hyperverse 0.0.1__py3-none-any.whl

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

@@ -0,0 +1,75 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_GRU_PERSIST_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_GRU_PERSIST_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::gru::LayerForward<SPEC>& layer, GROUP& group) {
+        // todo: forward implementation to Parameter struct
+        auto weights_input_group = create_group(device, group, "weights_input");
+        save(device, layer.weights_input, weights_input_group);
+        auto biases_input_group = create_group(device, group, "biases_input");
+        save(device, layer.biases_input, biases_input_group);
+        auto weights_hidden_group = create_group(device, group, "weights_hidden");
+        save(device, layer.weights_hidden, weights_hidden_group);
+        auto biases_hidden_group = create_group(device, group, "biases_hidden");
+        save(device, layer.biases_hidden, biases_hidden_group);
+        auto initial_hidden_state_group = create_group(device, group, "initial_hidden_state");
+        save(device, layer.initial_hidden_state, initial_hidden_state_group);
+        set_attribute(device, group, "type", "gru");
+        write_attributes(device, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::gru::LayerBackward<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::gru::LayerForward<SPEC>&)layer, group);
+        save(device, layer.post_activation, group, "post_activation");
+        save(device, layer.n_pre_pre_activation, group, "n_pre_pre_activation");
+        save(device, layer.output, group, "output");
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::gru::LayerGradient<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::gru::LayerBackward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::gru::LayerForward<SPEC>& layer, GROUP& group) {
+        auto weights_input_group = get_group(device, group, "weights_input");
+        bool success = load(device, layer.weights_input, weights_input_group);
+        auto biases_input_group = get_group(device, group, "biases_input");
+        success &= load(device, layer.biases_input, biases_input_group);
+        auto weights_hidden_group = get_group(device, group, "weights_hidden");
+        success &= load(device, layer.weights_hidden, weights_hidden_group);
+        auto biases_hidden_group = get_group(device, group, "biases_hidden");
+        success &= load(device, layer.biases_hidden, biases_hidden_group);
+        auto initial_hidden_state_group = get_group(device, group, "initial_hidden_state");
+        success &= load(device, layer.initial_hidden_state, initial_hidden_state_group);
+        return success;
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::gru::LayerBackward<SPEC>& layer, GROUP& group) {
+        bool success = load(device, (nn::layers::gru::LayerForward<SPEC>&)layer, group);
+        success &= load(device, layer.post_activation, group, "post_activation");
+        success &= load(device, layer.n_pre_pre_activation, group, "n_pre_pre_activation");
+        success &= load(device, layer.output, group, "output");
+        return success;
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::gru::LayerGradient<SPEC>& layer, GROUP& group) {
+        return load(device, (nn::layers::gru::LayerBackward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::gru::State<SPEC>& state, GROUP& group) {
+        save(device, state.state, group, "state");
+        save(device, state.step, group, "step");
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::gru::State<SPEC>& state, GROUP& group) {
+        bool success = load(device, state.state, group, "state");
+        success &= load(device, state.step, group, "step");
+        return success;
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/gru/persist_code.h

@@ -0,0 +1,199 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_GRU_PERSIST_CODE_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_GRU_PERSIST_CODE_H
