gluon2pytorch

Gluon to PyTorch model convertor with script generation.

Installation

git clone https://github.com/nerox8664/gluon2pytorch
cd gluon2pytorch
pip install -e . 

or you can use pip:

pip install gluon2pytorch

How to use

It's the convertor of Gluon graph to a Pytorch model file + weights.

Firstly, we need to load (or create) Gluon Hybrid model:

class ReLUTest(mx.gluon.nn.HybridSequential):
    def __init__(self):
        super(ReLUTest, self).__init__()
        from mxnet.gluon import nn
        with self.name_scope():
            self.conv1 = nn.Conv2D(3, 32)
            self.relu = nn.Activation('relu')

    def hybrid_forward(self, F, x):
        x = F.relu(self.relu(self.conv1(x)))
        return x


if __name__ == '__main__':
    net = ReLUTest()
    
    # Make sure it's hybrid and initialized
    net.hybridize()
    net.collect_params().initialize()

The next step - call the converter:

    pytorch_model = gluon2pytorch(net, [(1, 3, 224, 224)], dst_dir=None, pytorch_module_name='ReLUTest')

Finally, we can check the difference

    input_np = np.random.uniform(-1, 1, (1, 3, 224, 224))

    gluon_output = net(mx.nd.array(input_np))
    pytorch_output = pytorch_model(torch.FloatTensor(input_np))
    check_error(gluon_output, pytorch_output)

Supported layers

Layers:

• Linear
• Conv2d
• ConvTranspose2d (Deconvolution)
• MaxPool2d
• AvgPool2d
• Global average pooling (as special case of AdaptiveAvgPool2d)
• BatchNorm2d*
• Padding2d (constant, reflection, replication)

Reshape:

• Flatten

Activations:

• ReLU
• LeakyReLU
• Sigmoid
• Softmax
• SELU

Element-wise:

• Addition
• Concatenation
• Subtraction
• Multiplication

Misc:

• clamp
• BilinearResize2D
• LRN

Classification models converted with gluon2pytorch

Segmentation models converted with gluon2pytorch

Code snippets

Look at the tests directory.

License

This software is covered by MIT License.