Model.h File

Model — the simplified entry point for loading and running an MPK on the Modalix chip. More...

Included Headers

#include "builder/NodeGroup.h" #include "model/PreprocessPlan.h" #include "nodes/io/Input.h" #include "pipeline/BoxDecodeType.h" #include "pipeline/Run.h" #include "pipeline/TensorSpec.h" #include <cstddef> #include <array> #include <cstdint> #include <memory> #include <optional> #include <string> #include <unordered_map> #include <vector>

Namespaces Index

namespace	simaai
namespace	neat
namespace	internal

Classes Index

class	Model
	Loaded form of an MPK; the simplified entry point to run inference on Modalix. More...
struct	ModelInfo
	Diagnostic snapshot of how the route planner resolved the model. More...
struct	RouteNeeds
	What the model fundamentally needs in its pre/post chain (derived from manifest dtypes). More...
struct	RouteCapabilities
	What pre/post adapters the MPK provides (read from manifest stages). More...
struct	RouteSelection
	What the planner actually included in the materialized route. More...
struct	OutputTopology
	Output topology: how many tensors the model emits, and whether they're physically packed. More...
struct	InferenceTerminalPolicy
	Where the inference pipeline should terminate. More...
struct	PreprocessRequirements
	Concrete preprocess parameters resolved by the planner from the MPK manifest. More...
struct	Options
	User-provided options at Model construction time. More...
struct	SessionOptions
	Options for Model::session() — controls how the model assembles into a NodeGroup. More...
class	Runner
	Long-lived execution handle returned by Model::build(...). More...

Description

Model — the simplified entry point for loading and running an MPK on the Modalix chip.

Model is the user-facing wrapper around an MPK (model pack) .tar.gz. It loads the file, extracts and validates the manifest, runs the route planner, and exposes ready-to-use NodeGroup fragments (preprocess, inference, postprocess) plus convenience run() and build() methods that drive a one-shot inference. Internally a Model is a Session wrapper: the same composition, validation, and runtime machinery the Session API exposes powers Model underneath. New users start with Model::run(input); advanced users compose their own Session from model.session() plus extra Nodes.

See Also

"Models" in the design deep dive (§0.7 — The main concepts)

See Also

"Sessions: the assembly contract" (§0.12) for what Model wraps

See Also

"MPK contract" (§0.16) for the file format Model loads

File Listing

The file content with the documentation metadata removed is:

1

18#pragma once

19

20#include "builder/NodeGroup.h"

21#include "model/PreprocessPlan.h"

22#include "nodes/io/Input.h"

23#include "pipeline/BoxDecodeType.h"

24#include "pipeline/Run.h"

25#include "pipeline/TensorSpec.h"

26

27#include <cstddef>

28#include <array>

29#include <cstdint>

30#include <memory>

31#include <optional>

32#include <string>

33#include <unordered_map>

34#include <vector>

35

36#if defined(SIMA_WITH_OPENCV)

37#include <opencv2/core/mat.hpp>

38#endif

39

40namespace simaai::neat {

41

42namespace internal {

43struct ModelAccess;

44} // namespace internal

45

47using TensorSpec = TensorConstraint;

48

77class Model {

78public:

86 struct ModelInfo {

88 struct RouteNeeds {

89 bool pre_quantization =

90 false;

91 bool pre_tessellation =

92 false;

93 bool pre_cast = false;

94 bool post_detessellation =

95 false;

96 bool post_dequantization =

97 false;

98 bool post_cast = false;

99 };

100

102 struct RouteCapabilities {

103 bool has_pre_quantization = false;

104 bool has_pre_tessellation =

105 false;

106 bool has_pre_cast = false;

107 bool has_post_detessellation =

108 false;

109 bool has_post_dequantization = false;

110 bool has_post_cast = false;

111 bool has_post_boxdecode =

112 false;

113 };

114

116 struct RouteSelection {

117 bool include_preprocess_stage = true;

118 bool include_postprocess_stage = true;

119 bool infer_only = false;

120 std::string preprocess_graph;

122 std::string selected_post_kind;

124 };

125

127 struct OutputTopology {

128 std::size_t physical_outputs = 0U;

129 std::size_t logical_outputs =

130 0U;

131 bool packed_outputs = false;

132 };

133

134 std::string mpk_json_path;