+#include "layer.h"
+#include "../../../containers/matrix/persist_code.h"
+#include <sstream>
+#include "../../../persist/code.h"
+#include "../../../containers/matrix/persist_code.h"
+#include "../../../nn/capability/persist_code.h"
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools {
+    namespace nn::layers::gru::persist_code{
+        template<typename DEVICE, typename SPEC>
+        rl_tools::persist::Code finish(DEVICE& device, nn::layers::gru::LayerForward<SPEC> &layer, std::string name, rl_tools::persist::Code input, bool const_declaration=true, typename DEVICE::index_t indent=0){
+            using TI = typename DEVICE::index_t;
+            std::stringstream indent_ss;
+            for(TI i=0; i < indent; i++){
+                indent_ss << "    ";
+            }
+            std::string ind = indent_ss.str();
+            using TI = typename DEVICE::index_t;
+            std::stringstream ss, ss_header;
+            ss_header << input.header;
+            ss_header << "#include <rl_tools/nn/layers/gru/layer.h>\n";
+            ss << input.body;
+            std::string TI_string = containers::persist::get_type_string<typename SPEC::TI>();
+            ss << ind << "namespace " << name << " {\n";
+            ss << ind << "    using TYPE_POLICY = " << to_string(typename SPEC::TYPE_POLICY{}) << ";\n";
+            ss << ind << "    using CONFIG = " << "RL_TOOLS""_NAMESPACE_WRAPPER ::rl_tools::nn::layers::gru::Configuration<"
+               << "TYPE_POLICY, "
+               << TI_string << ", "
+               << SPEC::HIDDEN_DIM << ", "
+               << get_type_string_tag(device, typename SPEC::PARAMETER_GROUP{}) << ", "
+               << (SPEC::FAST_TANH ? "true" : "false")
+               << ">; \n";
+            ss << ind << "    " << "using TEMPLATE = RL_TOOLS""_NAMESPACE_WRAPPER ::rl_tools::nn::layers::gru::BindConfiguration<CONFIG>;" << "\n";
+            ss << ind << "    " << "using INPUT_SHAPE = RL_TOOLS""_NAMESPACE_WRAPPER ::rl_tools::tensor::Shape<" << TI_string << ", " << get<0>(typename SPEC::INPUT_SHAPE{}) << ", " << get<1>(typename SPEC::INPUT_SHAPE{}) << ", " << get<2>(typename SPEC::INPUT_SHAPE{}) << ">;\n";
+            ss << ind << "    " << "using CAPABILITY = " << to_string(typename SPEC::CAPABILITY::template CHANGE_PARAMETERS<true, true>{}) << ";" << "\n";
+            ss << ind << "    " << "using TYPE = RL_TOOLS""_NAMESPACE_WRAPPER ::rl_tools::nn::layers::gru::Layer<CONFIG, CAPABILITY, INPUT_SHAPE>;" << "\n";
+            std::string initializer_list;
+            std::string forward_members = "{weights_input::parameters, biases_input::parameters, weights_hidden::parameters, biases_hidden::parameters, initial_hidden_state::parameters}";
+            std::string backward_members = "{" + forward_members + ", post_activation::container, n_pre_pre_activation::container, output::container}";
+            if constexpr(SPEC::CAPABILITY::TAG == nn::LayerCapability::Forward){
+                initializer_list = forward_members;
+            }
+            else{
+                if constexpr(SPEC::CAPABILITY::TAG == nn::LayerCapability::Backward){
+                    initializer_list = backward_members;
+                }
+                else{
+                    if constexpr(SPEC::CAPABILITY::TAG == nn::LayerCapability::Gradient){
+                        initializer_list = "{" + backward_members + "}";
+                    }
+                    else{
+                        utils::assert_exit(device, false, "Unknown capability");
+                    }
+                }
+            }
+            ss << ind << "    " << (const_declaration ? "constexpr " : "") << "TYPE module = " << initializer_list << ";\n";
+            ss << ind << "    " << "template <typename T_TYPE = TYPE>" << "\n";
+            ss << ind << "    " << (const_declaration ? "constexpr " : "") << "T_TYPE factory = " << initializer_list << ";" << "\n";
+            ss << ind << "    " << "template <typename T_TYPE = TYPE>" << "\n";
