sonibla/pytorch_keras_converter

pytorch_keras_converter

A PyTorch-Keras converter made for Cadene's pretrained models.

Also converts some simple PyTorch models. See supported layers for more details.

I don't have a lot if time for this project, feel free to contribute if you want to :)

Table of contents

• Installation
  • Prerequisites
  • Installing
  • Troubleshooting
• Usage
  • Quick examples
  • API documentation
• Supported models and layers
• Known issues
• Contributing
• Authors
• License
• Acknowledgments

Installation

These instructions will get you a copy of the project up and running on your local machine.

Prerequisites

Python : supported versions : >=3.5

You can also install python with Anaconda.

Installing

Download files:

git clone https://github.com/sonibla/pytorch_keras_converter.git
cd pytorch_keras_converter

Then install pytorch_keras_converter using setup.py

The following command should install automatically pytorch_keras_converter and every dependency:

python3 setup.py install --user

To install on a particular version of Python (here 3.7):

python3.7 setup.py install --user

To install on the entire system (requires administrator privileges):

sudo python setup.py install

Troubleshooting

Installing pip or setuptools

If modules pip or setuptools aren't installed on your Python environment:

curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
python get-pip.py --user

Manually installing dependencies

Dependencies should install automatically when running setup.py. But if it fails, install them manually:

python -m pip install torch pretrainedmodels tensorflow graphviz numpy h5py tqdm --user

Remove --user to install on the whole system, replace python with python3.7 to install on Python 3.7.

Usage

Quick examples

• To convert se_resnet50:

>>> import pytorch_keras_converter as pkc
>>> pkc.cadene_to_tf(['se_resnet50'], quiet=False)

• To convert bninception and fbresnet152:

>>> import pytorch_keras_converter as pkc
>>> pkc.cadene_to_tf(['bninception', 'fbresnet152'])

• To convert se_resnet152 with random weights:

>>> import pytorch_keras_converter as pkc
>>> pkc.cadene_to_tf(['se_resnet152(pretrained=None)'])

• To automatically convert many models:

Create a file containing on each line one model's name. For example:

resnet18
resnet34
resnet50(pretrained=None)
resnet50

Let's store this file in models.txt.

Now, in a Python shell:

>>> import pytorch_keras_converter as pkc
>>> pkc.cadene_to_tf('models.txt')

API documentation

pytorch_keras_converter.API.convert

def convert(model,
            input_shape,
            weights=True,
            quiet=True,
            ignore_tests=False,
            input_range=None,
            save=None,
            filename=None,
            directory=None):

Arguments:

• model: A Keras or PyTorch model or layer to convert
• input_shape (list, tuple or int): Input shape, without batchsize.
• weights (bool): Also convert weights. If set to False, only convert model architecture
• quiet (bool): If False, display a progress bar and some messages
• ignore_tests (bool): If tests should be ignored. If weights is False, tests are automatically ignored.
  • If set to True, converted model will still be tested by security. If models are not identical, it will only print a warning.
  • If set to False, and models are not identical, RuntimeWarning will be raised
• input_range: Optional. A list of 2 elements containing max and min values to give as input to the model when performing the tests. If None, models will be tested on samples from the "standard normal" distribution.
• save: If model should be exported to a hdf5 file.
• filename: Optional. Filename to give to model's hdf5 file. If filename is not None and save is not False, then save will automatically be set to True
• directory: Optional. Where to save model's hdf5 file. If directory is not None and save is not False, then save will automatically be set to True

Raises:

• RuntimeWarning: If converted and original model aren't identical, and ignore_tests is False

Returns:

If model has been exported to a file, it will return the name of the file. Else, it returns the converted model.

pytorch_keras_converter.API.convert_and_save

def convert_and_save(model,
                     input_shape,
                     weights=True,
                     quiet=True,
                     ignore_tests=False,
                     input_range=None,
                     filename=None,
                     directory=None):

Arguments:

