# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
#
"""Class definition for base model exporter used in OTX."""
from __future__ import annotations
import json
import logging as log
import os
import tempfile
from abc import abstractmethod
from pathlib import Path
from typing import TYPE_CHECKING, Any, Literal
from zipfile import ZipFile
from model_api.models import Model
from otx.core.exporter.exportable_code import demo
from otx.core.types.export import OTXExportFormatType, TaskLevelExportParameters
from otx.core.types.precision import OTXPrecisionType
if TYPE_CHECKING:
import onnx
import openvino
from otx.core.model.base import OTXModel
[docs]
class OTXModelExporter:
"""Base class for the model exporters used in OTX.
Args:
task_level_export_parameters (TaskLevelExportParameters): Collection of export parameters
which can be defined at a task level.
input_size (tuple[int, ...]): Input shape.
mean (tuple[float, float, float], optional): Mean values of 3 channels. Defaults to (0.0, 0.0, 0.0).
std (tuple[float, float, float], optional): Std values of 3 channels. Defaults to (1.0, 1.0, 1.0).
resize_mode (Literal["crop", "standard", "fit_to_window", "fit_to_window_letterbox"], optional):
A resize type for model preprocess. "standard" resizes images without keeping ratio.
"fit_to_window" resizes images while keeping ratio.
"fit_to_window_letterbox" resizes images and pads images to fit the size. Defaults to "standard".
pad_value (int, optional): Padding value. Defaults to 0.
swap_rgb (bool, optional): Whether to convert the image from BGR to RGB Defaults to False.
output_names (list[str] | None, optional): Names for model's outputs, which would be
embedded into resulting model. Note, that order of the output names should be the same,
as in the target model.
"""
def __init__(
self,
task_level_export_parameters: TaskLevelExportParameters,
input_size: tuple[int, ...],
mean: tuple[float, float, float] = (0.0, 0.0, 0.0),
std: tuple[float, float, float] = (1.0, 1.0, 1.0),
resize_mode: Literal["crop", "standard", "fit_to_window", "fit_to_window_letterbox"] = "standard",
pad_value: int = 0,
swap_rgb: bool = False,
output_names: list[str] | None = None,
) -> None:
self.input_size = input_size
self.mean = mean
self.std = std
self.resize_mode = resize_mode
self.pad_value = pad_value
self.swap_rgb = swap_rgb
self.task_level_export_parameters = task_level_export_parameters
self.output_names = output_names
@property
def metadata(self) -> dict[tuple[str, str], str]:
"""Collection of metadata to be stored in OpenVINO Intermediate Representation or ONNX.
This metadata is mainly used to support ModelAPI.
"""
return self.task_level_export_parameters.to_metadata()
[docs]
def export(
self,
model: OTXModel,
output_dir: Path,
base_model_name: str = "exported_model",
export_format: OTXExportFormatType = OTXExportFormatType.OPENVINO,
precision: OTXPrecisionType = OTXPrecisionType.FP32,
to_exportable_code: bool = False,
) -> Path:
"""Exports input model to the specified deployable format, such as OpenVINO IR or ONNX.
Args:
model (OTXModel): OTXModel to be exported
output_dir (Path): path to the directory to store export artifacts
base_model_name (str, optional): exported model name
format (OTXExportFormatType): final format of the exported model
precision (OTXExportPrecisionType, optional): precision of the exported model's weights
to_exportable_code (bool, optional): whether to generate exportable code
Returns:
Path: path to the exported model
"""
if export_format == OTXExportFormatType.OPENVINO:
if to_exportable_code:
return self.to_exportable_code(
model,
output_dir,
base_model_name,
precision,
)
return self.to_openvino(model, output_dir, base_model_name, precision)
if export_format == OTXExportFormatType.ONNX:
return self.to_onnx(model, output_dir, base_model_name, precision)
msg = f"Unsupported export format: {export_format}"
raise ValueError(msg)
[docs]
@abstractmethod
def to_openvino(
self,
model: OTXModel,
output_dir: Path,
base_model_name: str = "exported_model",
precision: OTXPrecisionType = OTXPrecisionType.FP32,
) -> Path:
"""Export to OpenVINO Intermediate Representation format.