+            ss << ind << "    " << (const_declaration ? "constexpr " : "") << "T_TYPE factory_function(){return " << initializer_list << ";" << "}\n";
+            ss << ind << "}\n";
+
+
+            return {ss_header.str(), ss.str()};
+        }
+    }
+    template<typename DEVICE, typename SPEC>
+    persist::Code save_code_split(DEVICE& device, nn::layers::gru::LayerForward<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0, bool finish=true){
+        using TI = typename DEVICE::index_t;
+        std::stringstream indent_ss;
+        for(TI i=0; i < indent; i++){
+            indent_ss << "    ";
+        }
+        std::string ind = indent_ss.str();
+        using TI = typename DEVICE::index_t;
+        std::stringstream ss, ss_header;
+        ss << ind << "namespace " << name << " {\n";
+        {
+            auto weights_input = save_code_split(device, layer.weights_input, "weights_input", const_declaration, indent+1);
+            ss_header << weights_input.header;
+            ss << weights_input.body;
+        }
+        {
+            auto weights_hidden = save_code_split(device, layer.weights_hidden, "weights_hidden", const_declaration, indent+1);
+            ss_header << weights_hidden.header;
+            ss << weights_hidden.body;
+        }
+        {
+            auto biases_input = save_code_split(device, layer.biases_input, "biases_input", const_declaration, indent+1);
+            ss_header << biases_input.header;
+            ss << biases_input.body;
+        }
+        {
+            auto biases_hidden = save_code_split(device, layer.biases_hidden, "biases_hidden", const_declaration, indent+1);
+            ss_header << biases_hidden.header;
+            ss << biases_hidden.body;
+        }
+        {
+            auto initial_hidden_state = save_code_split(device, layer.initial_hidden_state, "initial_hidden_state", const_declaration, indent+1);
+            ss_header << initial_hidden_state.header;
+            ss << initial_hidden_state.body;
+        }
+        ss << ind << "}\n";
+        if(finish){
+            return nn::layers::gru::persist_code::finish(device, layer, name, {ss_header.str(), ss.str()}, const_declaration, indent);
+        }
+        else{
+            return {ss_header.str(), ss.str()};
+        }
+    }
+    template<typename DEVICE, typename SPEC>
+    persist::Code save_code_split(DEVICE& device, nn::layers::gru::LayerBackward<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0, bool finish=true){
+        using TI = typename DEVICE::index_t;
+        std::stringstream indent_ss;
+        for(TI i=0; i < indent; i++){
+            indent_ss << "    ";
+        }
+        std::string ind = indent_ss.str();
+        using TI = typename DEVICE::index_t;
+        std::stringstream ss, ss_header;
+        auto previous = save_code_split(device, static_cast<nn::layers::gru::LayerForward<SPEC>&>(layer), name, const_declaration, indent, false);
+        ss_header << previous.header;
+        ss << previous.body;
+        ss << ind << "namespace " << name << " {\n";
+        auto post_activation = save_code_split(device, layer.post_activation, "post_activation", const_declaration, indent+1);
+        auto n_pre_pre_activation = save_code_split(device, layer.n_pre_pre_activation, "n_pre_pre_activation", const_declaration, indent+1);
+        auto output = save_code_split(device, layer.output, "output", const_declaration, indent+1);
+        ss_header << n_pre_pre_activation.header;
+        ss_header << post_activation.header;
+        ss_header << output.header;
+        ss << n_pre_pre_activation.body;
+        ss << post_activation.body;
+        ss << output.body;
+        ss << ind << "}\n";
+        if(finish){
+            return nn::layers::gru::persist_code::finish(device, layer, name, {ss_header.str(), ss.str()}, const_declaration, indent);
+        }
+        else{
+            return {ss_header.str(), ss.str()};
+        }
+    }
+
+    template<typename DEVICE, typename SPEC>
