Loading...
Searching...
No Matches
Runtime-owned ONNX session and reusable input/output storage.
More...
|
| | State (const std::string &model, const std::string &metadata, const NetworkOptions &policy) |
| | Validate the export contract before binding a reusable ONNX session.
|
| | ~State () |
| | Close the session while every borrowed member buffer is still alive.
|
|
| std::unique_ptr< ConditioningCache > | conditioning_cache |
| onnx_session | session {} |
| | Shared session borrowing this state's fixed buffers.
|
| unsigned | timeout_ms = 0 |
| | Whole-plan cooperative deadline; zero disables the local timer.
|
| std::array< float, DIPLA_COORDINATE_COUNT *points > | candidate {} |
| | Float32 graph input: eight X values, then eight Y, then eight Z, not interleaved XYZ.
|
| std::array< float, DIPLA_COORDINATE_COUNT *points > | predicted {} |
| | Noise prediction written by the ONNX graph for each reverse step.
|
| std::array< float, DIPLA_MAP_CHANNEL_COUNT *cells > | maps {} |
| | NCHW map input: mean, variance, marker; each channel uses row * 51 + column.
|
| std::array< float, 3 > | position {} |
| | Normalized current pose tensor, ordered east/north/altitude.
|
| std::array< float, 1 > | total {} |
| | Normalized scalar total-variance tensor.
|
| std::array< int64_t, 1 > | timestep {} |
| | Current reverse-diffusion step.
|
| std::array< double, DIPLA_DIFFUSION_STEPS > | alpha {} |
| | Trained cumulative-alpha schedule used by the deterministic DDIM update.
|
| Normalization | mean_normalization |
| | Center/scale pairs for mean, variance, and log1p(total variance), in that order.
|
| Normalization | variance_normalization |
| Normalization | total_normalization |
Runtime-owned ONNX session and reusable input/output storage.
Bind once and reuse across all twenty denoising calls and later plans. This avoids rebuilding application-side tensors in the hot loop; it does not promise allocation-free execution inside ONNX Runtime.
- Invariant
- The session borrows member buffers whose addresses remain stable until the session is closed by the destructor.
- Warning
- A State is single-owner mutable workspace, not a concurrent sampler.
◆ State()
| dipla::Network::State::State |
( |
const std::string & | model, |
|
|
const std::string & | metadata, |
|
|
const NetworkOptions & | policy ) |
|
inline |
Validate the export contract before binding a reusable ONNX session.
Reject mismatched geometry, schedules, and tensors at startup, when there is still a useful file error to report, rather than after a plausible-looking but mis-scaled route reaches the application.
- Parameters
-
| [in] | model | ONNX graph path. |
| [in] | metadata | Deployment JSON describing training-time preprocessing. |
| [in] | policy | Optional integrity descriptor, thread count, and deadline policy. |
- Precondition
- initialize_runtime() has successfully selected a compatible runtime.
- Exceptions
-
| std::exception | If reading, allocation, validation, or session creation fails. |
- Note
- Parser views are temporary. Copy only checked numeric values into members; nothing in the session may borrow the constructor's JSON arena.
- Todo
- Centralize the supported world and schedule contract. Changing these checks alone is not model support: update the public sizes, tensor shapes, grid indexing, sampler loop, and coordinate transforms together, then rerun the reference cases. Until then, reject drift.
- Todo
- Derive this shape from the same supported-grid contract as maps and Belief; do not independently make only the binding dynamic.
◆ ~State()
| dipla::Network::State::~State |
( |
| ) |
|
|
inline |
Close the session while every borrowed member buffer is still alive.
Destructor bodies run before member destruction. Explicit closure here preserves that order even though session is declared before its buffers.
◆ alpha
Trained cumulative-alpha schedule used by the deterministic DDIM update.
◆ candidate
Float32 graph input: eight X values, then eight Y, then eight Z, not interleaved XYZ.
◆ conditioning_cache
◆ maps
NCHW map input: mean, variance, marker; each channel uses row * 51 + column.
◆ mean_normalization
| Normalization dipla::Network::State::mean_normalization |
Center/scale pairs for mean, variance, and log1p(total variance), in that order.
◆ position
| std::array<float, 3> dipla::Network::State::position {} |
Normalized current pose tensor, ordered east/north/altitude.
◆ predicted
Noise prediction written by the ONNX graph for each reverse step.
◆ session
| onnx_session dipla::Network::State::session {} |
Shared session borrowing this state's fixed buffers.
◆ timeout_ms
| unsigned dipla::Network::State::timeout_ms = 0 |
Whole-plan cooperative deadline; zero disables the local timer.
◆ timestep
| std::array<int64_t, 1> dipla::Network::State::timestep {} |
Current reverse-diffusion step.
◆ total
| std::array<float, 1> dipla::Network::State::total {} |
Normalized scalar total-variance tensor.
◆ total_normalization
| Normalization dipla::Network::State::total_normalization |
◆ variance_normalization
| Normalization dipla::Network::State::variance_normalization |
The documentation for this struct was generated from the following file: