how to export to onnx/openvino AFTER training?

[betogulliver]

I'd like to export models I've trained using anomalib (custom-dataset) to onnx/openvino.

I understand I could set the following in my config.yaml file and re-train to get the models exported :

optimization:
export_mode : openvino

Is there a way I can skip re-training and simply jump straight to the EXPORT part?

thanks for any/all help

[samet-akcay]

Hi @betogulliver,

If you set the export_mode in config.yaml` file:

optimization:
    export_mode : openvino

When enabled, anomalib runs ExportCallback at the end of the training, as shown below:
https://github.com/openvinotoolkit/anomalib/blob/8a6c6ea115ad2c460a9beaee85736d1180eaa996/anomalib/utils/callbacks/export.py#L37-L50

anomalib basically uses export function to export it to onnx or openvino. You could utilize this function for your specific need.
https://github.com/openvinotoolkit/anomalib/blob/8a6c6ea115ad2c460a9beaee85736d1180eaa996/anomalib/deploy/export.py#L52-L81

[samet-akcay]

I'm converting this to a Q&A in discussions, so let's continue there.

[betogulliver]

Hi @samet-akcay

thanks for your suggestions.
I tried them by modifying tools/test.py into my tools/export.py (see below for details)

however, when I tried to run as follows :

python tools/export.py --model padim --config anomalib/models/padim/config.yaml --weight_file results/padim/mvtec/bottle/weights/model.ckpt

I got the following error (details below)

torch.onnx.symbolic_registry.UnsupportedOperatorError: Exporting the operator ::_upsample_nearest_exact2d to ONNX opset version 11 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.

any ideas/suggestions on how to fix my problem?
NOTE: using optimization : export_mode : openvino during training works OK and results in the onnx/openvino versions of the model

best regards.

log

./anomalib/config/config.py:175: UserWarning: config.project.unique_dir is set to False. This does not ensure that your results will be written in an empty directory and you may overwrite files.
  warn(
Transform configs has not been provided. Images will be normalized using ImageNet statistics.
Transform configs has not been provided. Images will be normalized using ImageNet statistics.
/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/utilities/prints.py:36: DeprecationWarning: From v0.10 an `'Binary*'`, `'Multiclass*', `'Multilabel*'` version now exist of each classification metric. Moving forward we recommend using these versions. This base metric will still work as it did prior to v0.10 until v0.11. From v0.11 the `task` argument introduced in this metric will be required and the general order of arguments may change, such that this metric will just function as an single entrypoint to calling the three specialized versions.
  warnings.warn(*args, **kwargs)
/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/utilities/prints.py:36: UserWarning: Metric `PrecisionRecallCurve` will save all targets and predictions in buffer. For large datasets this may lead to large memory footprint.
  warnings.warn(*args, **kwargs)
FeatureExtractor is deprecated. Use TimmFeatureExtractor instead. Both FeatureExtractor and TimmFeatureExtractor will be removed in version 2023.1
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
Traceback (most recent call last):
  File "tools/export.py", line 85, in <module>
    main_export()
  File "tools/export.py", line 75, in main_export
    do_export(model, config)
  File "tools/export.py", line 30, in do_export
    export(
  File "./anomalib/deploy/export.py", line 79, in export
    onnx_path = _export_to_onnx(model, input_size, export_path)
  File "./anomalib/deploy/export.py", line 98, in _export_to_onnx
    torch.onnx.export(
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/__init__.py", line 350, in export
    return utils.export(
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/utils.py", line 163, in export
    _export(
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/utils.py", line 1074, in _export
    graph, params_dict, torch_out = _model_to_graph(
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/utils.py", line 731, in _model_to_graph
    graph = _optimize_graph(
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/utils.py", line 308, in _optimize_graph
    graph = _C._jit_pass_onnx(graph, operator_export_type)
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/__init__.py", line 416, in _run_symbolic_function
    return utils._run_symbolic_function(*args, **kwargs)
  File "/home/user/anaconda3/envs/anomalib/lib/python3.8/site-packages/torch/onnx/utils.py", line 1421, in _run_symbolic_function
    raise symbolic_registry.UnsupportedOperatorError(
torch.onnx.symbolic_registry.UnsupportedOperatorError: Exporting the operator ::_upsample_nearest_exact2d to ONNX opset version 11 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.
(anomalib) users@wstation% ```

tools/export.py

```"""Test This script performs inference on the test dataset and saves the output visualizations into a directory."""

# Copyright (C) 2022 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

from argparse import ArgumentParser, Namespace

from pytorch_lightning import Trainer

from anomalib.config import get_configurable_parameters
from anomalib.data import get_datamodule
from anomalib.models import get_model
from anomalib.utils.callbacks import get_callbacks

# python tools/export.py --model padim --config anomalib/models/padim/config.yaml --weight_file results/padim/mvtec/bottle/weights/model.ckpt
USE_EXPORT = 1
if USE_EXPORT :
    import os
    from anomalib.deploy import ExportMode, export
    def do_export(model, config, export_root="/tmp/onnx", export_mode=ExportMode.OPENVINO) :
    # 	anomalib/tests/pre_merge/utils/callbacks/export_callback/test_export.py
#        print(type(model), model)
            input_size  = config.model.input_size
            os.makedirs(export_root, exist_ok=True)
            export(
                model=model,
                input_size=input_size,
                export_root=export_root,
                export_mode=export_mode,
            )
            if export_mode == ExportMode.OPENVINO:
                assert os.path.exists(os.path.join(tmp_dir, "openvino/model.bin")), "Failed to generate OpenVINO model"
            elif export_mode == ExportMode.ONNX:
                assert os.path.exists(os.path.join(tmp_dir, "onnx/model.onnx")), "Failed to generate ONNX model"
            else:
                raise ValueError(f"Unknown export_mode {export_mode}. Supported modes: onnx or openvino.")


def get_args() -> Namespace:
    """Get CLI arguments.

    Returns:
        Namespace: CLI arguments.
    """
    parser = ArgumentParser()
    parser.add_argument("--model", type=str, default="stfpm", help="Name of the algorithm to train/test")
    parser.add_argument("--config", type=str, required=False, help="Path to a model config file")
    parser.add_argument("--weight_file", type=str, default="weights/model.ckpt")

    args = parser.parse_args()
    return args


def main_export():
    """Export an anomaly classification and segmentation model that is initially trained via `tools/train.py`.

    The script is able to write the results into both filesystem and a logger such as Tensorboard.
    """
    args = get_args()
    config = get_configurable_parameters(
        model_name=args.model,
        config_path=args.config,
        weight_file=args.weight_file,
    )

    datamodule = get_datamodule(config)
    model = get_model(config)

    if USE_EXPORT :
        do_export(model, config)
        quit()
    
    callbacks = get_callbacks(config)

    trainer = Trainer(callbacks=callbacks, **config.trainer)
    trainer.test(model=model, datamodule=datamodule)


if __name__ == "__main__":
    main_export()

[betogulliver]

SOLVED!

FYI, the code above succesfully exports models without the need o re-training.

I've just recreate (start clean) my python venv and now exporting without re-training works as expected.

[chenmo2]

大佬怎么解决，

设置转openvino模型也有问题

[monkeycc]

I also need this transformation model function

The model has been trained

Need to convert onnx

Don't want to train

Script that needs to transform the model

@samet-akcay