+    persist::Code save_code_split(DEVICE& device, nn::layers::gru::LayerGradient<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0){
+        using TI = typename DEVICE::index_t;
+        std::stringstream indent_ss;
+        for(TI i=0; i < indent; i++){
+            indent_ss << "    ";
+        }
+        std::string ind = indent_ss.str();
+        using TI = typename DEVICE::index_t;
+        std::stringstream ss, ss_header;
+        auto previous = save_code_split(device, static_cast<nn::layers::gru::LayerBackward<SPEC>&>(layer), name, const_declaration, indent, false);
+        ss_header << previous.header;
+        ss << previous.body;
+        ss << ind << "namespace " << name << " {\n";
+        ss << ind << "}\n";
+        return nn::layers::gru::persist_code::finish(device, layer, name, {ss_header.str(), ss.str()}, const_declaration, indent);
+    }
+
+    template<typename DEVICE, typename SPEC>
+    std::string save_code(DEVICE& device, nn::layers::gru::LayerForward<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0){
+        auto code = save_code_split(device, layer, name, const_declaration, indent);
+        return code.header + code.body;
+    }
+    template<typename DEVICE, typename SPEC>
+    std::string save_code(DEVICE& device, nn::layers::gru::LayerBackward<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0){
+        auto code = save_code_split(device, layer, name, const_declaration, indent);
+        return code.header + code.body;
+    }
+    template<typename DEVICE, typename SPEC>
+    std::string save_code(DEVICE& device, nn::layers::gru::LayerGradient<SPEC> &layer, std::string name, bool const_declaration=true, typename DEVICE::index_t indent=0){
+        auto code = save_code_split(device, layer, name, const_declaration, indent);
+        return code.header + code.body;
+    }
+    template <typename DEVICE, typename SPEC>
+    std::string nn_analytics(DEVICE& device, nn::layers::gru::LayerGradient<SPEC>& layer) {
+        std::string data;
+        data += "{";
+        data += "\"weights_input\": " + nn_analytics(device, layer.weights_input) + ", ";
+        data += "\"biases_input\": " + nn_analytics(device, layer.biases_input) + ", ";
+        data += "\"weights_hidden\": " + nn_analytics(device, layer.weights_hidden) + ", ";
+        data += "\"biases_hidden\": " + nn_analytics(device, layer.biases_hidden) + ", ";
+        data += "\"initial_hidden_state\": " + nn_analytics(device, layer.initial_hidden_state) + ", ";
+        data += "\"n_pre_pre_activation\": " + json(device, layer.n_pre_pre_activation) + ", ";
+        data += "\"post_activation\": " + json(device, layer.post_activation) + ", ";
+        data += "\"output\": " + json(device, layer.output);
+        data += "}";
+        return data;
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+
+#endif

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

@@ -0,0 +1,3 @@
+// #include "dense/layer.h"
+// #include "sample_and_squash/layer.h"
+// #include "standardize/layer.h"

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/max_pool2d/layer.h

@@ -0,0 +1,143 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_MAX_POOL2D_LAYER_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_MAX_POOL2D_LAYER_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::max_pool2d {
+
+    template<typename T_TYPE_POLICY, typename T_TI,
+             T_TI T_KERNEL_HEIGHT, T_TI T_KERNEL_WIDTH = T_KERNEL_HEIGHT,
+             T_TI T_STRIDE_H = T_KERNEL_HEIGHT, T_TI T_STRIDE_W = T_KERNEL_WIDTH,
+             T_TI T_PADDING_H = 0, T_TI T_PADDING_W = T_PADDING_H>
+    struct Configuration{
+        using TYPE_POLICY = T_TYPE_POLICY;
+        using TI = T_TI;
+        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;
+    };
+
+    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, "MaxPool2d 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;
+        static constexpr TI OUTPUT_CHANNELS = INPUT_CHANNELS;
