caffe-tucker-decomposition

Caffe implementation of Tucker tensor decomposition for convolutional layers, as described on Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications.

For Caffe implementation of CP tensor decomposition for convolutional layers, refer to caffe-cp-decomposition.

For PyTorch implementation of both CP and Tucker decomposition, refer to PyTorch Tensor Decompositions.

Requirements

• pycaffe, which should include:
  • numpy
  • scipy
  • protobuf
• scikit-tensor (Does not support Python 3)

Usage

Clone this respository and import cnn_tucker to a python script.

The model decomposition is implemented in cnn_tucker.decompose_model(model_def_path, model_weights_path, layer_ranks), where:

• model_def_path: path to the .prototxt file.
• model_weights_path: path to the .caffemodel file.
• layer_ranks: dictionary of convolutional layers to be decomposed with its corresponding ranks.

The ranks can be chosen arbitrarily, but estimation of ranks via VBMF is provided as a utility function cnn_tucker.util.estimate_ranks(weights), where:

• weights: tensor (ndarray) containing the weights of a convolutional layer.

Example of how these are used can be found in cnn_tucker_example.py.

Fast getting started example: Tucker decomposition of VGG-16

You can easily get started with Tucker decomposition of layers conv4_1, conv4_2 and conv4_3 of VGG-16 by following these 3 steps:

1. Clone respository

git clone https://github.com/ddfabbro/caffe-tucker-decomposition.git

2. Add it to PYTHONPATH

export PYTHONPATH=$PYTHONPATH:$(pwd)/caffe-tucker-decomposition

3. Run Tucker decomposition. NOTE: VGG-16 caffe model will be downloaded.

python caffe-tucker-decomposition/cnn_tucker_example.py

Aside from downloading VGG-16 caffe model, this example should take less than 1 minute.

Limitations

• Convolutional layer paramaters that are non-uniform (e.g kernel_h, kernel_w, pad_h, pad_w, stride_h and stride_w ) are not supported. However, you can easily modify the code to your needs.
• Multi branch networks are not supported.

Additional resources

• Accelerating deep neural networks with tensor decompositions.
• Python implementation of Variational Bayes Matrix Factorization.