Contribution Guide

1. Issues

Currently, this repository is actively maintained by PINTO0309 and Hyunseok-Kim0. Therefore, our maintenance of the code and cooperation with your research requires a large time cost and strong motivation.

Simply put, there are five factors that motivate me,
1. Purpose: Personal development or Research or Product development. How much impact will solving this problem have on your efforts?
2. What: What events are occurring? What error logs are being displayed?
3. How: How did you attempt to solve the problem yourself?
4. Why: Why do you need this problem solved?
5. Resources: Material you have cited or referenced other than ONNX or JSON.
In other words, I am mainly interested in how interesting your project is to me and why you are in trouble and how big an impact it will have on your project. The issue template has all the required fields for the minimum information you want me to provide to motivate me. Please describe the information as politely and gentlemanly as possible without omission.
If your project's policy does not allow you to share ONNX files owned by your project, you may still politely submit your assignment and share only the ONNX files via email or other private means. The onnx files shared through private means will be used only to investigate the issue and will not be released to the public by me. The publicly available onnx files that you share with us in issues and pull requests may be included as part of the regression tests that we run automatically in GitHub Actions. Non-publishable onnx files and onnx files shared through private means will not be included in the regression test of GitHub Actions. You and I both do not want to take the risk of a leak, so it would be very helpful if you could share the entire model, rather than sharing it, e.g., by generating a small ONNX file with the minimum amount of reproducibility of the problem. Like this. PINTO0309#22, PINTO0309#182, PINTO0309#193
When submitting an issue, please submit as much reproducible information as possible. I do not own the same onnx files as you guys, so I am almost unable to reproduce the problems you guys are having at hand in most cases.
Also, never post logs or error messages displayed in the console in an abbreviated form. Omission makes it difficult to understand the situation.
I am not very impressed with issues that do not feel motivated. Therefore, if there is no response for more than 5 days after the last reply, the bot will automatically close the issue. However, issues and pull requests to which I have assigned TODO, discussion, or Bug will be excluded from the bot decision.

https://github.com/PINTO0309/onnx2tf/issues

2. Pull Request

Basically any reasonable pull request is welcome. There are no strict rules.
I would be very happy if you could share the ONNX file for testing if possible, as I would like to add the ONNX file for testing to the GitHub Actions for regression testing.
When you issue a pull request, GitHub Actions will automatically run a regression test of the model transformation. Therefore, you can check the GitHub Actions execution log to see if your suggested modification affects previous code.
- https://github.com/PINTO0309/onnx2tf/wiki/model_status
- https://github.com/PINTO0309/onnx2tf/actions/workflows/test-models.yml
Pull requests from engineers other than the repository owner will always fail to post the model conversion status to the Wiki due to lack of permissions to reference secrets, but there is no need to be concerned about this.

To debug, tools must be installed as follows. If you do not want to destroy your environment, you can use Docker.

Docker

git clone https://github.com/PINTO0309/onnx2tf && cd onnx2tf
docker build -t onnx2tf_develop .

VSCode devcontainer (Remote Container)

HostPC install

pip install -U pip \
&& pip install -U onnx \
&& pip install -U nvidia-pyindex \
&& pip install -U onnx-graphsurgeon \
&& pip install -U onnxruntime \
&& pip install -U onnxsim \
&& pip install -U simple_onnx_processing_tools \
&& pip install -U onnx2tf \
&& pip install -U h5py==3.7.0
&& pip install tensorflow==2.12.0

https://github.com/PINTO0309/onnx2tf/pulls

3. Code Structure

3-1. Code hierarchy

.
├── CITATION.cff
├── CONTRIBUTING.md
├── Dockerfile
├── LICENSE
├── LICENSE_onnx-tensorflow
├── README.md
├── demo.yml
├── onnx2tf
│   ├── __init__.py
│   ├── __main__.py
│   ├── onnx2tf.py             ... Main process of onnx2tf
│   ├── ops                    ... Various OP conversion processes compatible with ONNX OP Type name
│   │   ├── Abs.py
│   │   ├── Acos.py
│   │   ├── Acosh.py
:   :   :    :
│   │   ├── Xor.py
│   │   ├── _LSTM.py
│   │   ├── _Loop.py
│   │   └── __init__.py
│   └── utils                  ... common processing
│       ├── __init__.py
│       ├── colors.py
│       ├── common_functions.py
│       └── enums.py
└── setup.py

3-2. Code block structure

onnx2tf/ops/xxx.py

@print_node_info
@inverted_operation_enable_disable
@get_replacement_parameter
def make_node(
    *,
    graph_node: gs.Node,
    tf_layers_dict: dict,
    **kwargs: dict,
):

    # 1. ONNX parameter read section
    # 2. Preserving Graph Structure (Dict) section
    # 3. Param replacement section
    # 4. Generation of TF OP section
    # 5. Generation of Debug Info section

