NOTICE: This repo is no longer actively maintained.
A tool for converting PyTorch models into MatConvNet.
Some of the useful pretrained models available in the torchvision.models module
have been converted into MatConvNet and are available for download at the link
below:
The ResNeXt family of models have also been imported and are available for download:
The conversion script requires Python (with PyTorch installed) and MATLAB.
Converting models between frameworks tends to be a non-trivial task, so it is
likely that modifications will be needed for unusual models. To get started,
see the importer.sh script (this can be modified to import new models).
The easiest way to use this module is to install it with the vl_contrib
package manager. mcnPyTorch can be installed with the following three commands from
the root directory of your MatConvNet installation:
vl_contrib('install', 'mcnPyTorch') ;
vl_contrib('setup', 'mcnPyTorch') ;
Dependencies:
Python3PyTorch- standard numerical Python modules (which should be easy to install with
conda) - The Cadene repo of pre-trained pytorch models (adds support for additional networks which are not included in the main torchvision module)
To run the imported networks, the following matconvnet modules are also required:
- autonn - automatic differenation
- mcnExtraLayers - extra MatConvNet layers
Both of these can be setup directly with vl_contrib (i.e. run vl_contrib install <module-name> then vl_contrib setup <module-name>).
- The normalisation used by the pretrained PyTorch models differs significantly from the typical matconvnet approach (see here for an example).
- The weights in the converted model are modified slightly from the originals to compensate for differences in certain computational blocks. For instance, PyTorch adds an
espilonterm to batch norm denominator during both training and inference.