135 std::string model_name;

136

137 RouteNeeds needs;

138 RouteCapabilities capabilities;

139 RouteSelection selection;

140 OutputTopology output_topology;

141

142 std::vector<std::string> pre_kernels;

143 std::vector<std::string> post_kernels;

144 std::vector<std::string>

145 warnings;

146 };

147

155 struct InferenceTerminalPolicy {

156 bool mla_only = false;

157 std::optional<std::size_t>

158 last_stage_index;

159 std::optional<std::string>

160 last_stage_name;

161 std::optional<std::string> last_plugin_id;

162 std::optional<std::string>

163 last_processor;

164 };

165

173 struct PreprocessRequirements {

174 bool has_preproc_stage = false;

175 bool quant_needed = false;

176 bool tess_needed = false;

177 std::string input_media_type;

179 std::string input_format;

180 std::string output_format;

181 std::string output_dtype;

183 std::vector<int> axis_perm;

184 std::vector<int> output_shape;

185 std::vector<int> slice_shape;

186 std::optional<double> q_scale;

188 std::optional<std::int64_t> q_zp;

189 };

190

204 struct Options {

206 PreprocessOptions preprocess;

207

208 // ── Postprocessing / detection ─────────────────────────────────────────────────────────

216 BoxDecodeType decode_type = BoxDecodeType::Unspecified;

217 float score_threshold = 0.0f;

218 float nms_iou_threshold =

219 0.0f;

220 int top_k = 0;

224 int boxdecode_original_width = 0;

228 int boxdecode_original_height = 0;

229

230 // ── Naming / wiring ────────────────────────────────────────────────────────────────────

233 std::string upstream_name = "decoder";

236 std::string name_suffix;

237

238 // ── Extraction lifecycle ───────────────────────────────────────────────────────────────

245 bool cleanup_extracted_model_data = true;

246

248 VerboseOptions verbose;

249

251 InferenceTerminalPolicy inference_terminal;

252

256 ProcessCvuOptions processcvu;

257

259 ProcessMlaOptions processmla;

260

262 PreparedRunnerOptions prepared_runner;

263

266 int async_queue_depth = 0;

267 };

268

275 struct SessionOptions {

276 bool include_appsrc = true;

277 bool include_appsink = true;

278 std::string upstream_name;

280 std::string name_suffix;

281 std::string

282 buffer_name;

283 VerboseOptions verbose;

290 std::string processcvu_requested_run_target = "AUTO";

291

295 ProcessCvuOptions processcvu;

296

299 ProcessMlaOptions processmla;

300

303 PreparedRunnerOptions prepared_runner;

304

307 int async_queue_depth = 0;

308 };

309

315 enum class Stage {

316 Preprocess,

317 Inference,

318 Postprocess,

319 Full

320 };

321

332 explicit Model(const std::string& mpk_path);

341 explicit Model(const std::string& mpk_path, const Options& opt);

342

343 Model(Model&&) noexcept;

344 Model& operator=(Model&&) noexcept;

345 ~Model();

346

347 // ── Stage composition ────────────────────────────────────────────────────────────────────

349 simaai::neat::NodeGroup preprocess() const;

351 simaai::neat::NodeGroup inference() const;

353 simaai::neat::NodeGroup postprocess() const;

356 simaai::neat::NodeGroup session() const;

362 simaai::neat::NodeGroup session(SessionOptions opt) const;

363

364 // ── Introspection ────────────────────────────────────────────────────────────────────────

366 TensorSpec input_spec() const;

368 std::vector<TensorSpec> input_specs() const;

370 TensorSpec output_spec() const;

372 std::vector<TensorSpec> output_specs() const;

380 int compiled_batch_size() const;

383 ResolvedPreprocessPlan resolved_preprocess_plan() const;

385 PreprocessRequirements preprocess_requirements() const;

387 ModelInfo info() const;

389 std::unordered_map<std::string, std::string> metadata() const;

390

391 // ── Advanced / graph composition ─────────────────────────────────────────────────────────

393 NodeGroup fragment(Stage stage) const;

396 std::string backend_fragment(Stage stage) const;

397

399 simaai::neat::InputOptions input_appsrc_options(bool tensor_mode) const;

401 std::vector<simaai::neat::InputOptions> input_appsrc_options_list(bool tensor_mode) const;