+        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 = tensor::Replace<SHAPE_HEIGHT, NEW_OW, length(NEW_INPUT_SHAPE{}) - 2>;
+        };
+        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 = 0;
+    };
+
+    template<typename SPEC_1, typename SPEC_2>
+    constexpr bool check_spec_memory =
+        SPEC_1::INPUT_HEIGHT == SPEC_2::INPUT_HEIGHT
+        && SPEC_1::INPUT_WIDTH == SPEC_2::INPUT_WIDTH
+        && SPEC_1::INPUT_CHANNELS == SPEC_2::INPUT_CHANNELS
+        && SPEC_1::KERNEL_HEIGHT == SPEC_2::KERNEL_HEIGHT
+        && SPEC_1::KERNEL_WIDTH == SPEC_2::KERNEL_WIDTH;
+
+    template<typename SPEC_1, typename SPEC_2>
+    constexpr bool check_spec =
+        check_spec_memory<SPEC_1, SPEC_2>
+        && 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{};
+    struct Buffer{};
+
+    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 NUM_WEIGHTS = SPEC::NUM_WEIGHTS;
+        static constexpr TI INTERNAL_BATCH_SIZE = SPEC::INTERNAL_BATCH_SIZE;
+        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;
+        template<bool DYNAMIC_ALLOCATION=true>
+        using Buffer = max_pool2d::Buffer;
+        template<bool DYNAMIC_ALLOCATION=true>
+        using State = max_pool2d::State;
+    };
+
+    template<typename SPEC>
+    struct LayerBackward: public LayerForward<SPEC>{};
+
+    template<typename SPEC>
+    struct LayerGradient: public 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::max_pool2d::Layer<CONFIG, CAPABILITY, INPUT_SHAPE>;
+    };
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/max_pool2d/operations_cuda.h

@@ -0,0 +1,80 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_MAX_POOL2D_OPERATIONS_CUDA_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_MAX_POOL2D_OPERATIONS_CUDA_H
+
+#include "../../../devices/cuda.h"
+#include "../../../nn/nn.h"
+#include "../../../mode/mode.h"
+#include "layer.h"
+
+#include <cudnn.h>
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools{
+    template<typename DEV_SPEC, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void evaluate(devices::CUDA<DEV_SPEC>& device, const nn::layers::max_pool2d::LayerForward<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<OUTPUT_SPEC>& output, nn::layers::max_pool2d::Buffer&, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        static_assert(nn::layers::max_pool2d::check_input_output<LAYER_SPEC, INPUT_SPEC, OUTPUT_SPEC>);
+        using T = typename OUTPUT_SPEC::T;
+        using TI = typename devices::CUDA<DEV_SPEC>::index_t;
+        constexpr TI N = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        constexpr TI IH = LAYER_SPEC::INPUT_HEIGHT, IW = LAYER_SPEC::INPUT_WIDTH, C = LAYER_SPEC::INPUT_CHANNELS;
+        constexpr TI OH = LAYER_SPEC::OUTPUT_HEIGHT, OW = LAYER_SPEC::OUTPUT_WIDTH;
+        constexpr TI KH = LAYER_SPEC::KERNEL_HEIGHT, KW = LAYER_SPEC::KERNEL_WIDTH;
+        constexpr TI SH = LAYER_SPEC::STRIDE_H, SW = LAYER_SPEC::STRIDE_W;
+        constexpr TI PH = LAYER_SPEC::PADDING_H, PW = LAYER_SPEC::PADDING_W;
+        constexpr cudnnDataType_t dt = nn::cuda::get_cudnn_dtype<T>();
+        cudnnTensorDescriptor_t xd, yd;
+        check_cudnn_call(device, cudnnCreateTensorDescriptor(&xd), "cudnnCreateTensorDescriptor maxpool.xd");
+        check_cudnn_call(device, cudnnSetTensor4dDescriptor(xd, CUDNN_TENSOR_NHWC, dt, N, C, IH, IW), "cudnnSetTensor4dDescriptor maxpool.xd");
+        check_cudnn_call(device, cudnnCreateTensorDescriptor(&yd), "cudnnCreateTensorDescriptor maxpool.yd");
+        check_cudnn_call(device, cudnnSetTensor4dDescriptor(yd, CUDNN_TENSOR_NHWC, dt, N, C, OH, OW), "cudnnSetTensor4dDescriptor maxpool.yd");
+        cudnnPoolingDescriptor_t pd;
+        check_cudnn_call(device, cudnnCreatePoolingDescriptor(&pd), "cudnnCreatePoolingDescriptor maxpool.pd");
+        check_cudnn_call(device, cudnnSetPooling2dDescriptor(pd, CUDNN_POOLING_MAX, CUDNN_NOT_PROPAGATE_NAN, KH, KW, PH, PW, SH, SW), "cudnnSetPooling2dDescriptor maxpool.pd");
+        float alpha = 1, beta = 0;
