kl_divergence_loss.hpp

Go to the documentation of this file.

 
#pragma once
 
#include "pulsatrix/device_backend.hpp"
#include "pulsatrix/reparameterize.hpp"  // ReparamGrad -- the shared (grad_mu, grad_log_sigma) pair
#include "pulsatrix/tensor.hpp"
 
namespace pulsatrix {
 
class KLDivergenceLoss {
public:
    explicit KLDivergenceLoss(DeviceBackend* backend);
 
    [[nodiscard]] float forward(const Tensor& mu, const Tensor& log_sigma);
 
    [[nodiscard]] ReparamGrad backward() const;
 
private:
    DeviceBackend* backend_;
    Tensor last_mu_;
    Tensor last_log_sigma_;
    bool has_forwarded_ = false;
};
 
}  // namespace pulsatrix