• model: A Keras or PyTorch model or layer to convert
• input_shape (list, tuple or int): Input shape, without batchsize.
• weights (bool): Also convert weights. If set to False, only convert model architecture
• quiet (bool): If False, display a progress bar and some messages.
• ignore_tests (bool): If tests should be ignored. If weights is False, tests are automatically ignored.
  • If set to True, converted model will still be tested by security. If models are not identical, it will only print a warning.
  • If set to False, and models are not identical, RuntimeWarning will be raised
• input_range: Optional. A list of 2 elements containing max and min values to give as input to the model when performing the tests. If None, models will be tested on samples from the "standard normal" distribution.
• filename: Optional. Filename to give to model's hdf5 file. If filename is not None and save is not False, then save will automatically be set to True
• directory: Optional. Where to save model's hdf5 file. If directory is not None and save is not False, then save will automatically be set to True

Returns:

Name of created hdf5 file

pytorch_keras_converter.cadene_to_tf

def cadene_to_tf(modelList=None, 
                 outputDirectory=None, 
                 quiet=True):

Arguments:

• modelList: A tuple or list of names of the models to convert OR A str telling the emplacement of a file containing names of models to convert (one model per line). If you want a particular config for each model, put it between parenthesis after model's name, for example: 'se_resnet50(pretrained=None)'
• outputDirectory (str): Optionnal. Where hdf5 files should be saved
• quiet (bool): If False, display a progress bar

Raises:

• TypeError: If modelList wasn't a str, list, or tuple

Returns:

A list of created files

Supported models and layers

Supported Cadene's models

• SE-ResNet50
• SE-ResNet101
• SE-ResNet152
• CaffeResNet101
• BNInception
• FBResNet152
• ResNet18
• ResNet34
• ResNet50
• ResNet101
• ResNet152

Other models

Some simple PyTorch models are supported. Supported layers are:

• torch.nn.AvgPool2d(count_include_pad=True)
• torch.nn.MaxPool2d(dilation=1)
• torch.nn.Batchnorm2d
• torch.nn.Conv2d(groups=1, padding_mode='zeros')
• torch.nn.Linear
• torch.nn.ZeroPad2d
• torch.nn.Dropout
• torch.nn.Sequential
• torch.nn.ReLU
• torch.nn.Sigmoid
• torch.nn.AdaptiveAvgPool2d(output_size=1)

To convert a custom PyTorch model, use pytorch_keras_converter.API.

Known issues

Fails to import pretrainedmodels

If pretrainedmodels module isn't available, you can still convert models manually.

1. Download pretrainedmodels from GitHub

git clone https://github.com/Cadene/pretrained-models.pytorch.git

2. Open a Python shell

cd pretrained-models.pytorch
python

3. Create a model

>>> import pretrainedmodels
>>> model = pretrainedmodels.se_resnet50()
>>> input_shape = pretrainedmodels.pretrained_settings['se_resnet50']['input_size']

4. Convert your model

>>> import pytorch_keras_converter as pkc
>>> pkc.API.convert_and_save(model, input_shape)

Can't run the tests

If you use pytorch_keras_converter.API or pytorch_keras_converter.cadene_to_tf it will only show a warning : Warning: tests unavailable!.

If you manually test models using pytorch_keras_converter.tests it will raise the exception InvalidArgumentError.

This is because some TensorFlow layers only support NHWC (i.e channels last) on CPU.

Using Anaconda solves this issue.

Contributing

Contributions are always welcome !

Feel free to create an issue if you found a bug or if you have an idea of improvement. If you want to solve an issue or implement a new feature by yourself, you can fork the repository and make a pull request

This repository is fully PEP8 compliant.

When reporting bugs, please specify your OS, your version of Python, versions of every dependency and if you're using a particular environment (Anaconda for example).

You can use pip freeze to see versions of your Python modules

Authors

• sonibla

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE file for details

Acknowledgments

• Rémi Cadene
• David Picard
• Pierre Jacob