+        get<length(typename INPUT_SPEC::SHAPE{})-1>(typename INPUT_SPEC::SHAPE{}) == LAYER_SPEC::INPUT_CHANNELS &&
+        get<length(typename INPUT_SPEC::SHAPE{})-2>(typename INPUT_SPEC::SHAPE{}) == LAYER_SPEC::INPUT_WIDTH &&
+        get<length(typename INPUT_SPEC::SHAPE{})-3>(typename INPUT_SPEC::SHAPE{}) == LAYER_SPEC::INPUT_HEIGHT &&
+        get<length(typename OUTPUT_SPEC::SHAPE{})-1>(typename OUTPUT_SPEC::SHAPE{}) == LAYER_SPEC::OUTPUT_CHANNELS &&
+        get<length(typename OUTPUT_SPEC::SHAPE{})-2>(typename OUTPUT_SPEC::SHAPE{}) == LAYER_SPEC::OUTPUT_WIDTH &&
+        get<length(typename OUTPUT_SPEC::SHAPE{})-3>(typename OUTPUT_SPEC::SHAPE{}) == LAYER_SPEC::OUTPUT_HEIGHT;
+
+    struct State{};
+
+    template <typename T_SPEC, bool T_DYNAMIC_ALLOCATION>
+    struct BufferSpecification {
+        using SPEC = T_SPEC;
+        static constexpr bool DYNAMIC_ALLOCATION = T_DYNAMIC_ALLOCATION;
+    };
+    template<typename T_SPEC>
+    struct Buffer{
+        using SPEC = typename T_SPEC::SPEC;
+        static constexpr bool DYNAMIC_ALLOCATION = T_SPEC::DYNAMIC_ALLOCATION;
+        using ACCUMULATOR_TYPE = typename SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        using TI = typename SPEC::TI;
+        using D_WEIGHTS_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS, SPEC::KERNEL_HEIGHT, SPEC::KERNEL_WIDTH, SPEC::INPUT_CHANNELS>;
+        using D_WEIGHTS_SPEC = tensor::Specification<ACCUMULATOR_TYPE, TI, D_WEIGHTS_SHAPE, DYNAMIC_ALLOCATION>;
+        Tensor<D_WEIGHTS_SPEC> d_weights_acc;
+        using D_BIASES_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS>;
+        using D_BIASES_SPEC = tensor::Specification<ACCUMULATOR_TYPE, TI, D_BIASES_SHAPE, DYNAMIC_ALLOCATION>;
+        Tensor<D_BIASES_SPEC> d_biases_acc;
+        using D_INPUT_SHAPE = tensor::Shape<TI, SPEC::INTERNAL_BATCH_SIZE, SPEC::INPUT_HEIGHT, SPEC::INPUT_WIDTH, SPEC::INPUT_CHANNELS>;
+        using D_INPUT_SPEC = tensor::Specification<ACCUMULATOR_TYPE, TI, D_INPUT_SHAPE, DYNAMIC_ALLOCATION>;
+        Tensor<D_INPUT_SPEC> d_input_acc;
+        using D_NORM_SHAPE = tensor::Shape<TI, SPEC::OUTPUT_CHANNELS>;
+        using D_NORM_SPEC = tensor::Specification<ACCUMULATOR_TYPE, TI, D_NORM_SHAPE, DYNAMIC_ALLOCATION>;
+        Tensor<D_NORM_SPEC> d_gamma_acc;
+        Tensor<D_NORM_SPEC> d_beta_acc;
+        void* backend_state = nullptr;
+    };
+
+    template<typename T_SPEC>
+    struct LayerForward {
+        using SPEC = T_SPEC;
+        using TYPE_POLICY = typename SPEC::TYPE_POLICY;
+        using TI = typename SPEC::TI;
+        static constexpr TI INPUT_HEIGHT = SPEC::INPUT_HEIGHT;
+        static constexpr TI INPUT_WIDTH = SPEC::INPUT_WIDTH;
+        static constexpr TI INPUT_CHANNELS = SPEC::INPUT_CHANNELS;
+        static constexpr TI OUTPUT_HEIGHT = SPEC::OUTPUT_HEIGHT;
+        static constexpr TI OUTPUT_WIDTH = SPEC::OUTPUT_WIDTH;
+        static constexpr TI OUTPUT_CHANNELS = SPEC::OUTPUT_CHANNELS;
+        static constexpr TI KERNEL_HEIGHT = SPEC::KERNEL_HEIGHT;
+        static constexpr TI KERNEL_WIDTH = SPEC::KERNEL_WIDTH;
+        static constexpr TI STRIDE_H = SPEC::STRIDE_H;
+        static constexpr TI STRIDE_W = SPEC::STRIDE_W;
+        static constexpr TI PADDING_H = SPEC::PADDING_H;
+        static constexpr TI PADDING_W = SPEC::PADDING_W;
+        static constexpr TI NUM_WEIGHTS = SPEC::NUM_WEIGHTS;
+        static constexpr TI INTERNAL_BATCH_SIZE = SPEC::INTERNAL_BATCH_SIZE;
+        static constexpr Normalization NORMALIZATION = SPEC::NORMALIZATION;
+        using INPUT_SHAPE = typename SPEC::INPUT_SHAPE;
+        template <typename NEW_INPUT_SHAPE>
+        using OUTPUT_SHAPE_FACTORY = typename SPEC::template OUTPUT_SHAPE_FACTORY<NEW_INPUT_SHAPE>::SHAPE;
+        using OUTPUT_SHAPE = typename SPEC::OUTPUT_SHAPE;
+
