mindspore-mindinsight/mindinsight/mindconverter/README.md

• MindConverter tutorial
• • Overview
  • Installation
  • Usage
  • • PyTorch Model Scripts Migration
    • • MindConverter Provides Two Modes for PyTorch：
    • TensorFlow Model Scripts Migration
    • ONNX Model File Migration
  • Scenario
  • Example
  • • AST-Based Conversion
    • Graph-Based Conversion
    • • PyTorch Model Scripts Conversion
      • TensorFlow Model Scripts Conversion
      • ONNX Model File Conversion
  • Caution
  • Unsupported situation of AST mode
  • • Situation1
    • Situation2
  • Requirements
  • Frequently asked questions
  • Appendix
  • • TensorFlow Pb model exporting
    • MindConverter Error Code Definition

MindConverter tutorial

查看中文

• MindConverter tutorial
  • Overview
  • Installation
  • Usage
    • PyTorch Model Scripts Migration
    • TensorFlow Model Scripts Migration
    • ONNX Model File Migration
  • Scenario
  • Example
    • AST-Based Conversion
    • Graph-Based Conversion
      • PyTorch Model Scripts Conversion
      • TensorFlow Model Scripts Conversion
      • ONNX Model File Conversion
  • Caution
  • Unsupported situation of AST mode
    • Situation1
    • Situation2
  • Requirements
  • Frequently asked questions
  • Appendix
    • Tensorflow Pb Model Exporting
    • MindConverter Error Code Definition

Overview

MindConverter is a migration tool to transform the model scripts and weights from PyTorch, TensorFlow or ONNX to MindSpore. Users can migrate their PyTorch, TensorFlow or ONNX models to MindSpore rapidly with minor changes according to the conversion report.

Installation

MindConverter is a submodule in MindInsight. Please follow the Guide here to install MindInsight.

Usage

MindConverter currently only provides command-line interface. Here is the manual page.

usage: mindconverter [-h] [--version] [--in_file IN_FILE]
                     [--model_file MODEL_FILE] [--shape SHAPE [SHAPE ...]]
                     [--input_nodes INPUT_NODES [INPUT_NODES ...]]
                     [--output_nodes OUTPUT_NODES [OUTPUT_NODES ...]]
                     [--output OUTPUT] [--report REPORT]
                     [--project_path PROJECT_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --version             show program version number and exit
  --in_file IN_FILE     Specify path for script file to use AST schema to do
                        script conversation.
  --model_file MODEL_FILE
                        PyTorch(.pth), Tensorflow(.pb) or ONNX(.onnx) model
                        file path is expected to do script generation based on
                        graph schema. When `--in_file` and `--model_file` are
                        both provided, use AST schema as default.
  --shape SHAPE [SHAPE ...]
                        Optional, expected input tensor shape of
                        `--model_file`. It is required when use graph based
                        schema. Both order and number should be consistent
                        with `--input_nodes`. Usage: --shape 1,512 1,512
  --input_nodes INPUT_NODES [INPUT_NODES ...]
                        Optional, input node(s) name of `--model_file`. It is
                        required when use TensorFlow and ONNX model. Both
                        order and number should be consistent with `--shape`.
                        Usage: --input_nodes input_1:0 input_2:0
  --output_nodes OUTPUT_NODES [OUTPUT_NODES ...]
                        Optional, output node(s) name of `--model_file`. It is
                        required when use TensorFlow and ONNX model. Usage:
                        --output_nodes output_1:0 output_2:0
  --output OUTPUT       Optional, specify path for converted script file
                        directory. Default output directory is `output` folder
                        in the current working directory.
  --report REPORT       Optional, specify report directory. Default is
                        converted script directory.
  --project_path PROJECT_PATH
                        Optional, PyTorch scripts project path. If PyTorch
                        project is not in PYTHONPATH, please assign
                        `--project_path` when use graph based schema. Usage:
                        --project_path ~/script_file/

PyTorch Model Scripts Migration

MindConverter Provides Two Modes for PyTorch：

1. Abstract Syntax Tree (AST) based conversion: Use the argument --in_file will enable the AST mode.
2. Computational Graph based conversion: Use --model_file and --shape arguments will enable the Graph mode.

