/usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/autograd
Edit: /usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/autograd/anomaly_mode.h (1525B)
#pragma once
#include
#include
#include
namespace torch { namespace autograd {
// forward declaration of Node from function.h
struct Node;
struct TORCH_API AnomalyMode {
static bool is_enabled() {
return _enabled;
}
static void set_enabled(bool enabled) {
_enabled = enabled;
}
private:
static bool _enabled;
};
/// A RAII guard that enables Anomaly Detection Mode.
///
/// Anomaly detection mode is useful for debugging problems happening
/// in the backward, such as unexpectedly modified tensors or NaNs
/// occuring in the backward.
///
/// The enabling of anomaly mode is global - as soon as there is one
/// such guard, it is enabled for all computation and threads. It also
/// comes with a significant performance penalty.
///
/// Example:
/// @code
/// auto x = torch::tensor({1.}, torch::requires_grad());
/// {
/// torch::autograd::DetectAnomalyGuard detect_anomaly;
/// auto x = torch::tensor({5.0}, torch::requires_grad());
/// auto y = x * x;
/// auto z = y * y;
/// y += 1;
/// z.backward();
/// }
/// @endcode
class TORCH_API DetectAnomalyGuard {
public:
DetectAnomalyGuard();
~DetectAnomalyGuard();
};
struct TORCH_API AnomalyMetadata {
virtual ~AnomalyMetadata();
virtual void store_stack();
virtual void print_stack(const std::string& current_node_name);
virtual void assign_parent(const std::shared_ptr& parent_node);
private:
std::string traceback_;
std::shared_ptr parent_;
};
}}