Args:
model (OTXModel): OTXModel to be exported
output_dir (Path): path to the directory to store export artifacts
base_model_name (str, optional): exported model name
precision (OTXExportPrecisionType, optional): precision of the exported model's weights
Returns:
Path: path to the exported model.
"""
[docs]
@abstractmethod
def to_onnx(
self,
model: OTXModel,
output_dir: Path,
base_model_name: str = "exported_model",
precision: OTXPrecisionType = OTXPrecisionType.FP32,
embed_metadata: bool = True,
) -> Path:
"""Abstract method for ONNX export.
Converts the given torch model to ONNX format and saves it to the specified output directory.
Args:
model (OTXModel): The input PyTorch model to be converted.
output_dir (Path): The directory where the ONNX model will be saved.
base_model_name (str, optional): The name of the exported ONNX model. Defaults to "exported_model".
precision (OTXPrecisionType, optional): The precision type for the exported model.
Defaults to OTXPrecisionType.FP32.
embed_metadata (bool, optional): Flag to embed metadata in the exported ONNX model. Defaults to True.
Returns:
Path: The file path where the ONNX model is saved.
"""
[docs]
def to_exportable_code(
self,
model: OTXModel,
output_dir: Path,
base_model_name: str = "exported_model",
precision: OTXPrecisionType = OTXPrecisionType.FP32,
) -> Path:
"""Export to zip folder final OV IR model with runable demo.
Args:
model (OTXModel): OTXModel to be exported
output_dir (Path): path to the directory to store export artifacts
base_model_name (str, optional): exported model name
precision (OTXExportPrecisionType, optional): precision of the exported model's weights
Returns:
Path: path to the exported model.
"""
work_demo_dir = Path(demo.__file__).parent
parameters: dict[str, Any] = {}
output_zip_path = output_dir / "exportable_code.zip"
Path.mkdir(output_dir, exist_ok=True)
is_ir_model = isinstance(model, Model)
with tempfile.TemporaryDirectory() as temp_dir, ZipFile(output_zip_path, "x") as arch:
# model files
path_to_model = (
self.to_openvino(model, Path(temp_dir), base_model_name, precision)
if not is_ir_model
else Path(model.inference_adapter.model_path)
)
if not path_to_model.exists():
msg = f"File {path_to_model} does not exist. Check the model path."
raise RuntimeError(msg)
if not is_ir_model and self.metadata is not None:
parameters["task_type"] = self.metadata.get(("model_info", "task_type"), "")
parameters["model_type"] = self.metadata.get(("model_info", "model_type"), "")
parameters["model_parameters"] = {
"labels": self.metadata.get(("model_info", "labels"), ""),
"labels_ids": self.metadata.get(("model_info", "label_ids"), ""),
}
elif is_ir_model:
model_info = model.get_model().rt_info["model_info"]
parameters["task_type"] = model_info["task_type"].value if "task_type" in model_info else ""
parameters["model_type"] = model_info["model_type"].value if "model_type" in model_info else ""
parameters["model_parameters"] = {
"labels": model_info["labels"].value if "labels" in model_info else "",
"labels_ids": model_info["label_ids"].value if "label_ids" in model_info else "",
}
arch.write(str(path_to_model), Path("model") / "model.xml")
arch.write(path_to_model.with_suffix(".bin"), Path("model") / "model.bin")
arch.writestr(
str(Path("model") / "config.json"),
json.dumps(parameters, ensure_ascii=False, indent=4),
)
# python files
arch.write(
work_demo_dir / "requirements.txt",
Path("python") / "requirements.txt",
)
arch.write(work_demo_dir / "demo.py", Path("python") / "demo.py")
arch.write(work_demo_dir / "setup.py", Path("python") / "setup.py")
arch.write(work_demo_dir / "README.md", Path("./") / "README.md")
arch.write(work_demo_dir / "LICENSE", Path("./") / "LICENSE")
# write demo_package
demo_package = work_demo_dir / "demo_package"
for root, _, files in os.walk(demo_package):
if root.endswith("__pycache__"):
continue
for file in files:
file_path = Path(root) / file
archive_path = file_path.relative_to(demo_package)
arch.write(file_path, Path("python") / "demo_package" / archive_path)
return output_zip_path
@staticmethod
def _embed_onnx_metadata(onnx_model: onnx.ModelProto, metadata: dict[tuple[str, str], Any]) -> onnx.ModelProto:
"""Embeds metadata to ONNX model."""