The AST mode will be enabled, if both --in_file and --model_file are specified.

For the Graph mode, --shape is mandatory.

For the AST mode, --shape is ignored.

--output and --report is optional. MindConverter creates an output folder under the current working directory, and outputs generated scripts, converted checkpoint file, weight map file and conversion reports to it.

Please note that your original PyTorch project is included in the module search path (PYTHONPATH). Use the python interpreter and test your module can be successfully loaded by import command. Use --project_path instead if your project is not in the PYTHONPATH to ensure MindConverter can load it.

Assume the project is located at /home/user/project/model_training, users can use this command to add the project to PYTHONPATH : export PYTHONPATH=/home/user/project/model_training:$PYTHONPATH
MindConverter needs the original PyTorch scripts because of the reverse serialization.

PyTorch(.pth) conversion only supports one input and one output model, it is recommended to convert multi-input or multi-output PyTorch script using ONNX conversion after converting PyTorch script to ONNX file.

TensorFlow Model Scripts Migration

MindConverter provides computational graph based conversion for TensorFlow: Transformation will be done given --model_file, --shape, --input_nodes and --output_nodes.

AST mode is not supported for TensorFlow, only computational graph based mode is available.

ONNX Model File Migration

MindConverter provides computational graph based conversion for ONNX: Transformation will be done given --model_file, --shape, --input_nodes and --output_nodes.

AST mode is not supported for ONNX, only computational graph based mode is available.

Scenario

MindConverter provides two modes for different migration demands.

1. Keep original scripts' structures, including variables, functions, and libraries.
2. Keep extra modifications as few as possible, or no modifications are required after conversion.

The AST mode is recommended for the first demand (AST mode is only supported for PyTorch). It parses and analyzes PyTorch scripts, then replace them with the MindSpore AST to generate codes. Theoretically, The AST mode supports any model script. However, the conversion may differ due to the coding style of original scripts.

For the second demand, the Graph mode is recommended. As the computational graph is a standard descriptive language, it is not affected by user's coding style. This mode may have more operators converted as long as these operators are supported by MindConverter.

Some typical image classification networks have been tested for the Graph mode. Note that:

1. The Dropout operator will be lost after conversion because the inference mode is used to load the PyTorch or TensorFlow model. Manually re-implement is necessary.
2. The Graph-based mode will be continuously developed and optimized with further updates.

Supported models list (Models in below table have been tested based on PyTorch 1.5.0 and TensorFlow 1.15.0, X86 Ubuntu released version).

Example

AST-Based Conversion

Assume the PyTorch script is located at /home/user/model.py, and outputs the transformed MindSpore script to /home/user/output, with the conversion report to /home/user/output/report. Use the following command:

mindconverter --in_file /home/user/model.py \
              --output /home/user/output \
              --report /home/user/output/report

In the conversion report, non-transformed code is listed as follows:

line <row>:<col> [UnConvert] 'operator' didn't convert. ...

For non-transformed operators, the original code keeps. Please manually migrate them. Click here for more information about operator mapping.

Here is an example of the conversion report:

 [Start Convert]
 [Insert] 'import mindspore.ops.operations as P' is inserted to the converted file.
 line 1:0: [Convert] 'import torch' is converted to 'import mindspore'.
 ...
 line 157:23: [UnConvert] 'nn.AdaptiveAvgPool2d' didn't convert. Maybe could convert to mindspore.ops.operations.ReduceMean.
 ...
 [Convert Over]

For non-transformed operators, suggestions are provided in the report. For instance, MindConverter suggests that replace torch.nn.AdaptiveAvgPool2d with mindspore.ops.operations.ReduceMean.

Graph-Based Conversion

PyTorch Model Scripts Conversion

Assume the PyTorch model (.pth file) is located at /home/user/model.pth, with input shape (1, 3, 224, 224) and the original PyTorch script is at /home/user/project/model_training. Output the transformed MindSpore script, MindSpore checkpoint file and weight map file to /home/user/output, with the conversion report to /home/user/output/report. Use the following command:

mindconverter --model_file /home/user/model.pth --shape 1,3,224,224 \
              --output /home/user/output \
              --report /home/user/output/report \
              --project_path /home/user/project/model_training

The Graph mode has the same conversion report as the AST mode. However, the line number and column number refer to the transformed scripts since no original scripts are used in the process.