403 std::string find_config_path_by_plugin(const std::string& plugin_id) const;

405 std::string find_config_path_by_processor(const std::string& processor) const;

408 std::string infer_output_name() const;

409

429 class Runner {

430 public:

432 Runner() = default;

434 explicit Runner(simaai::neat::Run run);

436 Runner(simaai::neat::Run run, const simaai::neat::InputOptions& tensor_input_opt);

439 Runner(simaai::neat::Run run, const simaai::neat::InputOptions& tensor_input_opt,

440 std::vector<std::string> ingress_names);

442 Runner(simaai::neat::Run run, std::vector<std::string> ingress_names);

443

445 explicit operator bool() const noexcept;

446#if defined(SIMA_WITH_OPENCV)

451 bool push(const std::vector<cv::Mat>& inputs);

452#endif

454 bool push(const simaai::neat::TensorList& inputs);

456 bool push(const simaai::neat::SampleList& inputs);

463 simaai::neat::SampleList pull(int timeout_ms = -1);

464#if defined(SIMA_WITH_OPENCV)

466 simaai::neat::TensorList run(const std::vector<cv::Mat>& inputs, int timeout_ms = -1);

467#endif

469 simaai::neat::TensorList run(const simaai::neat::TensorList& inputs, int timeout_ms = -1);

471 simaai::neat::SampleList run(const simaai::neat::SampleList& inputs, int timeout_ms = -1);

472

474 simaai::neat::MeasureScope start_measurement(const simaai::neat::MeasureOptions& opt = {});

481 int warmup(const simaai::neat::TensorList& inputs, int warm = -1, int timeout_ms = -1);

483 void close();

485 simaai::neat::RunStats stats() const;

487 simaai::neat::RunMeasurementSummary measurement_summary() const;

489 simaai::neat::RuntimeMetrics metrics(const simaai::neat::RuntimeMetricsOptions& opt = {}) const;

491 std::string metrics_report(

492 const simaai::neat::RuntimeMetricsOptions& opt = {},

493 simaai::neat::RuntimeMetricsFormat format = simaai::neat::RuntimeMetricsFormat::Text) const;

495 std::string metrics_report(simaai::neat::RuntimeMetricsFormat format) const;

497 simaai::neat::RunDiagSnapshot diag_snapshot() const;

499 std::string report(const simaai::neat::RunReportOptions& opt = {}) const;

501 void close_input();

502

503 private:

504 simaai::neat::Run run_{};

505 std::optional<simaai::neat::InputOptions> tensor_input_opt_for_cv_{};

506 std::vector<std::string> ingress_names_;

507 };

508

509private:

510 static const SessionOptions& default_session_options();

511 static const simaai::neat::RunOptions& default_run_options();

512

513public:

523 Runner build();

525 Runner build(const SessionOptions& opt);

527 Runner build(const simaai::neat::RunOptions& run_opt);

529 Runner build(const SessionOptions& opt, const simaai::neat::RunOptions& run_opt);

543 Runner build(const simaai::neat::TensorList& inputs,

544 const SessionOptions& opt = default_session_options(),

545 const simaai::neat::RunOptions& run_opt = default_run_options());

547 Runner build(const simaai::neat::SampleList& inputs,

548 const SessionOptions& opt = default_session_options(),

549 const simaai::neat::RunOptions& run_opt = default_run_options());

550#if defined(SIMA_WITH_OPENCV)

552 Runner build(const std::vector<cv::Mat>& inputs,

553 const SessionOptions& opt = default_session_options(),

554 const simaai::neat::RunOptions& run_opt = default_run_options());

555#endif

556

557 // ── One-shot execution (synchronous) ─────────────────────────────────────────────────────

569 simaai::neat::TensorList run(const simaai::neat::TensorList& inputs, int timeout_ms = -1);

571 simaai::neat::SampleList run(const simaai::neat::SampleList& inputs, int timeout_ms = -1);

572#if defined(SIMA_WITH_OPENCV)

574 simaai::neat::TensorList run(const std::vector<cv::Mat>& inputs, int timeout_ms = -1);

575#endif

576

577private:

578 friend struct internal::ModelAccess;

579 struct Impl;

580 std::unique_ptr<Impl> impl_;

581};

582

583} // namespace simaai::neat

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.