Guide of how to enable PyTorch nn.MaxUnpool2d in Intel OpenVINO.
There are two problems with OpenVINO and MaxUnpool at the moment of this guide creation:
- OpenVINO does not have Unpooling kernels
- PyTorch -> ONNX conversion is unimplemented for
nn.MaxUnpool2d
So following this guide you will learn
- How to perform PyTorch -> ONNX conversion for unsupported layers
- How to convert ONNX to OpenVINO Intermediate Respresentation (IR) with extensions
- How to write custom CPU layers in OpenVINO
MaxUnpool layer in PyTorch takes two inputs - input features from any layer and indices after MaxPool layer:
self.pool = nn.MaxPool2d(2, stride=2, return_indices=True)
self.unpool = nn.MaxUnpool2d(2, stride=2)
output, indices = self.pool(x)
# ...
unpooled = self.unpool(features, indices)If your version of PyTorch does not support ONNX model conversion with MaxUnpool, replace every unpool layer definition
self.unpool = nn.MaxUnpool2d(2, stride=2)to
self.unpool = Unpool2d()where Unpool2d defined in unpool.py. Also, replace op usage from
self.unpool(features, indices)to
self.unpool.apply(features, indices)See complete example in export_model.py.