Add Anomaly modelAPI changes to releases/1.4.0 (#2563)

* bug fix for legacy openvino models

* Apply otx anomaly 1.5 changes

* Fix tests

* Fix compression config

* fix modelAPI imports

* update integration tests

* Edit config types

* Update keys in deployed model

---------

Co-authored-by: Ashwin Vaidya <ashwinitinvaidya@gmail.com>
Co-authored-by: Kim, Sungchul <sungchul.kim@intel.com>

Author: 3 people

requirements/openvino.txt

@@ -2,7 +2,7 @@
 # OpenVINO Requirements.                                                      #
 nncf==2.6.0
 onnx==1.13.0
-openvino-model-api==0.1.3
+openvino-model-api==0.1.6
 openvino==2023.0
 openvino-dev==2023.0
 openvino-telemetry>=2022.1.0

src/otx/algorithms/action/adapters/openvino/model_wrappers/openvino_models.py

@@ -20,12 +20,9 @@
 
 import numpy as np
 from openvino.model_api.adapters import OpenvinoAdapter
+from openvino.model_api.adapters.utils import RESIZE_TYPES, InputTransform
 from openvino.model_api.models.model import Model
-from openvino.model_api.models.utils import (
-    RESIZE_TYPES,
-    Detection,
-    InputTransform,
-)
+from openvino.model_api.models.utils import Detection
 
 from otx.api.entities.datasets import DatasetItemEntity
 

src/otx/algorithms/anomaly/adapters/anomalib/exportable_code/__init__.py

@@ -22,7 +22,6 @@
 import subprocess  # nosec B404
 import tempfile
 from glob import glob
-from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple, Union
 from warnings import warn
 
@@ -36,7 +35,6 @@
     PostProcessingConfigurationCallback,
 )
 from omegaconf import DictConfig, ListConfig
-from openvino.runtime import Core, serialize
 from pytorch_lightning import Trainer
 
 from otx.algorithms.anomaly.adapters.anomalib.callbacks import (
@@ -47,6 +45,8 @@
 from otx.algorithms.anomaly.adapters.anomalib.data import OTXAnomalyDataModule
 from otx.algorithms.anomaly.adapters.anomalib.logger import get_logger
 from otx.algorithms.anomaly.configs.base.configuration import BaseAnomalyConfig
+from otx.algorithms.common.utils import embed_ir_model_data
+from otx.algorithms.common.utils.utils import embed_onnx_model_data
 from otx.api.entities.datasets import DatasetEntity
 from otx.api.entities.inference_parameters import InferenceParameters
 from otx.api.entities.metrics import NullPerformance, Performance, ScoreMetric
@@ -296,6 +296,8 @@ def export(
         self._export_to_onnx(onnx_path)
 
         if export_type == ExportType.ONNX:
+            self._add_metadata_to_ir(onnx_path, export_type)
+
             with open(onnx_path, "rb") as file:
                 output_model.set_data("model.onnx", file.read())
         else:
@@ -306,7 +308,7 @@ def export(
             bin_file = glob(os.path.join(self.config.project.path, "*.bin"))[0]
             xml_file = glob(os.path.join(self.config.project.path, "*.xml"))[0]
 
-            self._add_metadata_to_ir(xml_file)
+            self._add_metadata_to_ir(xml_file, export_type)
 
             with open(bin_file, "rb") as file:
                 output_model.set_data("openvino.bin", file.read())
@@ -319,40 +321,51 @@ def export(
         output_model.set_data("label_schema.json", label_schema_to_bytes(self.task_environment.label_schema))
         self._set_metadata(output_model)
 
-    def _add_metadata_to_ir(self, xml_file: str) -> None:
-        """Adds the metadata to the model IR.
+    def _add_metadata_to_ir(self, model_file: str, export_type: ExportType) -> None:
+        """Adds the metadata to the model IR or ONNX.
 
         Adds the metadata to the model IR. So that it can be used with the new modelAPI.
         This is because the metadata.json is not used by the new modelAPI.
         # TODO CVS-114640
         # TODO: Step 1. Remove metadata.json when modelAPI becomes the default inference method.
-        # TODO: Step 2. Remove this function when Anomalib is upgraded as the model graph will contain the required ops
+        # TODO: Step 2. Update this function when Anomalib is upgraded as the model graph will contain the required ops
         # TODO: Step 3. Update modelAPI to remove pre/post-processing steps when Anomalib version is upgraded.
         """
         metadata = self._get_metadata_dict()
-        core = Core()
-        model = core.read_model(xml_file)
+        extra_model_data: Dict[Tuple[str, str], Any] = {}
         for key, value in metadata.items():
-            if key == "transform":
+            if key in ("transform", "min", "max"):
                 continue
-            model.set_rt_info(value, ["model_info", key])
+            extra_model_data[("model_info", key)] = value
         # Add transforms
         if "transform" in metadata:
             for transform_dict in metadata["transform"]["transform"]["transforms"]:
                 transform = transform_dict.pop("__class_fullname__")
                 if transform == "Normalize":
-                    model.set_rt_info(self._serialize_list(transform_dict["mean"]), ["model_info", "mean_values"])
-                    model.set_rt_info(self._serialize_list(transform_dict["std"]), ["model_info", "scale_values"])
+                    extra_model_data[("model_info", "mean_values")] = self._serialize_list(
+                        [x * 255.0 for x in transform_dict["mean"]]
+                    )
+                    extra_model_data[("model_info", "scale_values")] = self._serialize_list(
+                        [x * 255.0 for x in transform_dict["std"]]
+                    )
                 elif transform == "Resize":
-                    model.set_rt_info(transform_dict["height"], ["model_info", "orig_height"])
-                    model.set_rt_info(transform_dict["width"], ["model_info", "orig_width"])
+                    extra_model_data[("model_info", "orig_height")] = transform_dict["height"]
+                    extra_model_data[("model_info", "orig_width")] = transform_dict["width"]
                 else:
                     warn(f"Transform {transform} is not supported currently")
-        model.set_rt_info("AnomalyDetection", ["model_info", "model_type"])
-        tmp_xml_path = Path(Path(xml_file).parent) / "tmp.xml"
-        serialize(model, str(tmp_xml_path))
-        tmp_xml_path.rename(xml_file)
-        Path(str(tmp_xml_path.parent / tmp_xml_path.stem) + ".bin").unlink()
+        # Since we only need the diff of max and min, we fuse the min and max into one op
+        if "min" in metadata and "max" in metadata:
+            extra_model_data[("model_info", "normalization_scale")] = metadata["max"] - metadata["min"]
+
+        extra_model_data[("model_info", "reverse_input_channels")] = False
+        extra_model_data[("model_info", "model_type")] = "AnomalyDetection"
+        extra_model_data[("model_info", "labels")] = "Normal Anomaly"
+        if export_type == ExportType.OPENVINO:
+            embed_ir_model_data(model_file, extra_model_data)
+        elif export_type == ExportType.ONNX:
+            embed_onnx_model_data(model_file, extra_model_data)
+        else:
+            raise RuntimeError(f"not supported export type {export_type}")
 
     def _serialize_list(self, arr: Union[Tuple, List]) -> str:
         """Converts a list to space separated string."""