+        // Weights: [OUTPUT_CHANNELS, KERNEL_HEIGHT, KERNEL_WIDTH, INPUT_CHANNELS] (NHWC filter layout)
+        using WEIGHTS_SHAPE = tensor::Shape<TI, OUTPUT_CHANNELS, KERNEL_HEIGHT, KERNEL_WIDTH, INPUT_CHANNELS>;
+        using WEIGHTS_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, WEIGHTS_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Weights, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST>;
+        typename SPEC::PARAMETER_TYPE::template Instance<WEIGHTS_PARAMETER_SPEC> weights;
+
+        // Biases: [OUTPUT_CHANNELS]
+        using BIASES_SHAPE = tensor::Shape<TI, OUTPUT_CHANNELS>;
+        using BIASES_PARAMETER_SPEC = typename SPEC::PARAMETER_TYPE::template Specification<TYPE_POLICY, TI, BIASES_SHAPE, typename SPEC::PARAMETER_GROUP, nn::parameters::categories::Biases, SPEC::DYNAMIC_ALLOCATION, SPEC::CONST>;
+        typename SPEC::PARAMETER_TYPE::template Instance<BIASES_PARAMETER_SPEC> biases;
+
+        // Normalization parameters (conditionally populated)
+        NormForward<SPEC::NORMALIZATION, SPEC> norm;
+
+        template<bool DYNAMIC_ALLOCATION=true>
+        using Buffer = Buffer<BufferSpecification<SPEC, DYNAMIC_ALLOCATION>>;
+        template<bool DYNAMIC_ALLOCATION=true>
+        using State = State;
+    };
+
+    template<typename SPEC>
+    struct LayerBackward: public LayerForward<SPEC>{
+        using PARENT = LayerForward<SPEC>;
+        using T = typename SPEC::TYPE_POLICY::template GET<numeric_types::categories::Activation>;
+        using TI = typename SPEC::TI;
+        // Pre-activations stores the conv output (before normalization and activation)
+        using PRE_ACTIVATIONS_SHAPE = tensor::Shape<TI, SPEC::INTERNAL_BATCH_SIZE, SPEC::OUTPUT_HEIGHT, SPEC::OUTPUT_WIDTH, SPEC::OUTPUT_CHANNELS>;
+        using PRE_ACTIVATIONS_SPEC = tensor::Specification<T, TI, PRE_ACTIVATIONS_SHAPE, SPEC::DYNAMIC_ALLOCATION, tensor::RowMajorStride<PRE_ACTIVATIONS_SHAPE>, SPEC::CONST>;
+        using PRE_ACTIVATIONS_TYPE = Tensor<PRE_ACTIVATIONS_SPEC>;
+        PRE_ACTIVATIONS_TYPE pre_activations;
+
+        // Cached normalization statistics (conditionally populated)
+        NormBackward<SPEC::NORMALIZATION, SPEC> norm_cache;
+    };
+
+    template<typename SPEC>
+    struct LayerGradient: public LayerBackward<SPEC>{
+        using PARENT = LayerBackward<SPEC>;
+        using T = typename SPEC::TYPE_POLICY::template GET<numeric_types::categories::Activation>;
+        using TI = typename SPEC::TI;
+        using OUTPUT_CONTAINER_SHAPE = tensor::Shape<TI, SPEC::INTERNAL_BATCH_SIZE, SPEC::OUTPUT_HEIGHT, SPEC::OUTPUT_WIDTH, SPEC::OUTPUT_CHANNELS>;
+        using OUTPUT_CONTAINER_SPEC = tensor::Specification<T, TI, OUTPUT_CONTAINER_SHAPE, SPEC::DYNAMIC_ALLOCATION, tensor::RowMajorStride<OUTPUT_CONTAINER_SHAPE>, SPEC::CONST>;
+        using OUTPUT_CONTAINER_TYPE = Tensor<OUTPUT_CONTAINER_SPEC>;
+        OUTPUT_CONTAINER_TYPE output;
+    };
+
+    template<typename CONFIG, typename CAPABILITY, typename INPUT_SHAPE>
+    using Layer =
+        typename utils::typing::conditional_t<CAPABILITY::TAG == nn::LayerCapability::Forward,
+            LayerForward<Specification<CONFIG, CAPABILITY, INPUT_SHAPE>>,
+        typename utils::typing::conditional_t<CAPABILITY::TAG == nn::LayerCapability::Backward,
+            LayerBackward<Specification<CONFIG, CAPABILITY, INPUT_SHAPE>>,
+        typename utils::typing::conditional_t<CAPABILITY::TAG == nn::LayerCapability::Gradient,
+            LayerGradient<Specification<CONFIG, CAPABILITY, INPUT_SHAPE>>, void>>>;
+
+    template <typename CONFIG>
+    struct BindConfiguration{
+        template <typename CAPABILITY, typename INPUT_SHAPE>
+        using Layer = nn::layers::conv2d::Layer<CONFIG, CAPABILITY, INPUT_SHAPE>;
+    };
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+
+#endif