In addition, input and output Tensor shape of unconverted operators shows explicitly (input_shape and output_shape) as comments in converted scripts to help further manual modifications. Here is an example of the Reshape operator (Not supported in current version):

class Classifier(nn.Cell):

    def __init__(self):
        super(Classifier, self).__init__()
        ...
        self.reshape = onnx.Reshape(input_shape=(1, 1280, 1, 1),
                                    output_shape=(1, 1280))
        ...

    def construct(self, x):
        ...
        # Suppose input of `reshape` is x.
        reshape_output = self.reshape(x)
        ...

It is convenient to replace the operators according to the input_shape and output_shape parameters. The replacement is like this:

from mindspore.ops import operations as P
...

class Classifier(nn.Cell):

    def __init__(self):
        super(Classifier, self).__init__()
        ...
        self.reshape = P.Reshape(input_shape=(1, 1280, 1, 1),
                                 output_shape=(1, 1280))
        ...

    def construct(self, x):
        ...
        # Suppose input of `reshape` is x.
        reshape_output = self.reshape(x, (1, 1280))
        ...

--output and --report are optional. MindConverter creates an output folder under the current working directory, and outputs generated scripts, MindSpore checkpoint file, weight map file and conversion reports to it.

Here is an example of the weight map:

{
    "resnet50": [
        {
            "converted_weight": {
                "name": "conv2d_0.weight",
                "shape": [
                    64,
                    3,
                    7,
                    7
                ],
                "data_type": "Float32"
            },
            "source_weight": {
                "name": "conv1.weight",
                "shape": [
                    64,
                    3,
                    7,
                    7
                ],
                "data_type": "float32"
            }
        }
    ]
}

Weight information in MindSpore (converted_weight) and that in source framework(source_weight) are saved in weight map separately.

TensorFlow Model Scripts Conversion

To use TensorFlow model script migration, users need to export TensorFlow model to Pb format first, and obtain the model input node and output node name. For exporting pb model, please refer to TensorFlow Pb model exporting.

Suppose the model is saved to /home/user/xxx/frozen_model.pb, corresponding input node name is input_1:0, output node name is predictions/Softmax:0, the input shape of model is 1,224,224,3, the following command can be used to generate the script:

mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \
              --input_nodes input_1:0 \
              --output_nodes predictions/Softmax:0 \
              --output /home/user/output \
              --report /home/user/output/report

After executed, MindSpore script, MindSpore checkpoint file, weight map file and report file can be found in corresponding directory.

Since the graph based scheme is a generative method, the original TensorFlow script is not referenced in the conversion process. Therefore, the code line and column numbers involved in the generated conversion report refer to the generated script.

In addition, for operators that are not converted successfully, the input and output shape of tensor of the node will be identified in the code by input_shape and output_shape. For example, please refer to the example in PyTorch Model Scripts Conversion section.

ONNX Model File Conversion

To use ONNX model file migration, user needs to obtain the model input node and output node name from ONNX model. To get input node and output node name, Netron is recommended.

Suppose the model is saved to /home/user/xxx/model.onnx, corresponding input node name is input_1:0, output node name is predictions/Softmax:0, the input shape of model is 1,3,224,224, the following command can be used to generate the script:

mindconverter --model_file /home/user/xxx/model.onnx --shape 1,3,224,224 \
              --input_nodes input_1:0 \
              --output_nodes predictions/Softmax:0 \
              --output /home/user/output \
              --report /home/user/output/report

After executed, MindSpore script, MindSpore checkpoint file, weight map file and report file can be found in corresponding directory.

Since the graph based scheme is a generative method, the original ONNX model is not referenced in the conversion process. Therefore, the code line and column numbers involved in the generated conversion report refer to the generated script.

In addition, for operators that are not converted successfully, the input and output shape of tensor of the node will be identified in the code by input_shape and output_shape. For example, please refer to the example in PyTorch Model Scripts Conversion section.

