Skip to content

Latest commit

 

History

History
 
 

capsule_em

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
mnist
mnist
 
 
norb
norb
 
 
CONTRIBUTING.md
CONTRIBUTING.md
 
 
DockerfileGPU
DockerfileGPU
 
 
README.md
README.md
 
 
__init__.py
__init__.py
 
 
em_layers.py
em_layers.py
 
 
em_model.py
em_model.py
 
 
experiment.py
experiment.py
 
 
layers.py
layers.py
 
 
model.py
model.py
 
 
requirements.txt
requirements.txt
 
 
run.sh
run.sh
 
 
run_local_gpu.sh
run_local_gpu.sh
 
 
simple_model.py
simple_model.py
 
 
utils.py
utils.py
 
 

README.md

Open source release of EMCAPSNet

This directory contains a preliminary release of the EMCAPSNet model and is under construction.

Training a model on smallNORB requires:

  • GPU (8)
  • TensorFlow
  • Numpy

To evaluate the pre-trained smallNORB checkpoint a GPU is required. Download the checkpoint from: https://storage.googleapis.com/capsule_toronto/norb_em_checkpoints.tar.gz Extract into: $HOME Download the smallNORB tfrecords from: https://storage.googleapis.com/capsule_toronto/smallNORB_data.tar.gz Extract into: $HOME/smallNORB/ The smallNORB tfrecords can also be generated from norb/norb_convert

To test on smallNORB, with a test error of 1.8: From the deepest google_research directory, run:

python -m capsule_em.experiment  --train=0 --eval_once=1 --eval_size=24300 --ckpnt=$HOME/model.ckpt-1 --final_beta=0.01 --norb_data_dir=$HOME/smallNORB/ --patching=False

To get 1.3 (1.4 in the paper), enable patching:

python -m capsule_em.experiment  --train=0 --eval_once=1 --eval_size=24300 --ckpnt=$HOME/model.ckpt-1 --final_beta=0.01 --norb_data_dir=$HOME/smallNORB/ --patching=True

A docker is now added to the repository. Please install docker with NVIDIA support. Then modify the run_local_gpu.sh with your smallNORB directory (-v option). Running ./run_local_gpu.sh will start training the model.