Model.h File

Model — the simplified entry point for loading and running a compiled model archive on Modalix. More...

Included Headers

#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 a compiled model archive; 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 contract 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 contract. More...
struct	Options
	User-provided options at Model construction time. More...
struct	RouteOptions
	Options for Model::graph() — controls how the model assembles into a Graph. More...
class	Runner
	Long-lived execution handle returned by Model::build(...). More...

Description

Model — the simplified entry point for loading and running a compiled model archive on Modalix.

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

See Also

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

See Also

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

See Also

"MPK contract" (§0.16) for the inference contract embedded in Model archives

File Listing

The file content with the documentation metadata removed is:

1

19#pragma once

20

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

45class Graph;

46

48using TensorSpec = TensorConstraint;

49

79class Model {

80public:

88 struct ModelInfo {

90 struct RouteNeeds {

91 bool pre_quantization =

92 false;

93 bool pre_tessellation =

94 false;

95 bool pre_cast = false;

96 bool post_detessellation =

97 false;

98 bool post_dequantization =

99 false;

100 bool post_cast = false;

101 };

102

104 struct RouteCapabilities {

105 bool has_pre_quantization =

106 false;

107 bool has_pre_tessellation =

108 false;

109 bool has_pre_cast =

110 false;

111 bool has_post_detessellation =

112 false;

113 bool has_post_dequantization =

114 false;

115 bool has_post_cast = false;

116 bool has_post_boxdecode =

117 false;

118 };

119

121 struct RouteSelection {

122 bool include_preprocess_stage = true;

123 bool include_postprocess_stage = true;

124 bool infer_only = false;

125 std::string preprocess_graph;

127 std::string selected_post_kind;

129 };

130

132 struct OutputTopology {

133 std::size_t physical_outputs = 0U;

134 std::size_t logical_outputs =

135 0U;

136 bool packed_outputs = false;

137 };

138

139 std::string mpk_json_path;

141 std::string model_name;

142

143 RouteNeeds needs;

144 RouteCapabilities capabilities;

145 RouteSelection selection;

146 OutputTopology output_topology;

147

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

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

150 std::vector<std::string>

151 warnings;

152 };

153

161 struct InferenceTerminalPolicy {

162 bool mla_only = false;

163 std::optional<std::size_t>

164 last_stage_index;

165 std::optional<std::string>

166 last_stage_name;

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

168 std::optional<std::string>

169 last_processor;

170 };

171

179 struct PreprocessRequirements {

180 bool has_preproc_stage = false;

181 bool quant_needed = false;

182 bool tess_needed = false;

183 std::string input_media_type;

185 std::string input_format;

186 std::string output_format;

187 std::string output_dtype;

189 std::vector<int> axis_perm;

190 std::vector<int> output_shape;

191 std::vector<int> slice_shape;

192 std::optional<double> q_scale;

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

195 };

196

210 struct Options {

212 PreprocessOptions preprocess;

213

214 // ── Postprocessing / detection ─────────────────────────────────────────────────────────

222 BoxDecodeType decode_type = BoxDecodeType::Unspecified;

223 float score_threshold = 0.0f;

224 float nms_iou_threshold =

225 0.0f;

226 int top_k = 0;

230 int boxdecode_original_width = 0;

234 int boxdecode_original_height = 0;

235

236 // ── Naming / wiring ────────────────────────────────────────────────────────────────────

239 std::string upstream_name = "decoder";

242 std::string name_suffix;

243

244 // ── Extraction lifecycle ───────────────────────────────────────────────────────────────

251 bool cleanup_extracted_model_data = true;

252

254 VerboseOptions verbose;

255

257 InferenceTerminalPolicy inference_terminal;

258

262 ProcessCvuOptions processcvu;

263

265 ProcessMlaOptions processmla;

266

268 PreparedRunnerOptions prepared_runner;

269

272 int async_queue_depth = 0;

273 };

274

282 struct RouteOptions {

283 bool include_input = false;

284 bool include_output = false;

297 bool expose_all_outputs = false;

298 std::string upstream_name;

300 std::string name_suffix;

301 std::string

302 buffer_name;

303 VerboseOptions verbose;

310 std::string processcvu_requested_run_target = "AUTO";

311

315 ProcessCvuOptions processcvu;

316

319 ProcessMlaOptions processmla;

320

323 PreparedRunnerOptions prepared_runner;

324

327 int async_queue_depth = 0;

328 };

329

335 enum class Stage {

336 Preprocess,

337 Inference,

338 Postprocess,

339 Full

340 };

341

352 explicit Model(const std::string& model_path);

361 explicit Model(const std::string& model_path, const Options& opt);

362

363 Model(Model&&) noexcept;

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

365 ~Model();

367

368 // ── Stage composition ────────────────────────────────────────────────────────────────────

370 simaai::neat::Graph preprocess() const;

372 simaai::neat::Graph inference() const;

374 simaai::neat::Graph postprocess() const;

376 simaai::neat::Graph graph() const;

379 simaai::neat::Graph graph(RouteOptions opt) const;

380

381 // ── Introspection ────────────────────────────────────────────────────────────────────────

383 TensorSpec input_spec() const;

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

387 TensorSpec output_spec() const;

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

397 int compiled_batch_size() const;

400 ResolvedPreprocessPlan resolved_preprocess_plan() const;

402 PreprocessRequirements preprocess_requirements() const;

404 ModelInfo info() const;

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

407

408 // ── Advanced / graph composition ─────────────────────────────────────────────────────────

410 Graph fragment(Stage stage) const;

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

414

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

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

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

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

426 std::string infer_output_name() const;

427

447 class Runner {

448 public:

450 Runner() = default;

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

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

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

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

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

461

463 explicit operator bool() const noexcept;

464#if defined(SIMA_WITH_OPENCV)

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

470#endif

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

474 bool push(const simaai::neat::Sample& inputs);

481 simaai::neat::Sample pull(int timeout_ms = -1);

482#if defined(SIMA_WITH_OPENCV)

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

485#endif

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

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

490

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

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

501 void close();

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

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

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

509 std::string metrics_report(

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

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

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

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

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

519 void close_input();

520

521 private:

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

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

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

525 };

526

527private:

528 static const RouteOptions& default_route_options();

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

530

531public:

541 Runner build();

543 Runner build(const RouteOptions& opt);

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

547 Runner build(const RouteOptions& opt, const simaai::neat::RunOptions& run_opt);

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

562 const RouteOptions& opt = default_route_options(),

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

565 Runner build(const simaai::neat::Sample& inputs,

566 const RouteOptions& opt = default_route_options(),

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

568#if defined(SIMA_WITH_OPENCV)

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

571 const RouteOptions& opt = default_route_options(),

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

573#endif

574

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

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

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

590#if defined(SIMA_WITH_OPENCV)

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

593#endif

594

595private:

596 friend struct internal::ModelAccess;

597 struct Impl;

598 std::unique_ptr<Impl> impl_;

599};

600

601} // namespace simaai::neat

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.