Caution

1. PyTorch, TensorFlow are not an explicitly stated dependency libraries in MindInsight. The Graph conversion requires the consistent PyTorch or TensorFlow version as the model is trained. (For MindConverter, PyTorch 1.5.0 is supported while PyTorch 1.4.x is unsupported; PyTorch 1.6.x and PyTorch 1.7.x are untested.).
2. This script conversion tool relies on operators which supported by MindConverter and MindSpore. Unsupported operators may not be successfully mapped to MindSpore operators. You can manually edit, or implement the mapping based on MindConverter, and contribute to our MindInsight repository. We appreciate your support for the MindSpore community.
3. MindConverter converts dynamic input shape to constant one based on --shape while using graph based scheme, as a result, it is required that inputs shape used to retrain or inference in MindSpore are the same as that used to convert using MindConverter. If inputs shape has changed, rerunning MindConverter with new --shape or fixing shape related parameters in old script manually is necessary.
4. MindSpore script, MindSpore checkpoint file and weight map file are saved in the same file folder path.

Unsupported situation of AST mode

Situation1

Classes and functions that can't be converted:

1. The use of .shape, .ndim and .dtype member of torch.Tensor.
2. torch.nn.AdaptiveXXXPoolXd and torch.nn.functional.adaptive_XXX_poolXd().
3. torch.nn.functional.Dropout.
4. torch.unsqueeze() and torch.Tensor.unsqueeze().
5. torch.chunk() and torch.Tensor.chunk().

Situation2

Subclassing from the subclasses of nn.Module

e.g. (code snip from torchvision.models.mobilenet)

from torch import nn

class ConvBNReLU(nn.Sequential):
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        padding = (kernel_size - 1) // 2
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
            nn.BatchNorm2d(out_planes),
            nn.ReLU6(inplace=True)
        )

Requirements

For users using MindConverter, in addition to install the TensorFlow or PyTorch that can satisfy the model loading, inference and training requirements, users also need to pip install the following third party package (tf2onnx is not required for users that convert PyTorch model definition script to MindSpore):

onnx>=1.8.0
tf2onnx>=1.7.1
onnxruntime>=1.5.2
onnxoptimizer>=0.1.2

For some models, if the onnx or tf2onnx error message appears during the conversion process, please try to upgrade the onnx, tf2onnx or onnxoptimizer in the environment to the latest version.

Frequently asked questions

Q1. terminate called after throwing an instance of 'std::system_error', what(): Resource temporarily unavailable, Aborted (core dumped):

Answer: This problem is caused by TensorFlow. First step of conversion process is loading TensorFlow model into memory using TensorFlow module, and TensorFlow starts to apply for needed resource. When required resource is unavailable, such as exceeding max process number of Linux system limit, etc., TensorFlow will raise an error from its C/C++ layer. For more detail, please refer to TensorFlow official repository. There are some known issue for reference only:
TF ISSUE 14885, TF ISSUE 37449

Q2. Can MindConverter run on ARM platform?

Answer: MindConverter supports both x86 and ARM platform. Please ensure all required dependencies and environments installed in the ARM platform.

Q3. Why did I get message of Error detail: [NodeInputMissing] ... when converting PyTorch model?

Answer: For PyTorch model, if operations in torch.nn.functional.xxx, torch.xxx, torch.Tensor.xxx were used, node parsing could be failed. It's better to replace those operations with torch.nn.xxx.

Q4. Why does the conversion process take a lot of time (more than 10 minutes), but the model is not so large?

Answer: When converting, MindConverter needs to use protobuf to deserialize the model file. Please make sure that the protobuf installed in Python environment is implemented by C++ backend. The validation method is as follows. If the output is "python", you need to install Python protobuf implemented by C++ (download the protobuf source code, enter the "python" subdirectory in the source code, and use python setup.py install --cpp_implementation to install). If the output is "cpp" and the conversion process still takes a long time, please add environment variable export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=cpp before conversion.

from google.protobuf.internal import api_implementation
print(api_implementation.Type())

Appendix

TensorFlow Pb model exporting

If build model with Keras API, user can refer to this tutorial.

MindConverter Error Code Definition

Error code defined in MindConverter, please refer to LINK.