1. ONNX parameter read section

NCW to NWC, NCHW to NHWC, NCDHW to NDHWC, etc... transposition determination

It has no effect on most OPs, but is used to determine whether to transpose or not transpose the input of an OP after dimensional compression or dimensional expansion has occurred in Reshape. Determines if the ONNX output shape of the previous OP and the TensorFlow output shape are the same, returning True if they are different and False if they are the same. True indicates that the input data transposition operation must be performed; False indicates that the input data transposition operation need not be performed. tf_layers_dict[before_op_name]['before_op_output_shape_trans'] stores the output shape comparison result of the previous OP as True/False. The shape determination process is performed in the @inverted_operation_enable_disable decorator.
```
before_op_output_shape_trans_1 = \
    tf_layers_dict.get(graph_node.inputs[0].name, {}).get('before_op_output_shape_trans', True)
before_op_output_shape_trans_2 = \
    tf_layers_dict.get(graph_node.inputs[1].name, {}).get('before_op_output_shape_trans', True)
before_op_output_shape_trans = \
    before_op_output_shape_trans_1 \
    and before_op_output_shape_trans_2
```
Transposition of input values and conversion process to Numpy.ndarray

Transposes input data based on before_op_output_shape_trans True/False; if before_op_output_shape_trans=True, transposition is performed; if before_op_output_shape_trans=False, no transposition is performed. In addition, if the parameter type of ONNX is Initializer, it is automatically converted to Numpy.ndarray.
```
graph_node_input_1 = get_constant_or_variable(
    graph_node.inputs[0],
    before_op_output_shape_trans,
)
graph_node_input_2 = get_constant_or_variable(
    graph_node.inputs[1],
    before_op_output_shape_trans,
)
```

Type annotation for debugging efficiency

Although it does not affect the operation of the tool itself, type annotations are added to make the best use of Lint when debugging with IDEs such as VSCode.

boxes = tf_layers_dict[graph_node_input_1.name]['tf_node'] \
    if isinstance(graph_node_input_1, gs.Variable) else graph_node_input_1
scores = tf_layers_dict[graph_node_input_2.name]['tf_node'] \
    if isinstance(graph_node_input_2, gs.Variable) else graph_node_input_2

Read the attribute values that the OP has

Obtain the attribute values that the OP of ONNX has. If no value exists, the default value follows the description in the official ONNX specification. Operator Schemas
```
center_point_box = graph_node.attrs.get('center_point_box', 0)
```

2. Preserving Graph Structure (Dict) section

Generates metadata. It is needed to maintain the graph structure of the entire model.

graph_node_output: gs.Variable = graph_node.outputs[0]
shape = graph_node_output.shape
dtype = graph_node_output.dtype

tf_layers_dict[graph_node_output.name] = {
    'optype': graph_node.op,
    'shape': shape,
    'dtype': dtype,
}

3. Param replacement section

@get_replacement_parameter Overwrites the ONNX value with the parameter for extrapolation read by the decorator. Reads a JSON format file written by the user and replaces ONNX parameters according to the described rules. Parameter replacement. All information read from JSON is stored in kwargs['op_rep_params']. param_target can be either 'inputs' or 'attributes' or 'outputs' or 'op'.

boxes = replace_parameter(
    value_before_replacement=boxes,
    param_target='inputs',
    param_name=graph_node.inputs[0].name,
    **kwargs,
)
scores = replace_parameter(
    value_before_replacement=scores,
    param_target='inputs',
    param_name=graph_node.inputs[1].name,
    **kwargs,
)

4. Generation of TF OP section

Generate a TensorFlow OP and store it in tf_layers_dict[graph_node_output.name]['tf_node']. The entire graph structure is preserved in dict format.

tf_layers_dict[graph_node_output.name]['tf_node'] = \
    tf.math.abs(
        x=input_tensor,
        name=graph_node.name,
    )

The structure of dict is assumed to be as follows when this is processed.

tf_layers_dict[graph_node_output.name] = {
    'optype': graph_node.op,
    'shape': shape,
    'dtype': dtype,
    'tf_node': (TensorFlow Op),
}

5. Generation of Debug Info section

Generates model conversion log information. tf_op_type can be set to TensorFlow OP as is or to any string. tf_inputs and tf_outputs can be set to any key value. The @print_node_info decorator automatically reads the information and outputs the conversion log to the console.

tf_layers_dict[graph_node_output.name]['tf_node_info'] = \
    make_tf_node_info(
        node_info={
            'tf_op_type': tf.image.non_max_suppression,
            'tf_inputs': {
                'boxes': tf_boxes,
                'scores': tf_scores,
                'max_output_boxes_per_class': max_output_boxes_per_class,
                'iou_threshold': iou_threshold,
                'score_threshold': score_threshold,
            },
            'tf_outputs': {
                'output': tf_layers_dict[graph_node_output.name]['tf_node'],
            },
        }
    )

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

CONTRIBUTING.md

CONTRIBUTING.md

Contribution Guide

1. Issues

2. Pull Request

3. Code Structure

3-1. Code hierarchy

3-2. Code block structure

1. ONNX parameter read section

2. Preserving Graph Structure (Dict) section

3. Param replacement section

4. Generation of TF OP section

5. Generation of Debug Info section

Files

CONTRIBUTING.md

Latest commit

History

CONTRIBUTING.md

File metadata and controls

Contribution Guide

1. Issues

2. Pull Request

3. Code Structure

3-1. Code hierarchy

3-2. Code block structure

1. ONNX parameter read section

2. Preserving Graph Structure (Dict) section

3. Param replacement section

4. Generation of TF OP section

5. Generation of Debug Info section