+        check_cudnn_call(device, cudnnPoolingForward(device.cudnn_handle, pd, &alpha, xd, input._data, &beta, yd, output._data), "cudnnPoolingForward maxpool");
+        check_cudnn_call(device, cudnnDestroyPoolingDescriptor(pd), "cudnnDestroyPoolingDescriptor maxpool.pd");
+        check_cudnn_call(device, cudnnDestroyTensorDescriptor(xd), "cudnnDestroyTensorDescriptor maxpool.xd");
+        check_cudnn_call(device, cudnnDestroyTensorDescriptor(yd), "cudnnDestroyTensorDescriptor maxpool.yd");
+        check_status(device);
+    }
+
+    template<typename DEV_SPEC, typename LAYER_SPEC, typename INPUT_SPEC, typename OUTPUT_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(devices::CUDA<DEV_SPEC>& device, nn::layers::max_pool2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<OUTPUT_SPEC>& output, nn::layers::max_pool2d::Buffer& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        evaluate(device, static_cast<const nn::layers::max_pool2d::LayerForward<LAYER_SPEC>&>(layer), input, output, buffer, rng, mode);
+    }
+
+    template<typename DEV_SPEC, typename LAYER_SPEC, typename INPUT_SPEC, typename D_OUTPUT_SPEC, typename MODE = mode::Default<>>
+    void backward(devices::CUDA<DEV_SPEC>& device, nn::layers::max_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<D_OUTPUT_SPEC>& d_output, nn::layers::max_pool2d::Buffer&, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+    }
+
+    template<typename DEV_SPEC, 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(devices::CUDA<DEV_SPEC>& device, nn::layers::max_pool2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<INPUT_SPEC>& input, Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_INPUT_SPEC>& d_input, nn::layers::max_pool2d::Buffer&, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        using T = typename D_OUTPUT_SPEC::T;
+        using TI = typename devices::CUDA<DEV_SPEC>::index_t;
+        constexpr TI N = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        constexpr TI IH = LAYER_SPEC::INPUT_HEIGHT, IW = LAYER_SPEC::INPUT_WIDTH, C = LAYER_SPEC::INPUT_CHANNELS;
+        constexpr TI OH = LAYER_SPEC::OUTPUT_HEIGHT, OW = LAYER_SPEC::OUTPUT_WIDTH;
+        constexpr TI KH = LAYER_SPEC::KERNEL_HEIGHT, KW = LAYER_SPEC::KERNEL_WIDTH;
+        constexpr TI SH = LAYER_SPEC::STRIDE_H, SW = LAYER_SPEC::STRIDE_W;
+        constexpr TI PH = LAYER_SPEC::PADDING_H, PW = LAYER_SPEC::PADDING_W;
+        constexpr cudnnDataType_t dt = nn::cuda::get_cudnn_dtype<T>();
+        cudnnTensorDescriptor_t xd, yd;
+        check_cudnn_call(device, cudnnCreateTensorDescriptor(&xd), "cudnnCreateTensorDescriptor maxpool_bwd.xd");
+        check_cudnn_call(device, cudnnSetTensor4dDescriptor(xd, CUDNN_TENSOR_NHWC, dt, N, C, IH, IW), "cudnnSetTensor4dDescriptor maxpool_bwd.xd");
+        check_cudnn_call(device, cudnnCreateTensorDescriptor(&yd), "cudnnCreateTensorDescriptor maxpool_bwd.yd");
+        check_cudnn_call(device, cudnnSetTensor4dDescriptor(yd, CUDNN_TENSOR_NHWC, dt, N, C, OH, OW), "cudnnSetTensor4dDescriptor maxpool_bwd.yd");
+        cudnnPoolingDescriptor_t pd;
+        check_cudnn_call(device, cudnnCreatePoolingDescriptor(&pd), "cudnnCreatePoolingDescriptor maxpool_bwd.pd");
+        check_cudnn_call(device, cudnnSetPooling2dDescriptor(pd, CUDNN_POOLING_MAX, CUDNN_NOT_PROPAGATE_NAN, KH, KW, PH, PW, SH, SW), "cudnnSetPooling2dDescriptor maxpool_bwd.pd");
+        float alpha = 1, beta = 0;
+        check_cudnn_call(device, cudnnPoolingBackward(device.cudnn_handle, pd, &alpha, yd, layer.output._data, yd, d_output._data, xd, input._data, &beta, xd, d_input._data), "cudnnPoolingBackward maxpool");
+        check_cudnn_call(device, cudnnDestroyPoolingDescriptor(pd), "cudnnDestroyPoolingDescriptor maxpool_bwd.pd");
+        check_cudnn_call(device, cudnnDestroyTensorDescriptor(xd), "cudnnDestroyTensorDescriptor maxpool_bwd.xd");
+        check_cudnn_call(device, cudnnDestroyTensorDescriptor(yd), "cudnnDestroyTensorDescriptor maxpool_bwd.yd");
+        check_status(device);
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif