linear_module.hpp

Go to the documentation of this file.

 
#pragma once
 
#include <optional>
#include <initializer_list>
#include <vector>
 
#include "pulsatrix/module.hpp"
 
namespace pulsatrix {
 
class LinearModule : public Module {
public:
    LinearModule(int64_t in_features, int64_t out_features, DeviceBackend* backend, DeviceType device);
 
    LinearModule(int64_t in_features, int64_t out_features, DeviceBackend* backend);
 
    [[nodiscard]] Tensor backward(const Tensor& grad_output) override;
 
    [[nodiscard]] OpType op_type() const override { return OpType::Linear; }
 
    void set_weight(std::initializer_list<float> values);
 
    void set_bias(std::initializer_list<float> values);
 
    void set_weight(const std::vector<float>& values);
 
    void set_bias(const std::vector<float>& values);
 
    [[nodiscard]] const Tensor& weight() const { return weight_; }
    [[nodiscard]] const Tensor& bias() const { return bias_; }
    [[nodiscard]] const Tensor& weight_grad() const { return weight_grad_; }
    [[nodiscard]] const Tensor& bias_grad() const { return bias_grad_; }
 
    [[nodiscard]] Tensor propagate_relevance(const Tensor& relevance_out, const LRPRuleConfig& config) override;
 
    [[nodiscard]] bool supports_lrp_rule(LRPRule) const override { return true; }
 
    [[nodiscard]] std::vector<NamedParamRef> named_parameters() override {
        return {{"weight", {&weight_, &weight_grad_}}, {"bias", {&bias_, &bias_grad_}}};
    }
 
 
    [[nodiscard]] std::optional<DeviceType> compute_device() const override { return weight_.device(); }
 
protected:
    [[nodiscard]] Tensor forward_impl(const Tensor& input) override;
 
private:
    int64_t in_features_;
    int64_t out_features_;
    DeviceBackend* backend_;
    Tensor weight_;
    Tensor bias_;
    Tensor weight_grad_;
    Tensor bias_grad_;
    Tensor last_input_;
    Tensor last_pre_bias_output_;
    bool has_forwarded_ = false;
};
 
}  // namespace pulsatrix