Comparing Traditional CNNs and Capsule Networks with Greebles Classification.
Implementation of the capsule network is adapted from naturomics/CapsNet-Tensorflow.
Install anaconda or miniconda. Then create the environment with conda.
# Use TensorFlow with GPU
$ conda env create -f environment-gpu.yml
Change the dataset flag in config.py as needed. Run the following script to download and extract data files.
$ python data.py
For affNIST and smallnorb, also run the scripts:
$ python affnist.py
$ python smallnorb.py
For Greebles, generate a greebles dataset with the script found at sherrybai/greebles-generator.
Extract the resulting zip file into the directory data/greebles.
Then, run the script:
$ python greebles.py
Add a new entry to datasets.yml if you wish to test on another dataset not already included.
Training a model:
$ python {model}_train.py
Testing a model:
$ python {model}_eval.py
i.e. python cnn_train.py
Logs and train/validation/test csv files can be found under logs/ and summary/.
Visualize plots with tensorboard:
$ tensorboard --logdir=logdir