for item in metadata:
meta = onnx_model.metadata_props.add()
attr_path = " ".join(map(str, item))
meta.key = attr_path.strip()
meta.value = str(metadata[item])
return onnx_model
@staticmethod
def _embed_openvino_ir_metadata(ov_model: openvino.Model, metadata: dict[tuple[str, str], Any]) -> openvino.Model:
"""Embeds metadata to OpenVINO model."""
for k, data in metadata.items():
ov_model.set_rt_info(data, list(k))
return ov_model
def _extend_model_metadata(self, metadata: dict[tuple[str, str], str]) -> dict[tuple[str, str], str]:
"""Extends metadata coming from model with preprocessing-specific parameters.
Model's original metadata has priority over exporter's extra metadata
Args:
metadata (dict[tuple[str, str], str]): existing metadata for export
Returns:
dict[tuple[str, str] ,str]: updated metadata
"""
mean_str = " ".join(map(str, self.mean))
std_str = " ".join(map(str, self.std))
extra_data = {
("model_info", "mean_values"): mean_str.strip(),
("model_info", "scale_values"): std_str.strip(),
("model_info", "resize_type"): self.resize_mode,
("model_info", "pad_value"): str(self.pad_value),
("model_info", "reverse_input_channels"): str(self.swap_rgb),
}
extra_data.update(metadata)
return extra_data
def _postprocess_openvino_model(self, exported_model: openvino.Model) -> openvino.Model:
if len(exported_model.outputs) == 1 and len(exported_model.outputs[0].get_names()) == 0:
# workaround for OVC's bug: single output doesn't have a name in OV model
exported_model.outputs[0].tensor.set_names({"output1"})
# name assignment process is similar to torch onnx export
if self.output_names is not None:
if len(exported_model.outputs) >= len(self.output_names):
if len(exported_model.outputs) != len(self.output_names):
msg = (
"Number of model outputs is greater than the number"
" of output names to assign. Please check output_names"
" argument of the exporter's constructor."
)
log.warning(msg)
for i, name in enumerate(self.output_names):
traced_names = exported_model.outputs[i].get_names()
name_found = False
for traced_name in traced_names:
if name in traced_name:
name_found = True
break
name_found = name_found and bool(len(traced_names))
if not name_found:
msg = (
f"{name} is not matched with the converted model's traced output names: {traced_names}."
" Please check output_names argument of the exporter's constructor."
)
log.warning(msg)
exported_model.outputs[i].tensor.set_names({name})
else:
msg = (
"Model has less outputs than the number of output names provided: "
f"{len(exported_model.outputs)} vs {len(self.output_names)}"
)
raise RuntimeError(msg)
if self.metadata is not None:
export_metadata = self._extend_model_metadata(self.metadata)
exported_model = self._embed_openvino_ir_metadata(exported_model, export_metadata)
return exported_model
def _postprocess_onnx_model(
self,
onnx_model: onnx.ModelProto,
embed_metadata: bool,
precision: OTXPrecisionType,
) -> onnx.ModelProto:
if embed_metadata:
metadata = {} if self.metadata is None else self._extend_model_metadata(self.metadata)
onnx_model = self._embed_onnx_metadata(onnx_model, metadata)
if precision == OTXPrecisionType.FP16:
from onnxconverter_common import float16
onnx_model = float16.convert_float_to_float16(onnx_model)
return onnx_model
