README.md

Guide of how to enable PyTorch nn.MaxUnpool2d in Intel OpenVINO.

Description

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

Get ONNX model

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.