Skip to content

Replace some weights matrices of linear layers of RNNs with matrices in tensor train format. Get better generalization and a lot less parameters.

Notifications You must be signed in to change notification settings

philip-bl/tensor_train_rnn

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Tensor Train Decomposition for Recurrent Neural Networks

This was supposed to be a replication of Tensor Train Recurrent Neural Networks for Video Classification. See report.pdf for description of what we've done.

Experiments

Different experiments were performed by different people and use different libraries, so please read about each of them separately. You might even need to make different python environments for them.

SVHN Dataset

You can find a script with the experiment in tt_svhn.py. To download the dataset run load_svhn.sh.

You may want to change path to the dataset stored in a variable data_path in tt_svhn.py before running the experiment. To run the experiment simply run:

$ python tt_svhn.py

Youtube Celebrities Faces

You can find everything related to this experiment in ytcelebfaces. In particular, read ytcelebfaces/README.md to learn what libraries you need to install and how to download the dataset.

Deep Shot Boundary Detection Dataset

You can simply start evaluation procedure by calling from the root of the repository:

$ sh deepsbd/test.sh

This command will run test script for 9 transition examples, located at deepsbd/examples. To reproduce the full experiment you should download DeepSBD dataset(https://nsl.cs.sfu.ca/projects/DeepSBD/dataset/DeepSBD.tar.gz, more than 33GB in total) and run deepsbd/train.py with appropriate command line parameters.

About

Replace some weights matrices of linear layers of RNNs with matrices in tensor train format. Get better generalization and a lot less parameters.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published